Traditionally, publishers such as the Wall Street Journal sold advertising space by closing large deals with ad agencies. The agency was the key player in the middle, connecting publishers to brands. This intermediary had the knowledge of how to run advertising campaigns while as they purchased ad inventory in bulk, they also typically could offer brands a better deal.

In the internet era, this value chain started to fragment considerably and a wide variety of new players emerged. First, there are the ad networks, which dominate large swaths of the online advertising market. The two dominant ad networks, Google and Meta, are amongst the largest companies in the world. Basically, on these ad networks, advertisers and brands can buy inventory themselves to run ad campaigns.

Secondly, there is the open internet, where publishers open up their ad inventory for real-time auctions with each impression being sold to the best bidder. Both price and quality of the advertiser can be taken into account here. For example, Netflix will prefer to show an ad from Nike as opposed to one from OnlyFans. To run this auction process, key players in the middle are the supply side platforms (SSPs) and the demand side platforms (DSPs).

So, when a user visits a website or smartphone app, the publisher’s server will detect that an ad impression is available. The server then sends a request to the publisher’s SSP which collects information about the consumer. Data collected can include age, gender, interests, browsing history, browsing behavior, the smartphone device in use, typical engagement metrics, and geolocation. The richer and more precise the dataset on the user, the more valuable the ad impression becomes.

The SSP forwards this impression to the DSPs, which evaluate the impression to place their bids. To more accurately value this impression, the DSP can make use of additional data it has on the end user, which may have been collected by the DSP, the brand itself, or a third-party data supplier. Data collection on internet users is a fairly complicated industry, with tons of continuously evolving regulations.

The Trade Desk (TTD) is the dominant DSP as it has both the widest and best quality ad inventory available, some of the richest data sets on consumers, the best platform capabilities, and as it is an independent and neutral arbiter in the middle. This final point means that there is no conflict of interest between the advertiser and the DSP. Unlike Amazon’s or Google’s DSPs, TTD doesn’t own ad inventory itself, so it won’t try to unfairly promote its own impressions to advertisers. The sole aim of TTD is to identify the best available impressions for its customers.

Typically, advertisers make use of Amazon’s DSP when they want to run a campaign on amazon.com or Prime Video, while Google’s DSP is mainly leveraged to run campaigns on Youtube. However, the aim of large brands is to re-target potential end customers across a variety of platforms and with different ads. So they might show you one ad on TikTok, another on Spotify, the Financial Times, X, LinkedIn, Netflix, etc. To realize such a multi-platform campaign you need TTD.

An additional advantage is that a good DSP keeps track of which ads have already been shown to each end user. Showing a potential customer the same ad for the twentieth time is just wasted money. So, a clever DSP can create substantial cost savings for customers as it can figure out when to change ads to re-target a customer, or on the other hand forego the impression altogether.

This is a former manager at TTD explaining the company’s strong position in the DSP market (via Tegus):

“The Trade Desk’s value proposition is largely based on being independent as a neutral platform in the middle, while having access to large swaths of inventory across pretty much every streaming player within the connected TV (CTV) space, audio streaming, ads on mobile apps, and every publisher within the programmatic space. Then, The Trade Desk has all sorts of connections to make it easy for the client to bring in their own and third-party data, while also additional data from a publisher itself can be integrated directly. For example, Disney might be willing to share some insights about their audience with The Trade Desk which can be tied to the available impressions.

In CTV, advertisers have been migrating their programmatic buying over to The Trade Desk. Historically, Roku was someone that you would have to work with directly and Netflix was in the same camp, but both are now opening up. And you can also buy Hulu inventory over The Trade Desk. There’s a lot of opportunity in CTV.

Audio is another channel that is often forgotten, and players like Spotify and Pandora are integrated with the Trade Desk as well. Audio pales still in comparison with media spend on Connected Television, but it is growing because the audio platforms have authenticated and logged-in users, and having those types of identifiers is helpful in digital advertising.

The average client on The Trade Desk typically has an annual digital media budget upwards of $10 to $12 million, with some that are spending more than several hundred million a year. Probably more than half of these budgets on The Trade Desk are agency driven, both from major holdcos and smaller specialized agencies, but there are an increasing number of clients that are working with The Trade Desk directly. These have their own in-house teams of programmatic trading desks, but this is only for large companies because it’s very costly.

The Trade Desk is now introducing AI to automate the setup of campaigns, effectively taking over that manual work that’s placed on agencies or in-house trading teams. This creates an opportunity for smaller advertisers to work with The Trade Desk directly, as you don’t have to rely anymore on sophisticated programmatic buyers to set up and manage your campaigns. You can now do this by pressing a few buttons and then smaller advertisers can also get into additional channels like connected television. Minimizing the burden of programmatic buying will allow more advertisers to move away from the ad networks of Google and Facebook—where you can activate a campaign with a couple of buttons and a credit card—to a platform like The Trade Desk.”

To further strengthen its strong position in the programmatic value chain, TTD has been building direct connections with the large publishers themselves. The attraction for publishers is that most brands are running campaigns on TTD anyways and so now they can forego having to pay the SSP its take rate, which can be substantial. This new feature is what TTD calls OpenPath, and the attraction for TTD is that they can offer the widest and best quality inventory from the open internet on their platform. For example, one smaller agency mentioned in a Tegus interview that it was impossible for them to obtain ad impressions for the Olympics on NBC, however, these easily became available over TTD.

Traditionally, we had the below value chain for digital ads on the open internet. An advertiser uses a DSP to connect to a SSP and then bid on an impression from a publisher:

However, with OpenPath, TTD is putting itself at the center to connect advertisers and publishers on the open internet:

Over the past 1.5 years, TTD shares sold off heavily due to a variety of factors. We can see that EPS performance actually still has been positive, with negative share price momentum being driven by multiple compression. The shares are now trading on a 10x forward PE:

Is TTD a buying opportunity or is the story structurally broken? We’ll go through a variety of interviews with industry insiders as well as insights from sell side analysts to evaluate..

The Trade Desk Disrupted?

Read more

We’ve been big software bulls over the past few decades, but the rise of AI and LLMs will likely have a profound impact on the industry. At this stage it’s not clear yet to what extent positives will outweigh negatives. Having recently done a large coding project, working together with LLMs probably boosted productivity with around 4 to 5x. Thus, an obvious bear case would be that increasing productivity levels in the coming decade will lead to a reduction of software seats that enterprises will have to purchase. Another bear case is that SaaS apps have an abundance of functionality built-in through massive logic-based code bases in the back, however, if users start interacting more with LLMs via natural language, a lot of this code might actually become obsolete. Basically, the business application of the future could be more like a portal where a user interacts with an AI agent, which carries out all tasks while interacting with a database in the back.

This is Satya Nadella, Microsoft’s CEO, on the potential disruption of SaaS on the B2G podcast:

“Business applications are essentially CRUD databases with a bunch of business logic. The business logic is all going to these agents and they’re going to update multiple databases, and all the logic will be in the AI tier so to speak.”

In a subsequent interview with The Economic Times, Nadella stated:

“A lot of the business logic will move to a new tier, which then will be a multi-agent tier that needs to be orchestrated. It's going to be an agent that will orchestrate across multiple SaaS applications. Just like I can build a spreadsheet, I will build thousands, hundreds of agents that are working to help me streamline my own work”

However, it’s not all doom and gloom in the outlook for SaaS. Strong controls will have to be implemented, likely via business logic, for example, ‘this user is only allowed to interact with this part of the database’ etc. It’s also likely that common workflows where a human needs to be involved can easier be done via pre-built business logic. For example, even though the AI will be writing a lot of code, we still need an IDE because humans need to validate and adapt the code. Similarly, we still need software to manage and deploy the software, which is largely running on traditional logic. In the same vein, we still need Microsoft Office. Even though AI might be preparing a lot of Powerpoints and Excel sheets, humans again still need to review and adapt them. Thus, the future we see in the coming five years is one where AI will be increasing automation in a wide variety of fields, but traditional software will still be needed to manage overall workflows.

Another positive is that lost user seats from higher levels of automation can be taken up by companies in new industries that will emerge, which has been the historical trend. A few hundred years ago, 90%+ of the population was employed in farming. Large gains in agricultural productivity, and subsequently industrial productivity, didn’t lead to mass unemployment. Instead, a vast services industry emerged where an abundance of skilled workers were needed ranging from finance to tech and health care. Similarly, software apps are also introducing automated workflows within their environments thanks to AI and this compute time can start being monetized over time.

Another positive from AI for the SaaS names is that their skilled workforces are highly expensive, resulting in massive salaries and share based compensation (SBC). If worker productivity can be improved with 2x or more over time, there’s a lot of potential for SaaS companies to become leaner organizations and boost earnings margins. Not only can coding productivity see a massive boost, but also services productivity as more customer queries can get handled by an AI. Similarly, for new software implementations, an AI teacher can get new users up and running in much faster ways, and the AI can also write customized software extensions which enterprises often need.

When it comes to software investing, historically, most of the outperformance has been driven by strong top line growth rates. However, there is a good reason to believe that with AI, investors can also start being much more enthusiastic when it comes to margins.

The Bull Case for the Cloud Data Market

In the remainder of this article, we will zoom in on the cloud data market for analytics. The best long term bull case for Snowflake, and its peers such as Databricks, was probably given by Intel’s former CEO a few years ago, who was running virtualization software company VMware previously:

“We are in year 20 of the public cloud, and you have 60%-plus of compute in the cloud, but 80%-plus of data remains on-prem.”

For premium subscribers, we will discuss Snowflake’s current battle with Databricks for the cloud data market, and the company’s future opportunities. We will also give the detailed insights from a large number of well known, blue chip enterprises on how they’re using cloud data systems and where they are increasingly moving workloads to. And we’ll discuss the same topics with a VP at Accenture, one of the largest system integrators in the world. Finally, we’ll do an analysis of the financials and valuations for Snowflake as well as names in the wider software space, and we’ll conclude with our stock picks.

Snowflake vs Databricks

Although historically both Snowflake and Databricks came from different backgrounds—Databricks was oriented towards running data science workloads in the cloud while Snowflake was largely a horizontally scalable SQL-based data warehouse—both companies have been moving into each other’s territories over the last years. Snowflake has been introducing support for unstructured and semi-structured data, in order to transform its cloud platform to become more of a data lakehouse like Databricks. Snowflake introduced also Snowpark so that data engineers can run data workloads on Spark-like clusters via a variety of popular data science languages such as Python, Java and Scala; which is really the bread-and-butter of Databricks. At the same time, Databricks has been introducing support for SQL and ACID transactions to allow its data lake to also function as a cloud-based data warehouse, naming it ‘the data lakehouse’. Traditional business analytics built on top of SQL and ACID transactions remains the core of Snowflake’s business today.

Read more

The three big public clouds were already amongst the best businesses in the world, and AI has again accelerated growth rates. For Google’s cloud platform, growth rates stepped up again towards the 30% range as from March last year:

Demand this year will be constrained by how fast Google is adding data center capacity, so clearly fast growth rates over the last year have taken them a bit by surprise as they haven’t been adding data centers fast enough. Getting a greenfield data center up and running takes around two-three years, and so sudden spikes in cloud demand growth can limit the availability of data center supply. This is Google’s CFO describing the current demand-supply environment:

“In Cloud, we're in a tight demand-supply environment. And given that revenues are correlated with the timing of deployment of new capacity, we could see variability in cloud revenue growth rates depending on capacity deployment each quarter. We expect relatively higher capacity deployment towards the end of 2025.”

So it’s no surprise that the company keeps stepping up capex, as demand growth has been outstripping supply growth. The company is forecasting this year to invest around $19 billion in capex every quarter.

Customers are migrating large amounts of data to the cloud to fine-tune AI models, as well as AI workloads by querying LLMs in the cloud over API. An API is basically a web address where your code can send requests to, and then the server will process your query and send a response back. For example, a front-end app that’s running on a smartphone can send customer queries to the servers, and then AI agents running on these servers will help the customer sort things out. This is Microsoft’s CEO Satya Nadella on the data migration to the cloud:

“Microsoft Fabric has more than 21,000 paid customers, up 80% year-over-year. Fabric brings together data workloads like data warehousing, data science, real-time intelligence, along with Power BI into one end-to-end solution. Real-time intelligence is now the fastest-growing workload in Fabric with 40% of customers already using it in just 5 months since becoming generally available. And the amount of data in our multi-cloud data lake, OneLake, has grown more than 6x over the past year.”

Microsoft Fabric is basically a Databricks-clone, which is the hottest app for data science workloads in the cloud. The advantage of Databricks is that you can run it in multiple cloud environments, reducing vendor lock-in risk for customers. But the drawback is that the large cloud players such as Microsoft, Google and Amazon could subsidize their own data and analytics platforms, as they’re making enough money elsewhere in their clouds anyways, such as in basic compute. This is Nadella on how customers are using Azure as a platform to train and run AI models:

“Now on to AI platform and tools. Foundry is the agent in AI app factory. It's now used by developers at over 70,000 enterprises and digital natives to design, customize and manage their AI apps and agents. We processed over 100 trillion tokens this quarter, up 5x year-over-year, including a record 50 trillion tokens last month alone. And 4 months in, over 10,000 organizations have used our new agent service to build, deploy and scale their agents.

This quarter, we also made a new suite of fine-tuning tools available to customers with industry-leading reliability, and we brought the latest models from OpenAI along with new models from Cohere, DeepSeek, Meta, Mistral and Stability to Foundry. And we've expanded our family of SLMs with new multimodal and mini models. Model capabilities are doubling in performance every 6 months, thanks to multiple compounding scaling laws.”

Note that half of all tokens processed over the last year were processed during the last month, which means that AI is exploding in the cloud. This all sounds promising for the hyperscalers and their suppliers, especially Nvidia where the market is currently assigning a bottom valuation multiple to its earnings, on the assumption that we’re about to peak in terms of AI spend. However, listening to both Google’s and Microsoft’s calls, this isn’t happening anytime soon.

Microsoft has been a beast in the cloud—you can see clearly how they’ve strongly outperformed Amazon Web Services (AWS) over the last two years, adding much more annually recurring net new revenues to their business:

Google Cloud Platform (GCP) is the smaller cloud, but is gradually picking up steam as well. The chart below shows how their revenue growth rates are slightly below Microsoft’s Azure. Thus, Azure is not only adding the most revenues on an absolute basis, but it is also growing the fastest percentage-wise.

Under Nadella’s leadership, Microsoft has transformed from a rather solid, but dull, software provider associated with Windows, Office and other enterprise software such as databases and ERP systems, into a developer friendly platform providing both the world’s leading cloud as well as the best tooling to develop and manage software with such as GitHub, VSCode and even Linux on Windows.

There was some panic in the markets around Microsoft reducing their investment outlook over the last few months, however, the capex outlook remains strong. This is the company’s CFO on their investment plans:

“We remain committed to investing against the strong demand signals we see for our services. So our earlier comments on FY '26 capital expenditures remain unchanged, we expect CapEx to grow. It will grow at a lower rate than FY '25 and will include a greater mix of short-lived assets, which are more directly correlated to revenue than long-lived assets. These investments will ensure we continue to lead in the cloud and AI opportunity ahead.”

Also Microsoft is seeing a tight demand-supply balance in the cloud, this is the CFO during the Q&A:

“We talked a little bit about pulling some of that space to be ready earlier and being able to deliver that to customers this quarter, which is really good work by the teams as we continue to get more and more efficient at that process. I did talk about in my comments, we had hoped to be in balance by the end of Q4, but we did see some increased demand as you saw through the quarter. So we are going to be a little short still, a little tight as we exit the year, but we are encouraged by that.”

This is Nadella on how they’re thinking about what’s needed in the data center buildout:

“There's a demand part to it, there is the shape of the workload part to it, and there is a location part to it. So you don't want to be upside down on having one big data center in one region when you have a global demand footprint. You don't want to be upside down when the shape of demand changes because, after all, with essentially pre-training plus test-time compute, that's a big change in terms of how you think about even what is training. The idea that you can have test-time compute plus pre-training and then some of the great optimization at inference-time that has all happened, for every Moore's Law change and movement, there's probably a 10x improvement because of software. That's sort of what's happening with these models.”

Spreading investments out globally makes a lot of sense as AI inference will become a large market in the coming years, and it’s necessary to run these workloads closer to the end customer. The main risk to being able to capture this strong and structural demand growth in the cloud, is having access to the necessary power supplies. A director at Microsoft via Tegus discusses current constraints the industry is facing:

“The power and space constraints for data centers are still there, and it's a growing problem. Earlier, the constraint was GPUs. Right now, we have power and space constraints, and the way we are dealing with it is long-term planning. We are going and building ahead on data centers aggressively. Microsoft is looking for new ways of power, Google is looking for new ways of power, and hyperscalers are partnering with multiple other organizations to get rental power for example. Recently, there have been a lot of innovations around nuclear power and small modular reactors. Heavy investments on renewable power with contracts ranging up to 10-15 years have been in place. In summary, we're doing a lot of thinking for long-term as well as for short-term needs, and we're renting out capacity, space or power from different colocation facilities. There is no demand risk for this capacity. Right now, the demand is more than the supply for sure.”

Subscribe now

For premium subscribers, we’ll dive deeper into current developments in hyperscaling and AI. We’ll discuss current developments in custom AI acceleration as well in AI apps, which long term will capture most value in the AI value chain. Historically, investing in differentiated software, ideally with network effects, has been the most lucrative area in the tech value chain. Examples include Google, Meta, Microsoft etc. This time, during the current AI revolution, of which we’re still in the early stages, this historical pattern will be repeated.

Note also that historically, during each new technology wave, the winners of the new wave eventually dwarfed the winners from previous technology waves. So, the winners from the PC and client-server era became much larger than the winners from the mainframe era, and the winners from the internet and cloud eras again dwarfed the tech winners from previous decades. Again, we think that this time will be no different as AI has the capabilities to automate large swaths of the economy. Thus, we think that the big winners in the AI era will dwarf the winners from the internet era in the coming decade. It’s a good bet that the first company with a market cap north of $10 trillion will leverage AI to automate parts of the global economy. Thus, this is a space we’re regularly keeping track of with the aim of identifying companies that are well positioned, with a competitive advantage, to capture these large opportunities.

Current developments in custom AI acceleration

Read more

Traditionally, publishers such as the Wall Street Journal sold advertising space in their media by closing large deals with ad agencies. The agency was the key player in the middle, connecting publishers to brands. This intermediary had the know-how to run advertising campaigns and as they purchased ad inventories in bulk, they also typically could offer brands a better deal.

In the internet era, this value chain started to fragment considerably and a wide variety of new players emerged. First, there are the ad networks, which dominate large swaths of the online advertising market. This is why the two dominant ad networks, Google and Meta, are amongst the largest companies in the world. Basically, on these ad networks, advertisers or brands can buy inventory themselves and run ad campaigns.

Secondly, there is the open internet, where publishers open up their ad inventory for real-time auctions where each impression is sold to the best bidder. Both price and the quality of the advertiser can be taken into account here, for example, Netflix will prefer to show an ad from Nike as opposed to one from OnlyFans. To run this auction process, key players are the supply side platforms (SSPs) and the demand side platforms (DSPs).

So, when a user visits a website or smartphone app, the publisher’s server will detect that an ad impression has become available. The server then sends a request to the publisher’s SSP which collects information about the consumer. Data collected can include age, gender, interests, browsing history, browsing behavior, intent, the smartphone device in use, typical engagement metrics and geolocation. The richer and more precise the dataset on the user, the more valuable the ad impression becomes.

The SSP forwards this impression to the DSPs, which evaluate the impression to place their bids. To more accurately value the impression, the DSP can join in additional data it has on the end user, which can have been collected by the DSP, the brand itself, or a third-party supplier. Data collection on internet users is a complicated industry as well, with tons of continuously evolving regulations.

Basically, The Trade Desk (TTD) is the dominant DSP as it has both the widest and best quality ad inventory available, some of the richest data sets on consumers, the best platform capabilities, and because it is an independent and neutral arbiter in the middle. This final point means that there is no conflict of interest between the advertiser and the DSP. Unlike Amazon’s or Google’s DSPs, TTD doesn’t own ad inventory itself, so it won’t try to unfairly promote its own impressions to advertisers. The sole aim of TTD is to identify the best available impressions for its customers.

Typically, advertisers make use of Amazon’s DSP when they want to run a campaign on amazon.com and similarly, Google’s DSP is mainly leveraged to run campaigns on Youtube. However, the aim of large brands is to re-target potential end customers across a variety of platforms and with different ads. So they might show you one ad on TikTok, another on Spotify, the Financial Times, X, LinkedIn, etc. To realize a multi-platform campaign like this, you basically need TTD.

An additional advantage is that a good DSP keeps track of which ad has already been shown to which end user. Showing a potential customer the same ad for the twentieth time is usually just wasted money. So a clever DSP can create substantial cost savings for customers as it can figure out when to change ads to re-target a customer or forego the impression altogether.

This is a former manager at TTD explaining the company’s strong position in the DSP market:

“The Trade Desk's value proposition is largely based on being independent as a neutral platform in the middle, while having access to large swaths of inventory across pretty much every streaming player within the connected TV (CTV) space, audio streaming, ads on mobile apps and every publisher within the programmatic space. Then, The Trade Desk has all sorts of connections to make it easy for the client to bring in their own and third-party data, while also additional data from a publisher itself can be integrated directly. For example, Disney might be willing to share some insights about their audience with The Trade Desk which can be tied to the available impressions.

In CTV, advertisers have been migrating their programmatic buying over to The Trade Desk. Historically, Roku was someone that you would have to work with directly and Netflix was in the same camp, but both are now opening up. And you can also buy Hulu inventory over The Trade Desk. There's a lot of opportunity in CTV.

Audio is another channel that is often forgotten, and players like Spotify and Pandora are integrated with the Trade Desk as well. Audio pales still in comparison with media spend on Connected Television, but it is growing because the audio platforms have authenticated and logged-in users, and having those types of identifiers is helpful in digital advertising.

The average client on The Trade Desk typically has an annual digital media budget upwards of $10 to $12 million, with some that are spending more than several hundred million a year. Probably more than half of these budgets on The Trade Desk are agency driven, both from major holdcos and smaller specialized agencies, but there are an increasing number of clients that are working with The Trade Desk directly. These have their own in-house teams of programmatic trading desks, but this is only for large companies because it's very costly.

The Trade Desk is now introducing AI to automate the setup of campaigns, effectively taking over that manual work that's placed on agencies or in-house trading teams. This creates an opportunity for smaller advertisers to work with The Trade Desk directly, as you don't have to rely anymore on sophisticated programmatic buyers to set up and manage your campaigns. You can now do this by pressing a few buttons and then smaller advertisers can also get into additional channels like connected television. Minimizing the burden of programmatic buying will allow more advertisers to move away from the ad networks of Google and Facebook—where you can activate a campaign with a couple of buttons and a credit card—to a platform like The Trade Desk.”

To further strengthen its strong position in the programmatic value chain, TTD has been building direct connections with the large publishers themselves. The attraction for publishers is that most brands are running campaigns on TTD anyways and so this way they can forego having to pay the SSP its take rate, which can be substantial. This new feature is what TTD calls OpenPath, and the attraction for TTD is that they can offer the widest and best quality inventory from the open internet on their platform. For example, one smaller agency mentioned in a Tegus interview that it was impossible for them to obtain ad impressions for the olympics on NBC, however, these easily became available to them via TTD.

So traditionally, we had the below value chain for digital ads on the open internet, i.e. an advertiser uses a DSP, to connect to a SSP and bid on an impression from a publisher:

However, with OpenPath, TTD is aiming to establish the below landscape, putting itself at the center to connect advertisers and publishers on the open internet:

The Trade Desk’s Future Growth Drivers

With the rise of the internet, TTD has been a spectacular growth story over the last decade. Quarterly revenues climbed from $18 million in 2015 to $741 million last quarter:

The company’s long term outlook remains strong. Obviously, OpenPath is a big opportunity for TTD to put itself at the center of programmatic advertising, and another large opportunity remains traditional media which is mainly comprised of linear TV these days. This is still a $300 billion global market, which continues to dwarf the small dark blue segment in the global ad market where TTD plays:

Basically, only $12 billion in annual ad spend is flowing through TTD currently while linear TV remains ripe for disruption, more on this in a bit.

To summarize the chart above, there are four blocks in the global ad market— search is dominated by Google, social is dominated by Meta, then there is traditional media which is mainly still linear TV, and finally there are the online video, audio and display (websites) markets, which are the bread-and-butter of TTD.

As the median age of linear TV viewers in the US is currently 65 years, this whole segment of the ad market will continue to move online in the coming decades. So traditional media will continue to shrink as eyeballs continue to move online. The other interesting feature of the online market is that ad pricing can be higher due to better targeting capabilities. TTD mentions that connected TV (CTV) typically sees a cost-per-mille (CPM, i.e. the price of 1,000 impressions) of $20 compared to only $10 for traditional TV:

The next bull argument is that TTD has a lot of runway for international expansion. Currently 88% of company revenues are coming from North-America, but the rest of the world actually contributes 60% of the global ad market:

Jeff Green, TTD’s founder, has regularly been discussing the company’s long term outlook during previous conference calls:

“A tailwind that is setting up The Trade Desk for continued outsized gains is the move to AI. Our AI product, Koa, is a great copilot for traders. But this is only the beginning, there are endless possibilities for us as we have one of the best data assets on the open internet. We're looking across our entire suite of products and are asking how they all can be advanced by AI.

Another factor that is changing the entire landscape of digital advertising are the changes happening at Google. Google has an incredibly dominant money-making business in search and another one in YouTube, with now also an incredible opportunity in cloud and AI. Because of these opportunities, Google could continue to downgrade their network business to focus on those prospects. After all, they operate at a disadvantage on the buy side of the open internet because Google lacks objectivity.

Google also has a pending antitrust trial that has created massive ripples throughout the global ad tech ecosystem. The outcome of the trial itself is less important than the change in behavior that will likely come from it, Google will remain under tremendous scrutiny from regulators around the world. The convoluted role they play in their network business, laid out very clearly by the US Department of Justice, is likely to result in more fair play. This presents a tremendous opportunity for us and for the broader ad tech ecosystem.

CTV is creating incremental secular tailwinds in the market. It's both our largest channel and our fastest growing. The premium open internet is where consumers are increasingly spending most of their time. Companies like Netflix, Roku, NBC, Spotify, Disney and Fox have redefined the consumers' internet experience over the last few years. You may remember at our last Investor Day, that we showed a graph of what our clients were willing to pay for CTV ad impressions on the open market and in many cases, we've been significantly higher than the price that the media companies were negotiating themselves in direct deals.

The content arms race is more intense than ever and media companies need to monetize their content as effectively as possible. Two weeks ago, Variety reported that the top six media companies will increase content spending this year by 9% to a record $126 billion, all of which needs to be funded. Media companies now need to provide advertisers with more insights and tools like UID2 and OpenPath are helping provide that signal.

Many of the major CTV players around the world are aggressively implementing OpenPath now. One major news publisher saw their revenue increase by more than 25% with OpenPath. Disney is now working towards 75% of their ad sales being automated by 2027, with the vast majority of those impressions being biddable. Another CTV leader that has embraced OpenPath is VIZIO, which has more than 24 million active devices in the United States and more than 300 ad-supported CTV channels. VIZIO deployed OpenPath and saw a 39% improvement in revenue from our platform. Because of the number of deals that we've signed recently, we are extremely confident that 2025 will be the year when OpenPath enters the steep acceleration phase of its S-curve growth.

A next growth vector for us is audio. The biggest players in audio are global and they can capture one of the biggest opportunities in advertising and media today, which is the delta between time spent in the audio channel and the amount of ad budgets heading to that channel. Digital audio is at the early stages of its evolution and the channel is in a similar position to where CTV was a few years ago. Consumers in the US spend an average of 3 hours per day consuming digital audio, up significantly over the last 5 years, and advertisers are eager to capitalize on this emerging advertising channel. Spotify has been at the forefront of this mass consumer shift and is building out their programmatic capabilities.

Then we have a massive opportunity in retail media. Amazon has showcased the benefits of using purchase data to make retail businesses work better. And finally, we have live sports. Sports is some of the most premium content and media, it is scarce and highly sought out by brands, and it is built for programmatic advertising.”

The opening up of Google’s network business is a large opportunity. In this business segment, independent apps and websites can leverage Google’s adtech technology to show ads to their users. So this is a separate business from search and Youtube, and includes adtech such as AdSense, AdMob and Google Ad Manager. As in all of its markets, Google is very dominant in this space and this has been drawing attention from regulators around the world.

If Google would have to open up its network business by making these impressions on third-party apps and websites biddable, it represents a big opportunity for TTD. Google Network alone, i.e. Google’s smallest business unit, is more than 12x the size of TTD:

As mentioned, Google Network is the company’s smallest business unit:

An organic growth story for TTD is that there are a wider variety of popular apps today that audiences are spending large amounts of time in. Ten years ago, these types of dominant apps were mostly Facebook, Instagram and Youtube; but now you have as well TikTok and Netflix being major apps, and with also a number of other large platforms that continue to gain scale such as Spotify, X, Reddit, Hulu, Roku, LinkedIn, Amazon Prime, Pinterest, etc. The good news is that most of this inventory is accessible over TTD, allowing the leading DSP to gain share in the overall ads market. As a result, the share of the two dominant walled gardens in the digital ads world—Meta and Google—has been falling:

Audiences are spending especially more time in audio and connected TV apps:

Although both of these are largely monetized via subscription models currently. However, as 63% of Spotify’s active users are currently ad-supported, there is a large opportunity to increase revenues from this user base by integrating them into OpenPath. Similarly, for Netflix, a new ad-based revenue model is one of the key growth stories for the coming years. A compelling advantage with both the audio and CTV formats is that you can’t skip the ad, whereas on social media such as Facebook, many just quickly scroll past the ad while largely ignoring the content.

Also in CTV, most of the audience growth is in ad supported streaming:

Finally, retail media networks are a great way for brands to promote their products. For example, if you’ve launched a new coffee brand and want to boost sales on walmart.com, you can bid for prime ad inventory. Many e-commerce sites are now realizing this so this should remain an attractive opportunity.

Thus, all was looking really good for TTD and this was a consensus long for many active equity managers. However, the combination of a valuation that required perfect execution, the recent earnings miss, and the subsequent overall market sell-off due to macro fears caused the company’s share price to fall off a cliff:

The Q4 Miss

This is Jeff Green again, discussing what went wrong in the quarter:

“For the first time in 33 quarters as a public company, we fell short of our own expectations. For Q4, the reality is that we stumbled due to a series of small execution missteps while simultaneously preparing for the future. That said, we see a larger and faster-growing market than we originally expected, and this is why we are recalibrating our larger company for an even stronger future. I want to highlight four major changes we've made at The Trade Desk in the last few months.

First, we did the largest reorganization in company history in December. For most people in the company, we provided a much clearer view of their roles and responsibilities, and that mostly also meant a change in reporting structure. Additionally, we streamlined client-facing teams to reduce complexity. Now, some teams will focus on brands while others will focus on agencies. Our commitment to agencies remains strong, but we are also expanding brand direct relationships, particularly through joint business plans (JBPs) which grow 50% faster than the rest of our business.

Second, we understand that this moment requires us to scale our internal operations and we will hire senior talent to support long-term growth. For example, we've done all of this without a COO for some time and there is absolutely no reason why we shouldn't add a world-class COO to the team. There are opportunities for us to get more efficient internally and operationally more rigorous.

Third, we have increased our resource allocation on brands. Advertisers are becoming more strategic and data-driven in their media buying decisions and that's great for us. While this shift has caused short-term fluctuations, it's ultimately aligned with our long-term strength. This is evident in the growing number of JBPs that we've secured with over 100 of the world's leading brands, and many of them in the second half of last year.

Fourth, we revamped our product development process, shifting back to smaller and more agile development teams that release updates weekly instead of drifting towards waterfall methods. Our engineering team is now divided into nearly 100 scrum teams to more easily ship and collaborate with the business teams. I expect this to accelerate Kokai enhancements and to complete the transition of 100% of our clients from Solimar to Kokai during this calendar year.

During the quarter, there were a series of decisions we could have made to enhance the short-term performance of the company and neglect the long-term. We consistently chose to focus on the long-term opportunity as this is in the best interests of all our stakeholders.”

We already saw above in TTD’s revenue growth chart that growth remains robust. However, due to the miss and the weaker than expected Q1 guidance, the market is now looking for 18% top line growth in 2025. You can see below that estimated revenues for the year basically got downgraded from $2.97 billion to $2.89 billion, i.e. a downgrade of 2.7%:

While EPS were downgraded by 8.6% due to operating leverage:

Thus, most of the 52% share price fall has been due to multiple compression, with the shares now trading at their cheapest valuation since IPO:

For premium subscribers, we’ll dive into current developments in the adtech industry and we’ll share the insights from a VP at one of the world’s largest ad agencies, with over $16 billion in annual revenues, on both current developments in the industry and for TTD. Additionally, we’ll review the various bear cases that caused the shares to sell off sharply such as rising competition, concerns about take rates and whether TTD’s new platform rollout—Kokai—has been botched. Finally, we’ll review the financials and valuations of key stocks in this space and which ones we see as top picks here, and whether TTD should represent a buying opportunity or whether this is a falling knife.

Inside the AdTech Industry: The Trade Desk's Problems & Outlook

Read more

Companies that dominate their industries are often amongst the best performing assets. These types of companies usually have strong pricing power, which translates into high EPS growth due to high incremental margins. Subsequently, earnings can be used to buy back shares, which again fuels EPS growth. So it’s not uncommon to see these types of stocks compound over multiple decades, translating into outsized returns for shareholders.

Today, we will do a deep dive on a name that is on its way to consolidate its global industry. Barriers to entry are high, as the company managed to build its compelling platform by decades of investment. Currently, there isn’t a single competitor which can offer a suitable alternative. While valuation isn’t cheap, the stock has gotten a lot cheaper following the market sell off in the last few weeks, and the company is very profitable while continuing to drive high double-digit top line growth rates. Additionally, EPS growth is at 1.5x those top line growth rates as incremental margins are very high. Finally, there is very limited share-based compensation so investors are getting a clean earnings number.

The company definitely isn’t a household name, in fact, we’ve never seen it mentioned on social media such as X. Although investors reading a variety of business newspapers will have seen the occasional article on the company. It’s a mid-cap that has sufficient liquidity even for large funds to take a position. We see an IRR of 21% in this name in the coming years and with plenty of potential to exceed that due to the company’s new products, which have strong potential to surprise the market on the upside. This should be an interesting name both for shorter term investors, as well as for Warren Buffett-style long term investors who are happy to hold companies for 5-10 years.

An Emerging Monopolist

Read more

Manhattan Associates is a leading supply chain software provider to manage a company’s distribution operations. The company’s cloud platform is especially used in consumer goods, but it is also increasingly being adopted in other sectors such as in industrials, pharma and even energy. Its platform can scan and validate incoming shipments, optimize inventory throughout the supply chain, determine the best routing for transportation, manage orders and returns, track worker productivity and integrate with automated robotics systems. Gartner rates Manhattan as the number one in warehouse management systems:

And also Forrester sees Manhattan as the number one in order management:

Order and warehouse management have grown increasingly complex with the rise of digital. The warehouse system for example has to figure out whether a product can be picked up from a nearby store when a consumer is browsing the company’s website. Inventories have to be updated based on all sales via the different channels such as in-store collection, online delivery and in-store shopping. And keep track of any possible returns. Having a modern solution that provides the necessary information to both employees and customers of where and when products will be available is obviously a competitive advantage.

Speaking to several consultants, Manhattan is clearly the leading solution in the space. The company has invested substantially over the last decade to build out a modern solution in the cloud with wide capabilities. Competitors are playing catch up. Blue Yonder (Panasonic) for example is a number of years behind Manhattan and although Korber is still listed as a strong competitor by Gartner above, system integrators mention that they are far behind.

As a result, Manhattan is now shifting its customer base to the cloud. This is a highly attractive business model for investors. Cloud subscriptions bring in higher revenues than the classic ‘license + maintenance’ business model in software, while getting rid of lumpy license sales and moving these to annually recurring subscription revenues. This removes cyclicality for investors while giving high future visibility on the business. The below table illustrates the idea:

And subscriptions also allow for more regular price increases, which can be part of the contract. A nice feature of owning a business like this, is that ripping out and replacing a supply chain system takes the customer years and years. Often 3 to 4 years for even a mid-sized business, as the system has to be rolled out across all distribution centers and stores. And this across all countries.

So this is really a costly, risky and hugely time-consuming project for a customer. Something which a company will only try to do if the reason is really compelling, such as moving to a modern cloud-based solution from a legacy system. As a result, once Manhattan is built-in at the customer, its pricing power is strong.

The classic case study in moving to a subscriptions-based business model is Adobe. The company started to move its customers to subscriptions as from 2013, which gave rise to high revenue growth once the initial negative impact of evaporating license sales started to wean off, and subscription revenues started accumulating:

And investors rewarded the company with a higher multiple as this story became clear. So this was a great opportunity last decade — the combination of high sales growth and expanding multiples resulted in a 10-plus-bagger for shareholders. Manhattan is now going through a similar business model transition.

This is Manhattan’s long time CEO, Eddie Chapel, detailing the company at the Baird conference last year:

“We went public in 1998. For the first 10 years, we were a one product and one country company. We provided warehouse management systems for large distribution centers. Then we grew into transportation management, order management, inventory management, retail point-of-sale and so on. Today, we're about a $1 billion revenue company. We started our cloud investments about 10 years ago. We have re-engineered everything, every piece of code from the ground up to have a truly versionless, cloud-native solution.

We focus now on Tier 1 and Tier 2 companies around the world, essentially companies that are $0.5 billion in revenues and up. We have some smaller customers than that, but that's sort of the sweet spot for us, and we have about 1,200 customers worldwide. Our biggest vertical is retail as that's where you have the most complexity in the supply chain. Also food service, grocery, third-party logistics, life sciences, automotive and high tech are verticals that we are dominant in.

About 80% of our revenue comes from the Americas and 20% from international. About half of our revenue comes from warehouse management systems, about 25% to 30% from order management, and 15% to 20% from transportation management and others. We have a 20% market share in warehouse management systems and a 2% market share in supply chain management overall, so there is plenty of runway in front of us. We have a 75% win rate against our top 6 competitors and in the 25% of deals that we don't win, 90% of the reason is that we lose on price. We are the premium solution in the market. Our aspirations are to grow at double-digit rates year-over-year with about 100 basis points a year of margin expansion.

With our cloud-native solution, we've managed to extend our lead. The biggest reason that our customers move to the cloud with us, is getting access to innovation. In the old world, you implemented a big software upgrade every 5-6 years. Today in a version-less world, we deliver brand-new capabilities every 90 days. In the distribution world, things are moving much faster these days, there's the need for robotics and automation, and just an overall changing profile in distribution operations. So we feel in terms of our moat in WMS, we had a pretty big one to start with and with our new technology and capabilities over the last 3 or 4 years, we benefited from our competitors not making that move. They haven't started yet or they're starting just now.

And frankly, I know how how big of a lift that is and how much effort it takes in time, money, resources and strategy. So given that we started our cloud journey 10 years ago and we're not going to stop innovating, I feel pretty good about keeping that market-leading position. We're taking market share for sure and we've done a pretty nice job of diversifying beyond retail. For example, we started doing business with Nike about 15 years ago and this was originally 100% wholesale. Now we even handle distribution for Exxon.

SAP is sunsetting their legacy warehouse management product and also their ERP, and that's definitely creating some tailwinds for us. The ECC to S/4 HANA move that SAP is encouraging people to make, is not a reinvention of technology. They're essentially taking their on-premise software and moving it into the cloud. What that signals to a customer is that there's no innovation. And that's fine for the core SAP solutions such as financials, order to cash and procurement. But in the areas around those core capabilities such as supply chain, customers are looking for innovation and we're the beneficiary.

One of our more recent customers, Schneider Electric, which is a wall-to-wall SAP customer, they're moving to S/4 HANA but they've broken away warehouse and transportation management. They awarded us a global contract to roll out across all their distribution centers and transportation systems. And they're not the only one. L'Oreal, Nike, Asda in the UK and so forth, all are going through that same ECC to S/4 HANA migration and are breaking away supply chain pieces. We've been the beneficiary and so I expect that to continue. Many customers of SAP's legacy WM have pretty rudimentary needs and they're satisfied with SAP's capabilities. They don't need innovation and so forth. But even if there are 10% of those that are going to be applicable to us, it would be highly advantageous.

Historically, cross-selling is about 25% of our new software revenue every quarter. In the old world of selling licensed software, it would not be unusual to go to a client who wants to buy WMS and we'd say ‘how about buying TMS as well?’ And then they would say only if it's a great deal. In the license world, you could make it a great deal and then they'd buy WMS and TMS together. In the cloud world, obviously you'd be paying subscriptions so nobody will start paying for something if they're not going to be using it for a while. Most companies find it hard to execute on more than one of these initiatives at a time. So as a consequence of that, they buy when they're ready.

We have a unified set of solutions with common technology, so as we see more customers move over to the cloud with only one product initially, the likelihood to cross-sell is higher going forward. We have an account team that are focused on existing customers, understanding their roadmap and working with them to build that roadmap. That's code for selling them more software. So we are building new products because there are market needs.

For example, point of sale, I've been asked at conferences whether we will ever get into point of sale and the answer was no. But retail stores have changed. They were the same for hundreds of years. You walked into the store, you picked up a product and you paid for it. So the retail store was a single function facility with a little cash register to consummate the transaction. That's all it did. Today, the store is a multi-function facility. It's a billboard for the digital business, it's also a customer service center, maybe you buy online and return in-store, and it's now also a mini distribution center. So you buy online and pick it up in store or curbside. All these many functions cannot be supported by a cash register. So the time was right for a next generation point of sale platform and that's why we launched.

The same is true for supply chain planning. Planning has been around for decades and these systems were batch-based. Retail companies would send a shipment to their stores once a week and then the CEO would get the sales report overnight on what happened yesterday. But today, order management systems are better and distribution centers are faster. We believe that the time is right to take planning systems from being batch-based, to continuously optimize what to buy and where to disposition inventory. And this is also a cloud-native solution that we built.”

So there are a number of interesting growth angles here. SAP has about 6,000 customers that use their supply chain solutions, with around 3,000 to 5,000 using their warehouse management. As SAP’s customers are now being transitioned over to the cloud, if Manhattan could win 10% of those making use of warehouse management, this could result in 300 to 500 additional customers. Manhattan currently has only 1,200 customers, so this could provide a substantial tailwind.

Secondly, cross-selling modules should be much more lucrative in the cloud. As Eddie mentioned, historically additional modules were frequently sold as part of one large license deal. And modules that the customer didn’t need initially were heavily discounted. In the cloud, you can switch on subscriptions for modules as you start using them, so customers will start paying full prices. And there is clear evidence that this is playing out, more on this later.

The third opportunity is the cross-selling of the company’s new products — point of sale and planning. Point of sale makes Manhattan’s cloud platform available to store employees so that they can check inventories, take payments and manage returns:

And you can access this solution on various end devices such as tablets, smartphones and even payment terminals:

Point of sale is a crowded space, but it’s a large TAM and could be attractive for Manhattan customers to make an advanced solution available to their store employees. Adyen is deploying a similar strategy with the goal of offering a unified system for customers to take both online and in-store payments.

We haven’t seen Manhattan’s new demand planning solution yet but the first customer will go live this year. There are a number of strong players in supply chain planning with cloud-based platforms such as o9, Kinaxis and Blue Yonder, but for some customers it could make sense to integrate forecasting inside of Manhattan’s platform.

Finally, Manhattan also closed a partnership with Shopify to integrate Manhattan’s order management system into Shopify’s platform. Given that Shopify is really a great business and which is now also increasingly being used by large enterprises, this could provide an interesting opening for Manhattan to gain new customers.

So Manhattan is obviously an interesting company — they are a leader in warehouse and order management software, and there is an obvious path for growth by gaining new customers, cross-selling additional software and by moving customers to higher value cloud subscriptions.

However, post the company’s recent results, shares sold off dramatically. With the shares now down 35%:

And with the company’s forward PE now back towards the bottom of its historical trading range:

For premium subscribers, we’ll give the insights from a consultant in the supply chain software space and from a large enterprise customer about what’s going on. We’ll dive into Manhattan’s problems and we’ll also flag one listed quality software name where these industry experts are currently seeing an acceleration in spending. Finally, we’ll analyze the financials and valuation of the key companies involved, and give our thoughts on which names should be good investments here.

Manhattan Associates' Problems & One Winner

Read more

In late ‘22, the market reckoned that Meta had gotten itself in a precarious situation. Revenue growth had evaporated while the company was investing heavily in the Metaverse pipedream. In addition, there was a clampdown on user tracking by Apple and its ATT policy, and also with Google looking to ban cookies at some stage, which raised doubts about Meta’s future ad monetization growth.

Meta’s business performance actually wasn’t that bad. The covid lockdowns had created a boom for online platforms, and a cyclical pullback was always likely when the world returned back to normal. This in combination with Apple banning user tracking by apps, honestly, it was a strong performance that Meta managed to hold its revenues flat in 2022. It looks like the successful roll-out of business messaging over WhatsApp and Messenger saved the day, or otherwise share prices would have headed even further south. The new business-messaging product was already at a $9 billion annual revenue run rate by Q3 of ‘22, when the shares hit their 7-year low.

Also Reels was generating over $3 billion in annualized revenues already, but these new TikTok-like short form videos were actually cannibalizing Facebook’s Feed and Instagram’s Stories. At that time, the latter two enjoyed much superior ad monetization, causing a $2 billion revenue headwind for the overall business. Meta first tends to scale a product up and make it popular, and then afterwards starts to think about how to maximize profits.

What especially spooked the market were Meta’s heavy investments in R&D, and especially the Metaverse. The latter continues to burn huge amounts of cash up to this day.

However, we can see that post the 2022 trough, Meta’s family of apps saw spectacular profit growth. The share price reaction was the same, with the shares going up 7-fold since the market completely misread this situation in late ‘22.

And all of these additional profits are coming from better monetization strategies, as Meta’s user growth remains modest. The other point to note here is that most of this user growth is coming from poorer regions, basically from countries where parts of the population are only now able to purchase their first smartphone. So while these boost user growth, the business value of these users is currently slim. Meta makes a lot of money by showing users ads of ecommerce products, and then users go to the advertised site within a few days to make the purchase. The advertizer then attributes this sale to the Meta ad they showed, and then decides to purchase more ads.

So while the investments in the Metaverse have been a pipedream so far, the investments in R&D - and especially AI - are paying off. The basic idea is to create a more engaging user feed which increases time spent on the various apps, thereby increasing ad revenues. For example, already in Q2 of ‘22, a single AI advancement in Facebook’s recommendation model led to a 15% watch time gain for Reels. The other revenue driver is to serve better ads with AI. The basic idea here is that more customized ads for a particular user increase ad engagement.

These advancements led to a large jump in Meta’s AI capacity already in ‘22, and with another massive jump coming this year. Obviously this is bullish for Broadcom and Nvidia.

For premium subscribers, we’ll go much deeper into Meta’s AI capabilities. We’ll also discuss Amazon’s vision for the future of AI. And finally, we’ll give our thoughts on whether it makes more sense to invest in the three big clouds — Amazon AWS, Microsoft Azure and Google Cloud Platform — or in the neo AI-clouds such as Nebius.

A former data scientist at Meta gives further details on Meta’s AI capabilities, this is an expert via Tegus:

“The start of really putting emphasis on GPUs started around 2019 with the first super compute cluster, RSC. That's the primary cluster that was used to train all our content recommendation models, ads models and computer vision. That built-out was fairly complete in late '21. For inference, everything was on CPUs. Back then, the models were never hitting a billion parameters. Even the ad models, which were pretty much the most complex models that we had at the time, those were still in the hundreds of millions of parameters.

Read more

So it looks like the Chinese AI startup DeepSeek has made two major breakthroughs. The first is really a series of better algorithms that dramatically lower the cost of training. Reading through the paper, the most important of these look to be DeepSeek’s Multi-Head Latent Attention (MLA), a finer-grained mixture-of-experts (MoE) architecture, Multi-Token Prediction (MTP) and running many operations in training in 8-bit floating points as opposed to everything in 32-bit.

The results are huge gains in efficiency and according to DeepSeek, one could train their model in less than 2.8 million H800 hours. The H800 is the Nvidia Hopper GPU which China used to be allowed to import. Assuming a cost to rent one of these GPUs at $2 per hour, this leads to estimated overall training cost of around $5.6 million:

We will only be able to tell whether these assertions are fully true once these algorithms have been replicated in training. There’s no doubt that US-based AI labs will try these out so we should know fairly soon. But having read the papers, at this stage we don’t see any reason to doubt the above table.

While the first innovation should lower AI compute costs, the second innovation should actually increase compute costs as it will boost the demand for inference-time, more on this in a bit. So this second innovation is in DeepSeek’s DeepThink Model, also called R1, which was trained with reinforcement learning on the base model above. The goal of this to provide DeepThink with chain-of-thought capabilities, similar to OpenAI’s o1 model. Basically the way this process worked was that DeepThink was given rewards during training, so that if it gave a correct answer or it made a correct step in its thinking process, the model could update its parameters based on the size of the reward. The goal of this process is to learn by trial-and-error and large amounts of iteration so that the model can become really good at its tasks.

Reinforcement learning is nothing new and was actually the hottest area in AI just a few years ago until LLMs came along. It’s actually the way how both AlphaGo and AlphaFold were trained. The innovation here is that this is the first time that this type of unsupervised reinforcement learning was applied to an LLM. Obviously this type of training works very well for exact sciences types of problems, and this is also the reason why DeepThink is so good in coding and math.

The impact for AI datacenter stocks & Nvidia

So we discussed two sets of innovations above. The first one with the algorithmic improvements for AI workloads is what really caused the sell off in AI stocks. The basic theory is that given these huge efficiency improvements combined with the currently massive buildout in AI data centers, once these improvements have been introduced in the models of OpenAI, Google and Meta; six or twelve months from now we’ll have a large amount of excess AI capacity. Causing orders for Nvidia GPUs and other components in the AI data center supply chain to be halted. Obviously the main stocks at risk from DeepSeek’s breakthroughs are those exposed to the buildout of AI data center capacity.

The counter argument of the bulls is that due to these cost improvements, AI demand increases disproportionately resulting in a higher need for AI compute. While in the long term we think that this argument is valid, in the near term i.e. the coming one or two years or so, this will depend on the price elasticity of demand. The way to think about this is if DeepSeek’s improvements reduce the need for compute by 8x in the coming year, but demand only increases 2x, we’re still down 4x in terms of AI capacity which we thought we would need.

So there are two key questions. The first is, assuming all of DeepSeek’s claims are indeed valid, how much of an efficiency improvement these innovations give? And the second is, given these cost improvements, how elastic is AI demand?

On the first question, Dario Amodei — the CEO of Anthropic — thinks that the cost improvement in training is about 4x:

“I think a fair statement is ‘DeepSeek produced a model close to the performance of US models 7-10 months older, for a good deal less cost (but not anywhere near the ratios people have suggested)’. If the historical trend of the cost curve decrease is ~4x per year, that means that in the ordinary course of business — in the normal trends of historical cost decreases like those that happened in 2023 and 2024 — we’d expect a model 3-4x cheaper than 3.5 Sonnet/GPT-4o around now.

Since DeepSeek-V3 is worse than those US frontier models — let’s say by ~2x on the scaling curve, which I think is quite generous to DeepSeek-V3 — that means it would be totally normal, totally ‘on trend’, if DeepSeek-V3 training cost ~8x less than the current US models developed a year ago. I’m not going to give a number but it’s clear from the previous bullet point that even if you take DeepSeek’s training cost at face value, they are on-trend at best and probably not even that. For example this is less steep than the original GPT-4 to Claude 3.5 Sonnet inference price differential (10x), and 3.5 Sonnet is a better model than GPT-4.”

We disagree that DeepSeek V3 is 2x worse than GPT-4 or Sonnet. Having tested the model over the last ten days, the answers are impressive. So it’s probably a good assumption that DeepSeek’s innovations reduce the need for AI compute with 4 to 8x, with the latter being more likely.

On the second question of how elastic the demand for AI compute will be, we can look at history. Efficiency improvements have been a constant in AI research and Amodei estimates they are about 2 to 4x per year, with the rate increasing towards 4x more recently. At the same time we can see that despite these algorithmic improvements, compute requirements of AI models have continued to increase:

Most of these increases in compute demand have been driven by scaling laws, i.e. the observation that larger AI models trained on larger amounts of data and for larger amounts of time tend to perform better. Scaling laws are still alive and well as several of the leading players in LLMs are talking about 1 and 2 million GPU training clusters. And these will be Blackwell clusters. For comparison, OpenAI’s ChatGPT 4 was trained as 25,000 A100s.

And even more encouraging is that a second scaling law has emerged, this time in inference. The scaling law here is that a larger model that reasons for larger amounts of time tends to comes up with better solutions. And actually, DeepSeek’s breakthroughs with its R1 DeepThink model will benefit the second scaling law. Now that LLMs can develop chain-of-thought reasoning capabilities with unsupervised reinforcement learning, this will result in both more models and much more impressive models with inference-time capabilities. Anecdotally, we never used OpenAI’s o1 as it felt slow and boring, while DeepSeek’s R1 felt much faster and interesting as it displayed how the model was reasoning. So when it comes to the second scaling law, DeepSeek should strongly drive compute demand here.

And demand for inference-time will see a further boost once we head into the world of AI agents. The basic idea here is that you allocate a certain task to an AI assistant, who subsequently organizes and delegates all the work to be carried out to various AI agents with their own expertise. An example here could be writing a software program — one agent defines the overall software architecture, another agent writes the functions, the database-expert agent defines the database scheme and the API expert writes the API.

Zuckerberg on Meta’s call had a similar line of reasoning. While DeepSeek’s innovations can drive down the cost of training, more compute will be needed for inference:

“We expect to bring online almost a gigawatt of capacity this year and we're currently building a 2-gigawatt and potentially bigger AI data center. We're planning to fund all of this by AI advances that increase revenue growth.

On DeepSeek, there's a number of novel things that we're still digesting and that we hope to implement in our systems. It's too early to really have a strong opinion on what this means for the trajectory around infrastructure and CapEx. There's already a debate around how much of the compute infrastructure is going towards pre-training versus inference as you get more of these reasoning models. It already seemed pretty likely that the largest pieces aren't necessarily going towards pre-training. But that doesn't mean that you need less compute because one of the new properties that's emerged is the ability to apply more compute at inference-time in order to generate a higher level of intelligence and a higher quality of service.

The field continues to move quickly and I continue to think that investing very heavily in CapEx and infrastructure is going to be a strategic advantage over time. At this point, I would bet that the ability to build out that kind of infrastructure is going to be a major advantage for both the quality of the service and being able to serve the scale that we want to.”

And also Microsoft’s Satya Nadella continues to see AI scaling laws in both training and inference:

“AI scaling laws continue to compound across both pre-training and inference-time compute. We ourselves have been seeing significant efficiency gains in both training and inference for years now. On inference, we have typically seen more than 2x price performance gain for every hardware generation and more than 10x for every model generation due to software optimizations. And as AI becomes more efficient and accessible, we will see exponentially more demand. Therefore, much as we have done with the commercial cloud, we are focused on continuously scaling our fleet globally and maintaining the right balance across training and inference as well as geo distribution.

Azure is the infrastructure layer for AI. We continue to expand our data center capacity in line with both near-term and long-term demand signals. We have more than doubled our overall data center capacity in the last 3 years, and we have added more capacity last year than any other year in our history. And with OpenAI's APIs exclusively running on Azure, customers can count on us to get access to the world's leading models. And OpenAI has a lot more coming soon, so stay tuned.”

So while the long term outlook for AI continues to be bullish, especially in scaling inference-time, the risk for investors is that due the magnitude of DeepSeek’s improvements, we could see an excess in AI capacity 6 to 12 months from now. Resulting in a temporary halt in orders for GPUs and other AI data center components. There’s little doubt that Nvidia and other names in the AI data center supply chain will underperform in that scenario.

However, looking at Nvidia’s valuation, the current forward PE of 30x is already pricing in an earnings correction. The way this works is that cyclical names tend to trade on bottom multiples when the market reckons that earnings are on a cyclical high. So when earnings subsequently go through a cyclical correction, you start to see multiple expansion as the market starts factoring in the next upcycle.

A recent example of that typical scenario is Texas Instruments. As the company’s earnings went into a downturn, the stock’s PE multiple expanded:

And this limited downside in the share price over the 2023 - 2024 semi contraction. So despite the cyclical nature of TI’s business, the shares held up much better.

So basically what we’re seeing here is that for quality companies with an attractive long-term growth outlook, investors are willing to hold the shares through the cycle, which reduces share price volatility. That being said, the downside for Nvidia would be if the anticipated earnings correction turns out to be uglier than expected, i.e. a dramatic decline in orders for a substantial period. However, due to the strong scaling laws we’re currently seeing, where especially those in inference-time have a very promising outlook, this risk should be limited.

Nvidia continues to have a long term attractive outlook. The long term CAGR in AI demand should be steep due to the breadth of applications where AI can be applied, the huge sizes of some of these markets, and scaling laws. Nvidia is a stock which has really been climbing a wall of worries over the last few years. Ever since the AI boom started and Nvidia gave its first huge growth outlook in May of ‘23, many were already screaming bubble. Since this share price pop, the shares have tripled again even taking into account the sell off of last week.

Investors can still make an attractive IRR in Nvidia. Putting fiscal year ‘27 consensus numbers from Wall Street on a 30x forward PE results in an IRR of 45%. And we know from a channel check with Amazon that indeed a large increase in orders is coming again this year, although the largest increase in orders will be for their own internally developed accelerators. Cloud customers tend to use Nvidia whereas the hyperscalers themselves are trying to move AI workloads to their own silicon.

We delved deeper into the market share battle between Nvidia and ASIC accelerators below. The hyperscalers are obviously huge clients, and who are extremely good at writing software stacks to abstract away the hardware layers underneath. These software stacks are becoming ever more advanced and with programming languages that resemble Nvidia’s CUDA, making it easier for AI engineers to get up to speed with them.

So Nvidia continues to have attractions here. It has basically locked in the enterprise AI market which should see a high CAGR for the foreseeable future. That being said, a lot of investors will have made substantial money in this name already, so it’s not unwise to lock in some profits and allocate these to other names with perhaps a better risk-reward.

In this article, we will tour a large variety of AI-exposed names which should get a boost from DeepSeek. Two obvious areas are companies developing AI apps and then the semis deployed at the edge to run those AI apps. Both of these should benefit due to the increases in AI demand coming from DeepSeek lowering the cost curve, as well as from the rise of chain-of-thought reasoning.

In the following chapters of this article, we’ll flag our favorite stock picks as well as well-known names which we would avoid at this stage. Finally, we’ll create an AI basket from our buy ideas. Whereas most analysts have been purely focused on AI-datacenter-capex related stocks, we’re taking a much broader approach by also investing in AI software, robotic applications and edge semis. Long term, the first two should turn out to be the largest investment opportunities given the huge sizes of these markets.

The most lucrative long term area, AI apps & robotics

Read more

“For version 13 of FSD, we expect to see roughly a further five or six-fold improvement in miles between interventions compared to 12.5. Looking at the year as a whole, the improvement in miles between interventions will be at least 3 orders of magnitude. So that's a very dramatic improvement in the course of the year and we expect that trend to continue next year. The current internal expectation for FSD having more miles between interventions than a human is for the second quarter of next year. It may end up being in the third quarter, but it seems extremely likely to be next year, and then to continue with rapid improvements thereafter. We also published our Q3 vehicle safety report, which shows one crash for every seven million miles of autopilot. That compares to the US average of one crash for roughly every 700,000 miles so FSD is currently showing a 10x safety improvement relative to the US average.” — Elon Musk

Both anecdotally and on Tesla’s own data, the company has been making huge improvements in their full-self-driving (FSD) autopilot, with a Tesla vehicle now being able to drive long distances and navigate complex situations without driver interventions. Altimeter made a nice chart summarizing the above stats and if we can get another one to two orders-of-magnitude improvement next year, i.e. 10 to 100x, FSD would surpass the capabilities of human drivers:

According to YipitData, also Waymo is now getting more gross bookings than Lyft in the San Francisco area where it operates:

Tech investors have enjoyed many secular tailwinds over the last few decades, resulting in tremendous performances for companies that are able to dominate these spaces. If you’re a monopoly or oligopoly player in a vast new market, you’ll do very well. A few examples of these new markets which emerged include search, social media, the public cloud, SaaS, AI accelerators and leading edge foundries. However, looking ahead, the future robotaxi market could easily dwarf all these previous vast new markets. Even capturing a small share of the number of miles being driven on a daily basis would propel the winners to become the largest market caps in the world.

Two of the best positioned names in automated driving are Tesla and Waymo (Google). In this article we’ll dive into the underlying tech by speaking with a current engineer at Tesla who was previously at Waymo. We’ll also review the outlook for both companies, calculate the financial opportunity from this new market and discuss the current financials and valuation for both names. Obviously Tesla shares have been on a tear over the last six months and we’ll discuss whether the valuation still makes sense here, or whether it’s time to start locking in profits..

Insider insights

Read more

The above world to connect corporations (the advertisers) with various media (the publishers) was fairly simple with only one type of player in the middle, i.e. the agencies. The only other option available was to go direct, where for example a small company would contact a local newspaper to place its ad. Or perhaps a large corporation such as Louis Vuitton would know exactly which publisher they wanted to use, and so they could contact for example Vogue to pre-buy a certain amount of ad inventories for the coming year.

In the internet era, this landscape in the middle to connect advertisers and publishers started to fragment considerably, with a wide variety of players emerging. The largest and simplest new type of player are the ad networks, where dominant internet platforms such as Google, Meta and TikTok make their own inventory available for advertisers. Cost-per-Mille (CPM) and Cost-per-Click (CPC) are two of the most commonly used pricing models, where the ad network charges the advertiser a price per one-thousand impressions in the case of CPM or per user-click in the case of CPC. The number of exact clicks generated will depend on the quality of the ad, so the ad network will provide the advertiser with an estimate beforehand. Another factor to boost user-clicks is better audience segmentation, so if you know exactly which audience to target then the click-through-rates (CTRs) will be higher.

CPMs are also dependent on the media format. Ads on Google are priced higher as the platform already know the search intent of the consumer for example to buy a sofa. Similarly, ads on Youtube have best-in-class CPMs as well — Google has tons of data on what a viewer is interested in while there is also no opportunity to quickly scroll past the ad like you can do on X or Instagram, without giving it much if any attention. CPMs also depend on the quality of each publisher, so for example Instagram will have much higher CPMs than a secondary competitor.

The rise of digital ads has been one of the most interesting growth stories for investors over the last few decades, and there is still further room for growth as traditional media such as linear TV will gradually disappear over time.

However, given the large share of digital media already, it’s also clear that growth will continue to slow down. This can easily be seen in the annual growth rates of both Google Search and Meta, which have gradually been trending down over time — with the exception of the online boom during the covid lockdowns..

For premium subscribers, we’ll dive deep into the other parts of the value chain for programmatic ads, and we’ll flag companies which are strongly positioned in their niches. We’ll also give the insights on the industry from a former manager at The Trade Desk, an expert available on the Tegus network, and we’ll discuss which stocks currently offer the best risk-reward for investors. Finally, we’ll update our thoughts on Applovin, which was our first investment idea we wrote about at the start of this year and is since up 10-fold.

Read more

Currently there is a strong drive to move to modern solutions. Legacy apps in the space offer a non-live view of the supply chain, where data is transferred with batches between the various point solutions which creates significant time lags. This is supplemented by workflows in Excel spreadsheets, creating a Byzantine web of data.

Modern SaaS solutions can offer a live and unified view of the entire supply chain. This brings in a substantial ROI — inventories can be optimized, products get better geographically allocated based on demand, and sourcing risk can be reduced by ordering from a variety of manufacturing sites. Going forward, all these benefits will only increase with the rise of a smarter AI.

The demand for AI will likely speed up the adoption of state of the art supply chain solutions. A consultant for supply chain operations mentioned that legacy apps are struggling to implement AI into their platforms, while it brings in compelling advantages for customers. Such as most notably increasing levels of automation and smarter insights. So his view is that new cloud-based solutions will be the winners in this market.

An executive at Bristol-Myers Squibb, the global pharmaceutical giant, explains the challenges that multinational enterprises have to deal with in the management of their supply chain:

“For new products, it's typically unknown how much you need to make and how much inventory you need. You're always at risk of either producing too much or not enough. It's also the distribution of your goods, if you've got a global launch for your product, the biggest issue is how much are you going to have located in different countries and different regions. Because when you have your product label for one market ready, no matter what the issues are, you're not allowed to distribute a regulatory product to another country. So we have to follow local regulations and it could be a pretty big issue depending on the product and the volumes in countries.

Then you always have to plan for unforeseen events that you have no control over. We need to be able to supply products to customers and patients with a high degree of reliability. There's an aspect of reputational cost, some of those products are critical in nature and so if you have issues to supply to a certain market, people there are dependent on these products.

Having dual sourcing is crucial, which means that we don't tend to keep all our product manufacturing at a single site. As an example, you recently heard about the Pfizer facility in North Carolina that was impacted by a tornado that took out the whole site, and they were manufacturing there some critical sterile products. So we have to plan around those things and make sure that we have a suitable supply in place, as well as the proper levels of network capacity and inventory buffers.”

The company which we’re going to look at is a play on the above themes. So this is a leading and real-time software in the supply chain management space, that is especially strong in its ability to connect to a wide variety of enterprise systems and extract data on a live basis. The company has been investing heavily to make AI a central part of their platform, both to automate workflows and derive better insights and recommendations. Additionally, large system integrators are investing into the platform as well, which is usually a bullish indicator for future revenue growth given their wide relationships with the enterprise world.

As a result, analysts’ forward EPS projections have been strongly trending upwards — the green line in the chart below — while the share price hasn’t been responding with the PE multiple derating to historical lows:

Taking reasonable assumptions, we get to a 25% IRR (annualized return) over the coming five years. To get a deep understanding of this space, we’ll also talk to a number of executives from blue chip companies on how they’re using these solutions to operate their supply chains and we flag one name to be on the lookout for when it IPOs..

Read more

Software has been an excellent space for investing over the last decade, with the space strongly outperforming both the S&P 500 and Russell 2000 indices:

This isn’t a surprise. The software industry frequently exhibits high switching costs for users in combination with network effects, resulting in strong pricing power for the typically few number of providers that dominate a niche. As customers are running their operations on a particular software, the software firm can implement a ‘land and expand’ strategy at the customer by creating more modules on the software platform. Customers can then purchase these to get access to the additional software functionality and then run ever more workloads and operations over the platform. This is an obvious flywheel for revenue growth for software companies.

And even when a particular software architecture gets disrupted over time, such as the mainframe or client-server models, these companies continue to print cash as share losses are only gradual while these also can be offset with price increases to a certain extent, depending on the particular situation. On top of that, many of the software leaders of the client-server era have successfully transitioned to the cloud by also offering SaaS versions of their platform.

A negative is that these attractive characteristics of the industry have increasingly gained the appreciating of investors. Ten years ago, you could frequently pick up quality software businesses for around 15 to 30 times forward earnings or so, but now these valuation ranges have moved up for most of them. Good quality software is now usually trading well above 30x.

A few good examples here are Cadence and Synopsys, two software businesses that dominate the market for semiconductor design. While an acceleration in top line growth gives scope for further multiple appreciation, there is also a risk that if growth slows down we’ll see multiple compression over time.

This is actually not a phenomenon unique to software. As the commodity supercycle collapsed around ten years ago, driven by the shale revolution in the energy markets and the slowdown of the Chinese economy more generally, investors woke up again that commodity markets are highly cyclical. Additionally, as the Sino-American geopolitical rivalry took off, and combined with the generally low interest rate environment, Western investors started looking for US-based quality assets that can bring in a safe and predictable growth rate. Two examples here are Costco and Moody’s:

Today we’ll be looking at a software name that is dominant in its space, where currently cyclical headwinds have been slowing down growth, but that has a number of interesting opportunities ahead of itself..

.. while its forward PE valuation is again near the bottom of its trading range:

Which makes the cyclical slowdown in growth well reflected in the share price:

So the investment case is fairly simple — as the cycle picks up again, we should see an acceleration in top line and EPS growth, and combined with multiple re-rating. Additionally, as the company has been expanding successfully into new fields, there are substantial revenue opportunities from new verticals.

A classic compounder at the bottom of the cycle

Read more

Tom Shea, OneStream’s founder, is a typical serial software entrepreneur. With each new start-up he founded, he iterated on the learnings from the previous ones, improving the software’s design and architecture. The first product was Magma, developed in the late 1980s, which was an add-in for Lotus 1-2-3, the Excel-like killer app for MS-DOS at the time. Magma allowed users to consolidate financial data for a large company owning a variety of subsidiaries, and use this output for budgeting, financial reporting and analytical purposes. At this stage, Tom Shea was still working as a financial analyst, daydreaming of becoming one day a CFO at a blue chip name. So Magma was more of a side project, automating his own workflow while allowing him to make some money on the side by selling the software.

During the nineties, Hyperion became the dominant enterprise software for these types of financial consolidation and reporting tasks, named after one of the Greek titans. Shea became an avid Hyperion user. While the tool was extremely strong at its core tasks, Shea noticed that many organizations were struggling to get all their data into Hyperion. The problem was that the data had to be sourced from a wide variety of software systems such as ERP tools, accounting software and spreadsheets, and this from a wide variety of subsidiaries. In-house developed processes were usually manual, error prone and time consuming, involving in the worst case a ton of spreadsheets and in the best case a number of outdated software tools. This area would form the target market for the second software tool developed by Shea.

At the start of the millenium, Shea founded Upstream so that enterprises could automate this cumbersome work. Upstream would extract the data from an enterprise various systems and computer files, automatically transform it where needed, subsequently validate it, to finally upload the data into Hyperion. As from now, data became available in real-time for Hyperion users. Upstream was so successful that Hyperion acquired Shea’s business in 2006. One year later, Hyperion itself was acquired by Larry Ellison for $3.3 billion, so like many other enterprise systems, it finally ended up somewhere inside of Oracle.

This is Tom Shea looking back at these years:

“From 2001 to 2006, life was a blur. UpStream was growing 90 to 100 percent a year, and we were building integrations to all of Hyperion’s products and some competing products, such as Comshare and OutlookSoft. In 2003, we settled in and became an exclusive partner to Hyperion. The UpStream engineering team and I started working a lot closer with Bob Powers and Hyperion’s engineering team at this time. We collaborated on tighter engine integration, product UI consistency, performance, and overall technology advancement.

Bob and I were forming a mutual respect for each other’s engineering skills and we liked collaborating. In 2006, Hyperion came to me and said ‘We have been dating long enough, we would like to purchase UpStream and make you part of the family.’ After careful consideration, we felt this was the best outcome for all stakeholders since the company was really built for and targeted towards Hyperion’s products and community. UpStream became Hyperion FDM in mid-2006, and we started to really integrate UpStream’s engine into the Hyperion product suite. Bob and I began to work more closely together, and we often talked about what a next-generation CPM product could be.

In 2007, acquisitions became prevalent in the CPM space, and in a matter of months, all the pillar companies were scooped up by larger enterprise software companies. In mid-2007, Oracle acquired Hyperion, then quickly after that SAP acquired Business Objects and IBM acquired Cognos. Just like that, the standalone CPM market was gone. At this time, Bob and I both stepped away from the CPM market. Bob went on to become the CTO of a SaaS company, and I took some time off so that I could spend more time with my family.”

The Shea - Powers relationship is important as both engineers would develop some years later the core of OneStream together. Shea would take on the role of CEO at OneStream with Powers as the CTO, while all the hard coding work would be split between them. This is Shea again reflecting on this period:

“In 2010, OneStream was born. It started with a call to Bob, where we discussed the potential to create a unified platform that encompassed the capabilities of all the products that we had seen in the evolving CPM space. This space had expanded over the years to go way beyond Financial Planning and Reporting to now include solutions such as Account Reconciliations, Financial Close Management, and many other domain-centric products for the office of finance.

I remember being on the phone with Bob and pacing around my front yard when we came to the conclusion that we would embark on a software architecture and coding journey that would be the equivalent of two people trying to eat an elephant for dinner. We discussed the idea that we needed to build a platform that could basically replace the 15 or so different CPM products with a single unified platform. The name OneStream comes from the idea that these 15 products needed to become one platform with infinite extensibility. This started another blurry phase in my life as both Bob and I went into coding overdrive. We literally divided up the engines that needed to be coded, and just set out coding until we had the foundation of a product.

In 2011, we began to have some core CPM capabilities in place, and it was time to get other key people involved to formally launch the product. This was a massive product that would require a lot of testing and validation before we could sell it to a customer. The 2010 to 2012 years were really focused on R&D and technical validation, the engines had to be solid and the data had to be right. We managed to get our first customer, Federal Mogul, in 2012 and began the long process of building an operational company. Each time we added a key team member, they led us to other key people, and the list of rock stars grew. The one thing that I have always known is that if we created a great product and we could attract the best people in the CPM industry, there would be no stopping us.”

A unique feature among software start-ups is that OneStream had been profitable since its first customer was signed in 2012. Basically the company sold licenses at the time so that they would receive the cash upfront and as such, there was no need for venture backing. Over the years, Shea turned down several offers from outside investors. But as the desire started to grow to turn OneStream into a $100 million revenue company, Shea finally accepted an investment of $500 million from KKR in 2018. This is Shea commenting on the KKR deal in Crain’s Detroit Business News:

“The thesis for bringing on KKR wasn’t just capital, but raising our profile too. News of this deal helped us land a CFO in November, Bill Koefoed, with a background at Microsoft and other big tech firms. KKR brought us a level of excitement that has made hiring a lot easier, we’re scaling the business. Employment is at 420 now, up from 250 when the deal was announced. And the KKR funding will help the company expand into other geographies and hopefully help land big government contracts. David Petraeus, a retired U.S. Army general and former director of the Central Intelligence Agency, joined the board after the deal. We’re not disclosing revenue at this moment, but I won’t dispute a published report that the company had revenue of $88 million last year. And as of mid-year, we’ve had in excess of 60 percent growth. An IPO is the eventual goal, but we have to do the work and scale the business first.”

OneStream is among the many software start-ups that followed in Salesforce’s footsteps to modernize traditional client-server based and on-premises computing infrastructure. As an additional abstraction layer, enterprises would from now on be able to automatically scale software systems in the cloud, a service managed for them by the software provider. As such, this new business model would be called Software-as-a-Service (SaaS). Whereas Salesforce is the key SaaS tool for the sales office, Workday is so for HR and ServiceNow for IT. The initial goal of Onestream was to become the equivalent for the finance office. However, the platform is now much more than that, more like a Datadog for the enterprise, which we’ll dive into below.

For premium subscribers, we’ll review in-depth:

• OneStream’s platform and the company’s runway for growth.

• The competitive dynamics in the industry, and the variety of software tools being used in the finance office and for enterprise observability.

• Insights from a variety of directors at $30 billion-plus blue chip companies on how they’re using OneStream and other competing point solutions, with the strengths and weaknesses of each. We’ll also include the views on the space from a director at Accenture, the leading systems integrator for enterprises. Each of these industry experts is available via the Tegus network, with a single expert typically charging $500 to $1,000 for a one hour call, so the price of a Tech Investments subscription is obviously a really great deal to get access to these valuable insights.

• The opportunities and risks from AI for enterprise software in general, and OneStream in particular.

• A detailed analysis of OneStream’s current financials and valuation, with thoughts on whether it makes sense to invest here.

OneStream’s platform & runway for growth

Read more

Advanced Driver Assistance Systems (ADAS) are likely to be be one of the highest growth areas in tech in the coming decade. Currently penetration rates are still low, but due to the strong progress of Tesla’s FSD, where the vehicle is now increasingly able to drive long and complex routes with zero or minimal human interventions, other large OEMs are starting to feel the heat. Even BYD, which historically was dismissive of this technology, is now introducing autopilots based on Nvidia’s hardware and Baidu’s mapping.

The autopilot industry exhibits attractive characteristics for investors. Due to the need for high quality and vast AI training datasets, the necessary engineering talent of which there is shortage, and large R&D budgets, this industry is likely to consolidate with only a limited number of strong competitors over time. The advanced players in self-driving tech are currently implementing two different strategies, with the first one being to offer an integrated hardware and software solution. Obviously Tesla has a very impressive product in this category, and Elon has mentioned they’re in discussion with one other OEM who is potentially interested in adopting Tesla’s FSD. The other strong player in the West with this strategy is Mobileye, who mentions they’re in discussion with 14 OEMS, which represent 52% of automotive production. And they have some impressive wins, most notably with the Volkswagen Group who will introduce Mobileye’s next-generation autopilots in its premium brands, which include Porsche and Audi.

The second strategy being pursued by leading players in this space, is selling the bare AI hardware so that automotive OEMs can train their own algorithms to drive their vehicles. The two key players here are Nvidia and Qualcomm. Nvidia goes one step further as they have developed an entire virtual world that is simulating a large variety of real-life traffic conditions, where OEMs can train their neural nets in. The company made a very lucrative deal with Mercedes, where they sell the Nvidia chip at a bargain, but subsequently take a 50% revenue share of Mercedes’ Drive Pilot. With a subscription starting at $2,500 per year.

Mobileye should be an interesting name for investors who reckon that a large number of automotive OEMs won’t have the necessary tech talent, know-how and data to develop an advanced self-driving product. This means that they’ll have to rely on a specialized and scale player to source this system from.

For its next generation autopilots, Mobileye has developed a hands-off system called SuperVision, an eyes-off system called Chauffeur, and a robotaxi system called Drive. These systems can make use of surround cameras, radars and one front lidar:

The main idea is that the vehicle can leverage the radar and lidar input when visibility is poor, such as under heavy rain, fog or at night. Radar is also excellent to detect the speed and velocity of surrounding vehicles or other objects, giving the autonomous vehicle a complete dataset of its surroundings. Mobileye has also been working on turning these radar and lidar based inputs into a second independent decision making neural net, so that the vehicle can continuously take the option with the highest confidence score out of the two input models, either the camera based one or the radar-lidar based system.

Another obvious use case of lidar is for long range detection, with in the below example a pallet being detected on the highway at 237 meters from the vehicle. So already the vehicle can start to prepare to switch lanes:

The final input that Mobileye’s AI relies on is its HD mapping. The benefit is that the vehicle already knows at any geolocation what the typical course of action is. For instance, when the vehicle arrives at a complex intersection, the AI already knows the path it’s going to follow. So it only needs to check if other objects might get in the way of this typical path, such as with temporary construction works, and it is only then that the system would have to come up with a creative solution, such as driving around the construction site. Another example is when the car arrives at a complex set of traffic lights. The map will provide context which light belongs to which lane and which light is for the pedestrians:

Overall, the attraction of the HD map is that the vehicle will have to figure out much less about its surroundings, as it already has access to large amounts of situational data. This reduces the risk of making a crucial mistake.

This is Mobileye’s CEO on the recent call detailing the performance of the new SuperVision system:

“When we look at the mean time between critical events, we did a test with one of our partner OEMs of our SuperVision vehicle in the U.S., where we have very good coverage of our REM maps, and the mean time between critical events is about 50 hours. This is about 5x better than the latest FSD version 12.5, according to the small amount of data that has been released so far. In China, it's much less than that because we haven't yet finished the localization, the REM coverage is not as good. And these are with the old EyeQ5 technology, imagine what we will do when we'll start launching EyeQ6. EyeQ5 does not yet have generative AI and with the EyeQ6, we have lots of generative AI components. All of our online and offline testing so far shows that we'll get a two orders of magnitude improvement.”

EyeQ6 will start shipping in the first half of next year and pack 8 CPU cores, one GPU, and four more ASIC accelerators for AI processing:

Mobileye has access to large amounts of mapping data as it launched this REM product in 2018, with EyeQ chips collecting data from their surroundings. In 2022 alone, these EyeQ chips were integrated in 233 model launches from a large variety of OEMs, as part of a base ADAS system:

Mobileye mentioned that they currently have 250 petabytes of labeled video data, which is even more than Tesla. As AI training is all about data quantity and quality, clearly both these companies should be amongst the best positioned to develop self driving systems.

Current EyeQ chips are largely being sold as part of the Mobileye’s base ADAS system, which is still making up most of its revenues. Next generation hands-off and then later on eyes-off systems will be sold at much higher ASPs. For example, a legacy ADAS is being sold for $50 per system while hands-off SuperVision is priced at around $1,500 per system and Chauffeur at $2,500. So going forward, Mobileye is well exposed to a strong revenue growth story with premium automotive vehicles upgrading to advanced self-driving systems such as SuperVision and Chauffeur. The company has also developed a mid-range autopilot, which will give an ASP lift of 4x compared to base ADAS. This latter system basically brings SuperVision capabilities to highways only.

We’re only at the beginning of this story really. So far, only a number of brands and models from Chinese OEMs are shipping with SuperVision. And then we have Volkswagen Group which will start shipping SuperVision, Chauffeur and Drive based products in the coming years. Mobileye’s CFO mentioned that these three advanced systems will start to scale significantly as from mid-2026.

An overview of the company’s design wins so far at the start of this year:

Mobileye mentioned on their latest call that they currently have advanced product wins or are in advanced discussions with 14 OEMs, representing approximately 52% of industry production. This is the company’s VP of Product Development detailing what these processes look like, and the outlook of getting deals done:

“The process of nominating this system is relatively complicated and involves many different activities starting from technically evaluating the system, building prototypes, doing global expeditions and performance evaluations, reviewing the architecture changes that are needed by the OEM et cetera. In addition, there is obviously the commercial part that needs to be negotiated to get to the right price point, and the investments needed in order to build this program, especially when we talk about incumbent OEMs, that have dozens or even hundreds of different vehicle models.

So the fact that it takes a few months or a year, in some cases more than years, it's natural given the magnitude of these decisions. At the same time, once these decisions have been made, it's very likely that the selected solutions will continue to be the backbone for the OEM for many years to come. We believe that we are now in the process of acquiring these OEM partners for our next-generation platforms. Once we do this in the next few months, we believe it will give us stability for years to come afterwards in terms of locking ourselves in as the technology provider for a very good portion of the top 10 OEMs in the world today.”

SuperVision has already been introduced in premium vehicles in China, and the company is getting good feedback. This is the IR on a recent survey they conducted with users:

“In a survey of about 600 Zeekr users, 85% said that they were either very satisfied or satisfied with the system, and 60% said they'd be willing to pay $3,000 to continue to have the functionality. Now, the reality is that no OEM is really charging for this in China because it's become part of the price war. So if you are trying to sell cars in China in the $45,000 and above pricing segment, it's very difficult to sell cars if you don't have a hands-free point-to-point pilot where you plug in the address and the car drives for you. There's 5 or 6 of these types of systems on the road in China.”

Overall, Mobileye should be a pretty straightforward investment case: the company has a strong market position for its existing products, there is a strong growth path ahead as ASPs of advanced autopilots will massively increase, while the market has been trading the shares based on short term volume expectations with the shares down 75%:

The key to unlocking strong growth in the coming 3 to 5 years ahead will be announcing further SuperVision and Chauffeur wins with large OEMs. Or Drive wins with ride hailing companies such as Uber, Grab or Didi. So far, the company has already won some Drive deals for shuttle bus operations.

While Mobileye paints a rosy picture for deals to come through, in this article we’ll go through findings from inside the industry to verify whether this is really the case. We’ll also discuss the current cycle, strengths and weaknesses of the various competitors such as Nvidia, Qualcomm and Tesla; the potential for the robotaxi Drive system, we’ll construct a detailed 2030 P&L with expected IRRs, and review Mobileye’s current financials and valuation, with concluding thoughts on whether the shares provide attractive risk-reward here.

Inside the automotive industry

Read more

Data platforms have seen enormous evolutions over the last decades. Traditionally both databases and data warehouses were dominated by SQL server type products such as those from Oracle, IBM and Microsoft. However, with the explosion of data and the internet, these SQL servers weren’t able to scale sufficiently and deal with all this new and unstructured data. New generation data warehouses that could scale horizontally — by spreading the data over servers were introduced — such as Hadoop for on-premise workloads and Redshift in the AWS cloud. With these being first generation solutions post the rise of the internet, these brought with them substantial problems as well. Hadoop was extremely difficult to manage while with Redshift, computing and storage resources were coupled together on servers in the cloud. So if you were only running your analytics workloads for three hours in a day, you still would have to pay for the entire 24 hours as you were making use of compute resources on these servers.

Enter Snowflake. The first data warehousing system that decoupled storage and compute. Storage would happen in cheap and basic object stores in the cloud, offering unlimited scaling of data storage, while compute servers would only be spun up when a customer was running his business analytics workloads. A truly elastic and pay-as-you-go system, saving enterprises substantial dollars.

At around the same time, the data science world was seeing tremendous amounts of innovation, especially in the field of machine learning. To handle large amounts of data manipulations, Spark compute clusters were introduced. These could be deployed in the cloud making use of a simple notebook containing the developer’s coding from one of the popular data science languages such as Python. This is really the core of Databricks, being the originator of Spark.

So originally both Snowflake and Databricks had very different use cases and both companies even worked together, as clients tended to run them for different workloads. However, both companies now have the vision to become the full fledged data platform in the cloud. The reason is simple, data gravity. It is both costly and lengthy to move large amounts of data, and so the workloads come to where the data resides, not vice versa.

Obviously this is a hugely attractive business, as once an enterprise has consolidated all their crucial data on your platform, moving away from it would be a yearslong endeavour that would be a serious distraction and risk for the business in terms of costs and time invested. So these types of businesses tend to be extremely sticky enjoying strong pricing power. However, to make the above vision become a reality, Snowflake first needs to up its game..

Snowflake 2.0

The problem Snowflake has it that they have to come from behind in terms of data science capabilities. This is currently the strongest growth area in terms of cloud data workloads, and is really where Databricks excels. While Snowflake can deal with unstructured data, the whole platform was really designed around SQL, traditional business analytics and dashboards. Whereas Databricks is designed around coding, data science and data processing. Snowflake already has the data storage layer, so now they’re introducing this second layer which can handle these types of Databricks tasks.

This is Snowflake’s new CEO on the recent call providing an overview of their current priorities:

“We started out as a data warehouse, but increasingly, we are the data backplane that provides a single unified view. Obviously, this data comes from production systems, such as Salesforce and SAP, or any of the other applications that you use. And additional capabilities like machine learning and AI are able to act on that data to then drive operational systems. Our overall strategy is to make sure that all of the data workloads that a company has are satisfied by Snowflake. This is where data engineering has been an investment for us, and Iceberg becomes key because the universe of data that can be acted upon by Snowflake, goes through a large expansion as not all data needs to be ingested into Snowflake before things happen. Our bet is really that AI and machine learning are going to go where the data is. Data is going to have a strong gravity.

I've spent a lot of time on the road. Just this quarter, I've met with over 100 of our biggest customers one on one. And the thing that clearly stands out from like a core data capability, is that in data sharing we are bar none. When it comes to newer things like AI, we've been open about the fact that we were a little behind early last year. But even before my coming, the management team here recognized the opportunity and invested heavily in it. And I would say the change over the previous quarter is that we can tell our customers with confidence that our AI products are world class.

In the core analytics space, we are the best in the world that there is. Especially when people consider migrating from complex on-prem systems, we have a professional services team that is exceptionally skilled at it and a very large ecosystem of partners that have been battle tested with massive migrations. And we have done migrations from on-prem workloads that end up saving something like 60% of the cost the customer has to bear, and their implementations end up being very efficient and low maintenance. These kinds of data migrations from legacy systems remain an important part of both new customer acquisition and are also driving substantial consumption increases in existing customers.”

So Snowflake’s strategy is to become the data platform in the cloud for all workloads. Not only a data warehouse for structured data in the cloud making use of SQL, but a data lake with capabilities to store unstructured data..

.. and leverage this data to run Spark-like clusters for machine learning with the popular data science languages such as Python and Scala:

Needless to say, these new workloads are strongly expanding Snowflake’s TAM, and the company sees their TAM growing at a 18% CAGR over the coming five years. And as customers are still migrating from on-prem and first generation cloud data warehouses, this is a company which could be looking at a high growth rate this decade.

But first Snowflake needs to step up its investments in innovation. The first obvious area to invest is an ability for data scientists to deploy code in the cloud on compute clusters akin to Databricks so that AI workloads can be handled on the platform. Snowflake is clearly behind here, with this feature only now being in public preview. This is Snowflake’s CEO on this new module:

“Our Notebooks offering is also seeing great traction in public preview with more than 1,600 accounts using that feature. This is critical to engage with data scientists and will unlock new opportunities that we previously did not address. We are in the early innings of this opportunity, and we'll continue to bring new features to market.”

The company is already further along in deploying LLMs as a copilot and for customers to train their own LLMs on their data. This Snowflake’s Chief Product Officer on integrating AI on Snowflake:

“Cortex, which represents the different language models that are available in Snowflake, the adoption is quite strong. And we have lots of new use cases, such as text summarization and text sentiment analysis. We introduced at Summit, both Cortex Analyst and Cortex Search as a way to enable users across organizations to be able to chat and interrogate the data, whether it's structured or unstructured. We have quite a bit of demand for going to general availability. And maybe the last one that I would call out is the Snowflake Copilot, we see a lot of usage on customers getting assistance on how to write better SQL queries, which also drives consumption back into Snowflake.”

Data can now also be stored outside of Snowflake in open source Iceberg tables. This should be a growth driver for Snowflake as potential customers were reluctant to bring all their data into the platform and effectively be locked in. With Iceberg tables, data can be stored outside the platform so that Snowflake is only leveraged when compute is needed, which is where the company makes the bulk of its revenues anyways (90%). And it gives the customer peace of mind that this same data can still be deployed over other platforms as needed. So while Iceberg reduces customer lock-in, net-net this should be a growth driver as it increases usage of Snowflake’s compute platform.

This is Snowflake’s CEO on Iceberg:

“Iceberg is enabling us to play offense and address a larger data footprint. Many of our largest customers have indicated they will now leverage Snowflake for more of their workloads as a result of this functionality. More than 400 accounts are using Iceberg as of the end of Q2.”

Going through comments from data engineers at various companies provided via Tegus, commentary on Iceberg was universally positive and most saw the format as being ahead of Databricks’ Delta tables format due to its higher flexibility.

The final interesting innovation by Snowflake is the launch of container services within the platform. This allows customers to deploy code and apps on a lightweight virtual machine within the Snowflake platform and make use of the data on the platform. This is the company’s CEO on this topic:

“Especially after the incidence of the summer with CrowdStrike, one of the hot topics has been how we set up applications with Snowflake. This is now a battle-tested operation, and most customers are shocked to find out that we can run a full replica of an important deployment that's something like 15% the cost of the original deployment, because the replica is not running any of the workloads that are going on in the main one.”

So Snowflake is innovating, and it is crucial that these new capabilities become a success, as otherwise the company might hit a brick wall. At the same time, the company has to face a number of other challenges..

For premium subscribers, we’ll dive into:

• The current competition between Snowflake and Databricks, there will be unique insights and data which were collected via data engineers, executives, system integrators and consultants, which were made available via the Tegus platform.

• A similar exercise for Google BigQuery and Microsoft Fabric, with a conclusion of which one is at a competitive advantage and which one has a competitive disadvantage. And for which platform these will pose the greatest threat.

• Competition from open-source solutions where enterprises can shift data workloads to, and to what extent these can substitute for Snowflake. Again we’ll draw insights from industry insiders which were made available via Tegus.

• How consultants and system integrators are currently expanding resources to support integrations of the various cloud data platforms, which is an indicator of future revenue growth.

• A detailed analysis of the Snowflake’s financials and valuation, including thoughts and highlights from Q2, and whether valuation is sufficiently attractive here to take a position.

Snowflake vs Databricks, inside the current battle

Read more

The computer-aided-design (CAD) software space consists of a series of highly consolidated industries where a small number of software vendors dominate their verticals. For example, Autodesk is extremely strong in architecture and construction, Dassault Systemes is used to design industrial-type products such as cars and planes, Aveva and Hexagon are the specialists in the petrochemical supply chain, and Bentley Systems dominates large horizontal infrastructure projects such as utilities, roads, bridges, grids and train networks.

Below is a former manager at Bentley describing the CAD space. This industry expert was made available via Tegus and for institutional readers, you can book a further call with him via the platform.

“If you look at projects today, they use a best-of-breed type solution, where you might use Autodesk tools to do one thing, Hexagon tools for something else, and then Bentley tools for another task. Autodesk is dominant in the vertical building market. If you look at what we call heavy civils i.e. roads, rails and bridges, Bentley is dominant in that market. If you look at the plant or process market, that’s split into two. So if you are doing onshore development, most likely it's going to be an Intergraph (Hexagon) and if you're doing offshore, it's probably more AVEVA-based (owned by Schneider Electric). Bentley is there as well, but in a third place. Lastly, for geospatial type of applications such as pipelines and transmission distribution lines, ESRI has owned that space for the longest time.”

This is Bentley’s CEO Greg Bentley giving some background on the company and their key competitor:

“The company history started with five Bentley brothers, the other four are proper engineers. And we began 40 years ago focusing especially on software for civil engineering and the modeling of horizontal networks of infrastructure. We went on to acquire the simulation engines that are used with the modeling of such infrastructure engineering, and then moved into ProjectWise, our collaboration environment. Our principal competitor is Autodesk, which is concentrated in vertical infrastructure buildings, such as commercial and facilities, but that's only 5% of our business. The rest of infrastructure is horizontal, i.e. networks of roads, rail, grid and water, and we are most comprehensive in what we do.”

For premium subscribers, we’ll dive into:

• The Bentley Systems company and its outlook.

• A large opportunity for the company that is currently emerging, it is not AI, although this can provide a tailwind as well with further software automation.

• Bentley’s competitive advantage.

• And a detailed analysis of the company’s financials and valuation, with thoughts on whether it makes sense to invest here in this compounder.

Read more

Technological transitions can move very rapidly, with adoption rates moving from zero to maturity in a span of less than 10 years. Good recent examples here are smartphone usage, social media apps and streaming services, and looking at history, adoption cycles are only speeding up:

But even more than 100 years ago, the streets of New York went from horse and carriage to being packed with automobiles in a timespan of only ten years:

When these technological transitions are on the verge of taking off, the market frequently massively underestimates their potential. A few examples are Tesla, Nvidia, Amazon, Microsoft, ASML, Shopify, Netflix, ServiceNow etcetera, the list of examples is long. Thinking about barriers to entry and long term market shares are good tools to separate the wheat from the chaff. Companies without a sufficient moat won’t be able to seize the new opportunity as competitors quickly emerge, eroding returns in the industry despite high levels growth. Secondly, you want to pick new technologies that have a reasonably high chance of making it, or at least sizing your position accordingly. Having a large weighting in a long shot is obviously foolish.

Clearly tech is an extremely lucrative ground for stock pickers, although not an easy one to navigate, as usually an in-depth knowledge on the underlying technology and industry is required. However, the robotaxi and autonomous driving markets is a space where investors have met disappointment over the last eight years. In the years of 2016 to 2018, we were being promised by some of the largest tech companies that autopilots were just around the corner, with robotaxis due to launch shortly. Obviously this didn’t happen and now everyone seems to have forgotten about the space, myself included up to last year. I didn’t look for years at his space.

But this is another beauty about stock picking, the large cap tech names exposed to this theme did enormously well anyways in the subsequent years, as their core businesses performed extremely well despite the robotaxi moonshot not coming through. In the second part of this note, we will look to find names with similar asymmetric return profiles i.e. limited downside if we’re met with disappointment again, but large upside if the robotaxi opportunity can successfully be rolled out this time around.

Subscribe now

Transformers as a game changer in autonomous driving

The reason for having a fresh look at this space is that AI technology has evolved considerably since the 2016-2018 years, which we’ll now go through.

Firstly, with the rise of ChatGPT, clearly innovation is still progressing at an impressive space in the world of AI. The algorithms underlying ChatGPT were only first thought of by researchers at Google in 2017, i.e. transformers. Transformers have a mathematical methodology of remembering context over a wider time-span. Similar to how when you’re reading a text, you’re remembering context from previous texts you’ve read.

Previous AI algorithms only made use of sentences they’ve just read (RNNs). So just because basic convolutional neural nets combined with RNNs and lines of handwritten C++ code weren’t sufficient to drive cars in 2018, AI algorithms are still evolving rapidly and in particular, the addition of better context-aware methodologies will obviously provide a big help.

So let’s say a pedestrian is crossing the street behind a parked truck, if the cameras of the car had already detected the pedestrian earlier, the transformer can remember that a pedestrian is still on his way behind the truck. With a convolutional model this isn’t possible, as it is purely analyzing the images it is being fed. An RNN will do better, however, its time span is limited, so if you’re being held up somewhere for a longer period, the transformer will still be able to remember the important context whereas with the RNN, it will be dropped out of memory.

The other advantage is that transformers can absorb much larger amounts of data. Again, this is hugely important in understanding context. Let’s say there are two dogs on opposing sides of the road. If the model has seen situations like this before, it now knows that these dogs are extremely likely to start interacting with other, such as one dog suddenly running across the street. Automatically, the vehicle can start slowing down to prepare for a possibly needed halt.

A recent paper from the self-driving team at Google illustrates how transformers can be used to model behavior of agents in the surroundings:

“Inspired by recent language modeling approaches, we use a masking strategy as the query to our model, enabling one to invoke a single model to predict agent behavior in many ways, such as potentially conditioned on the goal or full future trajectory of the autonomous vehicle or the behavior of other agents in the environment. Our model architecture employs attention to combine features across road elements, agent interactions, and time steps. We evaluate our approach on autonomous driving datasets for both marginal and joint motion prediction, and achieve state of the art performance across two popular datasets. We show that our model can unify a variety of motion prediction tasks from joint motion predictions to conditioned prediction.”

A recent paper on Arxiv summarizes current developments in transformers and autonomous driving. Below I’ve highlighted the most relevant paragraphs, which should give you a good idea of how transformers are being used:

“Transformers have revolutionized Natural Language Processing (NLP), with models like BERT, GPT, and T5 setting new standards in language understanding (Alaparthi 20, Radford 18, Raffel 20). Their impact extends beyond NLP, as the Computer Vision (CV) community adopts Transformers for visual data processing. This shift from traditional Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to Transformers in CV signifies their growing influence, with early implementations in image recognition and object detection (Dosovitskiy 20, Carion 2020, Zhu 20) showing promising outcomes.

In Autonomous Driving (AD), Transformers are transforming a range of critical tasks, including object detection (Mao 23), lane detection (Han 22) and segmentation (Ando 23, Cakir 22), and can be combined with reinforcement learning (Seo, Vecchietti) to execute complex path finding. They excel in processing spatial and temporal data, outperforming traditional CNNs and RNNs in complex functions like scene graph generation (Liu 22) and tracking (Zhang 23). The self-attention mechanism of Transformers provides a more comprehensive understanding of dynamic driving environments, essential for the safe navigation of autonomous vehicles.

Transformers in Autonomous Driving function as advanced feature extractors, differing from CNNs by integrating information across larger visual fields for global scene understanding. Their capability to process data in parallel offers significant computational efficiency, essential for real-time processing in autonomous vehicles. The global perspective and efficiency make the Transformer highly advantageous for Autonomous Driving technology, enhancing system capabilities.

The application of Vision Transformer models has led to significant progress in 3D and general perception tasks within autonomous driving. Initial models such as DETR (Carion 2020) adopted an innovative method to object detection by framing it as a set prediction issue, employing pre-defined boxes and utilizing the Hungarian algorithm to predict sets of objects. This methodology was further refined in Deformable DETR (Zhu20), which incorporated deformable attention for improved query clarity and faster convergence. DETR3D (wang 2022) extended these principles to 3D object detection, transforming LiDAR data into 3D voxel representations. Additionally, Vision Transformers like FUTR (Gong 2022) and FUTR3D (Chen 2023) have broadened their scope to include multimodal fusion, effectively processing inputs from various sensors to enhance the overall perception capabilities.

Transformers are increasingly pivotal in autonomous driving, notably in prediction, planning, and decision-making. This progression marks a significant shift towards end-to-end deep neural network models that integrate the entire autonomous driving pipeline, encompassing perception, planning, and control into a unified system. This holistic approach reflects a substantial evolution from traditional models, indicating a move towards more comprehensive and integrated solutions in autonomous vehicle technology

In trajectory and behavior prediction, Transformer-based models like VectorNet (Gao 2020), TNT (Zhao 2021), DenseTNT (Gu 2021), mmTransformer (Liu 2021) and AgentFormer (Yuan 2021) have addressed the limitations of standard CNN models, particularly in long-range interaction modeling and feature extraction. VectorNet enhances the depiction of spatial relationships by employing a hierarchical graph neural network, which is used for high-definition maps and agent trajectory representation. TNT and DenseTNT refine trajectory prediction, with DenseTNT introducing anchor-free prediction capabilities. The mmTransformer leverages a stacked architecture for simplified, multimodal motion prediction. AgentFormer uniquely allows direct inter-agent state influence over time, preserving crucial temporal and interactional information. WayFormer (Nayakanti 2023) further addresses the complexities of static and dynamic data processing with its innovative fusion strategies, enhancing both efficiency and quality in data handling.

End-to-end models in autonomous driving have evolved significantly, particularly in planning and decision-making. TransFuser (Chitta 2022, Lai 2023) exemplifies this evolution with its use of multiple Transformer modules for comprehensive data processing and fusion. NEAT (Chitta 2021) introduces a novel mapping function for BEV coordinates, compressing 2D image features into streamlined representations. Building upon this, InterFuser (Shao 2023) proposes a unified architecture for multimodal sensor data fusion, enhancing safety and decision-making accuracy. MMFN (Zhang 2022) expands the range of data types to include HD maps and radar, exploring diverse fusion techniques. STP3 (Hu 2022) and UniAD (Hu 2023) further contribute to this field, with STP3 focusing on temporal data integration and UniAD reorganizing tasks for more effective planning. These models collectively mark a significant stride towards integrated, efficient, and safer autonomous driving systems, demonstrating the transformative impact of Transformer technology in this domain.”

Clearly the pace of innovation is extremely vibrant in this field, with a large number of developments only being introduced over the last few years.

An example of how transformers can be used for route planning of the autonomous vehicle while predicting the movements of the surrounding objects:

An example of how transformers can be used to detect the key objects in the surroundings of the autonomous vehicle:

The key challenge in autonomous driving is the difficulty for the vehicle to understand the full context of its surroundings. For example, this lane is purely for trams and a wide variety of other complex city environments, such as street workers walking around with road signs, animals on the road, an ambulance which needs to pass through, an agent on the road waving instructions with his hands, a bicycle being transported at the back of a car, or images of persons or animals being displayed on vehicles.

There are an extremely wide variety of edge cases. Transformers will clearly be a big help in understanding this context, although it remains to be seen whether it will be fully sufficient.

In the meanwhile, autonomous vehicles (AVs) are clearly performing better in the field. With FSD 12, Tesla has moved to a single end-to-end neural network for autonomous driving, including transformers to process vision input. This network was trained on millions of video clips compiled from the Tesla fleet’s best drivers and replaced over 300,000 lines of C++ code. User reviews from what I’ve seen have been universally positive, for example with both Brad Gerstner and Bill Gurley comparing it to a ChatGPT or iPhone moment. At the same time, robotaxis are expanding operations in more and more cities.

The final factor in what is different now compared to the previously failed rollout in 2018, is that thanks to our friends at ASML, TSMC and Nvidia, hardware has just gotten so much better. The big advantage of better hardware in terms of autonomous driving is that it allows for larger parameter models to be trained and inferenced, thereby allowing the vehicle to understand much better all the necessary context from its surroundings. This leads to an increased probability that the vehicle will continuously make the correct decision, which should lead to strong outperformance compared to human drivers over time.

Lastly, there is a movement against robotaxis as people see them as unsafe. However, human drivers on the road produce a plethora of dangers: speeding, distracted driving, drunk driving, careless driving, running red lights and stop signs, cognitively impaired drivers, tailgating, road rage, etc. The US National Highway Traffic Safety Administration reports that in 2022, more than 289,000 people were injured as a result of distracted driving alone, such as drivers using their phone while driving.

Long term, if the data shows that autonomous vehicles have become 10x safer than humans, governments could well start considering to ban humans from operating a vehicle on the road. Similar to how smoking has now been banned in public indoor environments. Not too long ago, you could smoke on a train, in a plane, in the office, in college hallways, but not anymore.

In conclusion, putting the three factors together: we have lots of ongoing improvements in underlying mathematical algorithms giving the autonomous vehicle better context awareness and better decision making capabilities, we’re seeing indeed better results in the field, and we’re continuously still getting more powerful semi hardware allowing for better context aware models. So we could be at the start of a classic sigmoid-curved adoption pattern for new technologies — see the first chart we went through in this article — i.e. slow at first, with an accelerating slope thereafter until high adoption rates have been reached.

For premium subscribers, we’ll review:

• Which companies are currently best positioned in the autonomous driving and robotaxi spaces.

And we’ll also go into one name in detail:

• The company currently has an interesting collection of assets that provide a growing and cash generative business while being cheaply valued.

• Its robotaxis are currently scaling up as we speak.

• We’ll calculate a possible size for its robotaxi market by 2030 and we’ll take reasonable assumptions, such as robotaxis taking a 15% market share in the number of miles driven, and show how much upside this can give in the share price.

A name with potentially large upside

Read more

Workday is one among a number of disruptive SaaS companies which followed in the footsteps of SalesForce to disrupt the traditional on-premise business software market. SaaS has a number of advantages, with the most important ones being that it is asset light, i.e. you can run it from the cloud without having to deploy in-house infrastructure, making it also easy to upgrade to new versions of the software. Traditional on-premise software was typically heavily customized by system integrators, making it a nightmare to deploy and move to newer versions. For example, companies which were in the process of implementing SAP or another ERP (Enterprise Resource Planning) system were notorious for having to announce earnings misses due to bloating costs.

As from the 2000s, waves of companies were founded with the aim of simplifying business software, and which only gathered momentum during the subsequent decades. Workday’s founders had deep expertise in how to accomplish this. David Duffield had already disrupted the mainframe market during the ‘90s by moving HCM (Human Capital Management or Human Resources / HR) software to a client-server on-premise based model. Client-server had numerous advantages vs the mainframe model, most importantly horizontal scalability and a decoupling of applications, allowing one set of servers to take care of the core software and other servers to handle the databasing. Mainframes on the other hand basically unify all computing resources on a single server, allowing only for vertical scaling in one central location where you have to install a more powerful server.

Duffield’s software company PeopleSoft was highly successful and was eventually sold to consolidator Oracle in 2005 for more than $10 billion. However, as Duffield correctly identified that software architectures were transitioning again, this time to the cloud, he founded together with his Vice Chair at PeopleSoft, Aneel Bhusri, Workday later in the same year. The goal of Workday was to build an HCM software in the cloud, something which it has been extremely successful at. The company grew firstly in the mid-market and gradually worked its way up to larger and larger accounts, also known as the enterprise market, the Coca Colas and Exons of the world. As most of US firms are service types companies, around 80% of them or so, Workday decided to integrate a Financials module within their core platform, basically allowing their clients to run their core business of people and financials from the same software. So an enterprise combining this with SalesForce has an extremely powerful cloud based solution to run customer relationships, human workforce and financials on.

This is Workday’s founder Aneel Bhusri in 2015 on the company’s strategy:

“We started with a clean sheet of paper trying to build a unified HR and Financials product. Oracle has an HR cloud and a talent cloud, one is Fusion and one is Taleo, just right there, you've got 2 user interfaces and 2 technologies. So their story is maybe slightly better than SAP, but it's still far from a great product that has happy customers. If you do surveys, you'll find that our customers are much happier than both.

And we're now on our 21st update of Workday, we've been at it for 7.5 years. So it's a product that is now well tested and it's really coming into its stride as a true legacy replacement solution. We also have a broad product line: accounting, cash and assets, billing and revenue recognition, projects, professional services automation, and inventory. So we have a complete financial solution for the non-manufacturing industry which in North America is the bulk of the market, probably 75% to 80% of the market.

For a non-manufacturing company, the incremental implementation cost is not as high. It's still a big implementation because you're turning off your legacy financial systems, but the big benefit if you look at a non-manufacturing company, their number one cost is inevitably people. And so if you want to understand what you're getting in terms of payback on your people, why not have the HR and finance system be one and the same, and why not do analytics on pay-for-performance?

We replaced SAP HR now over 100 times. We thought when we started the company that we'd be a natural for PeopleSoft customers, but we replaced almost as many large HR implementations that were SAP-based. And I just think it was an afterthought for SAP, it only became important to them because of Workday and that's when they bought SuccessFactors.

Upgrades for legacy systems are more expensive than implementations of a cloud system because they're so heavily customized. Every 5 years, if you're a Fortune 500 company, spending $30 million to $50 million to upgrade to the latest version of the legacy system, there had to be a better way. Which is why the cloud comes in, and we now take care of those upgrades for the customer. That is the bulk of where the cost savings come from, probably around 50% of cost savings. We probably develop 3 or 4x faster at Workday than we did at PeopleSoft and when a new version comes out, well, we moved a couple of weeks ago all 700 customers and all of their unique configurations in 6 hours. That's unheard of in the enterprise world.

We took customer service to a new level at Workday, we were great at customer service at PeopleSoft, but it's a different world. We'd hand them a software package, wish them luck, give them a card for a Deloitte or IBM services person and we'll see you at the user conference. In the SaaS model, we own the end-to-end delivery and it forces you to be closer to your customers. And that's a huge opportunity for cloud companies to differentiate themselves versus on-premise companies.

We have two formidable competitors in SAP and Oracle, but their strength today is mostly on distribution and less on product. Basically, we're in a race over the next 5 years to build out our distribution channels while they try to catch up on products. Whoever has the products plus the distribution channels first, they'll own this market.

Currently, they're very aggressive on price, in some cases, they'll give it away for free to keep an account. But that's nothing new, that's been the case for the last 4-5 years since they woke up to the threat of the cloud. In particular, SAP has great CEO relationships and so they'll go to those CEOs and say ‘hey, we've got something kind of like Workday's’. And most CEOs of Fortune 500 companies don't know the difference between the different cloud vendors. Bill McDermott in particular, he will go in at the CEO level and that will saw us out. So we might do really well with HR and IT, but we have to get very good at reaching CEOs, make investments in marketing that raises our awareness where we can actually get to the CEOs.

If I were SAP or Oracle, I'd actually give us HR and not let us get an inch into the world of ERP with Financials. And they've frankly done the opposite, it's kind of a head-scratcher to me, but the Financials world for cloud systems is pretty open for us. We're competing against legacy systems, not against other cloud systems. And I think the history of cloud computing versus legacy is pretty clear on who wins, if it's cloud versus legacy, cloud's going to win.”

For premium subscribers, we’ll take an in-depth look at:

• Workday’s current runway for growth

• The company’s current competitive position and the overall competitive landscape

• An analysis of the weak guide post Q1 and the subsequent sell off in the shares

• Thoughts on software’s disaster quarter in general

• A discussion of the risks and opportunities coming from AI for Workday

• A financial analysis of the company

• Thoughts on valuation and whether the shares should be a good investment

If you enjoy research like this, please hit the like button and share the post with others who might be interested. Without further ado, we’ll now go into Workday..

Workday’s current runway for growth

Read more

As a brief recap, Cloudflare runs a network of edge servers located in smaller datacenters on which they’ve been building a wide array of software services, from cybersecurity to networking and application services. More recently, the platform has also become a mini public cloud version of Amazon AWS where developers can run workloads and store data. A key advantage of running services at the edge is proximity to the end consumer. Cloudflare can usually reach an end user within 50 milliseconds, giving an extremely smooth customer experience.

This has been a good story, and business is firing on all cylinders. This is the CEO discussing last quarter’s highlights:

“We achieved revenue of $379 million, up 30% year-over-year. We added 122 new large customers, those that pay us more than $100,000 per year, and now have 2,878 large customers, up 33% year-over-year. We are successfully moving upmarket and becoming a larger and more strategic vendor.”

He also gave a list of examples of big deals they signed, I’ll just flag two here for brevity:

“The National Cyber Security Centre, the U.K.'s technical authority for cyber threats, signed a 3-year contract with Cloudflare to deliver its protective domain name service. PDNS protects over 1,400 U.K. organizations in central government, local government, health care and emergency services from malware and cyber threats. This was a very competitive process with several vendors and a rigorous technical evaluation.

A large international energy company signed a 5-year, $4.5 million contract. This new customer is going all in with Cloudflare's SASE platform, with 6,000 Zero Trust seats along with CASB, DLP, Browser Isolation, Magic WAN and Magic Firewall. Competing against a first-generation Zero Trust vendor, our focus on scalability and efficiency as well as the ability to consolidate several vendors due to the significant value in our overall portfolio were key factors delivering this win.”

Additionally, the company has been busy restructuring its sales organization with the aim of accelerating growth. We got an update on their latest efforts:

“Just 90 days ago, we announced that Mark Anderson will be joining Cloudflare as our new President of Revenue to accelerate our next phase of growth at scale. The number of applicants for strategic account and enterprise sales positions increased 56% in March versus February following Mark's appointment. Across all positions, we had over 350,000 applicants in Q1, up 47% over the same quarter last year. We added several senior go-to-market leaders with proven track records in their areas of expertise, including a new Chief Partner Officer, a new Head of Global Sales and multiple regional strategic account sales leaders. If you're a sales professional who wants to win with great products and world-class leadership, the word is out, Cloudflare is the place to bet the next stage of your career. Beyond incredible hiring, our sales productivity from existing team members improved year-over-year. Sales cycles were similar to last quarter and new pipeline attainment exceeded our expectations. I feel extremely confident and clear in the long-term opportunity that Cloudflare has in front of us.”

Anderson was previously the CEO of Alteryx, President at Palo Alto Networks, and Head of Sales at F5 Networks.

As a final positive, the company has been installing GPUs in their end servers so that Cloudflare’s edge network can now be leveraged to run AI models close to the consumer. These are typically smaller models, not the huge language models such as GPT-4, but more in the class of Llama-2 and 3, as well as more traditional neural nets. We got an update on this interesting story:

“We're ahead of schedule rolling out GPUs across our network and now have them running in more than 150 cities globally, making us what we believe is the most widely distributed AI cloud by a huge margin. Our next generation of servers that begin to roll out in Q2 have GPUs built in by default and will support faster inference and even larger, more complicated models. Developers are building incredible new applications using Workers AI, and we're making it increasingly easy for them. We rolled out our partnership with Hugging Face, making it one-click simple to deploy most of their catalog of models to Cloudflare's network. And we added other bleeding edge models, including releasing Meta's Llama-3 simultaneously the day it was announced.

One of the largest e-commerce platforms out there is starting to use Workers AI to do image classification and generate keywords for their shops. A start-up in the drone space is taking imagery and turning it into 3D models. A handful of the public platform companies are using Workers AI for video description, transcribing and translation. A multinational online food ordering company, every once in a while, a restaurant does submit a picture for one of the dishes. And so they're actually experimenting with Workers AI to generate pictures of the food that you're serving. Hopefully, they disclosed that it's not an actual picture. A handful of photo and video editing platforms are using Workers AI or evaluating it in order to do image enhancement and make that better. There's a startup that's generating songs in local languages. Large consulting companies, public companies, the big gaming companies, all of which are worried about shadow AI, and so they're actually using Workers AI in order to understand how their own teams are using AI. A start-up is using a series of enterprise summarization tools where they can look at a bunch of emails and very quickly summarize them into executive reports. There's a public company in the financial services space that's doing real-time fraud detection.”

Having a GPU in 30,000 servers could bring in $702 million in annual revenues and $140 million in EBIT, which would nearly double Cloudflare’s estimated EBIT for this calendar year. So AI on the edge can become a meaningful business for Cloudflare over time and it’s certainly an interesting story to keep an eye on.

However, following the Q1 release, Cloudflare’s shares dipped 17%. These are the words from the CEO that spooked the market:

“In the short term however, my crystal ball is less clear. We see a lot of signals based on our privileged position running a good chunk of the Internet. Even without that visibility, if you've been watching the news at all, it's clear that the near-term outlook for the world is uncertain, increasing tensions in the Middle East, no end in sight on the Russia-Ukraine war and potential signs of instability in Asia. It's not at all certain on anything we see that things will get worse. But we do know from even recent history that macro factors can impact short-term sales trends.”

For comparison, these are the CFO’s remarks:

“Ongoing mixed macroeconomic data points and heightened geopolitical uncertainty remind us that we continue to operate in a business environment that remains challenging to predict. As a result, we remain prudent in our outlook for 2024. As you know, conservative is not a word that we use. When we talk about giving guidance, we tend to be thoughtful and prudent about it. And of course, it's a balance of the tailwinds we see, but also the uncertainty that is out there and in every major go-to-market transformation. In our case, it's actually more an evolution. There is risk, and it's reflected in our guidance, too.”

So what they’re talking about are mostly concerns about the economy as a whole, with a particular focus on a number of regional geopolitical conflicts which could indeed spill over into wider conflicts. If this is what concerns you, all cyclicals and high-growth stocks are indeed at risk, and you could rotate a portion of the portfolio into military suppliers and commodity names for example.

The market’s reaction can be explained by the fact that Cloudflare was trading on 18x revenues. With a multiple like that, you not only need an impeccable quarter, but also an impeccable guide. And given that there was some hesitation from management, the shares traded meaningfully lower. Personally, I think they’re sandbagging a bit and with the above mentioned drivers — i.e. cybersecurity, developer services, a better sales force, and AI on the edge — that we’re actually looking at a strong number of years in this name. Later on, the CEO seemed to be walking back already some of his comments, highlighting how increased global tensions are benefitting their cybersecurity business:

“We're fortunate that we're in the cybersecurity space, perhaps one of the few sectors that can actually benefit from increased global tensions. We're already seeing that, especially in our government business. The short term is uncertain, the long term is bright and so in the medium term, we're going to thoughtfully invest in our go-to-market efforts, in great engineering and in disruptive new products that deliver incredible value.”

In my view, when these high-growth and good quality tech assets trade below 15x forward revenues, the risk-reward is starting to look attractive. Below is an example of how ServiceNow has traded on an EV-to-Sales basis and how the company has been growing its revenues. Even latecomers to the party in ‘16 who bought this stock at 8x sales still managed to catch a 10-bagger:

The market has gotten smarter since then and you can’t buy a name like this anymore on 8x forward Sales. In fact, the cheapest Cloudflare has ever traded is just below 11x:

Looking at consensus modelling, the forecasted margins in the coming years look very conservative. As a result, we’re looking at a very high consensus PE. However, with the above mentioned drivers and the conservative modelling, it looks like there is plenty of opportunity to do better:

Unity vs Applovin

Unity is the dominant engine for mobile games development. Historically, despite its strong positioning, this business has been poorly monetized. We previously discussed how the company under new leadership is now moving to a new pricing model, which should unlock value for shareholders in the coming years. Typically, computer aided design (CAD) tools such as Unity are high margin business and trade at premium multiples, as investors like the sticky, recurring revenues with substantial pricing power. So over time there is strong potential for Unity to transition from being somewhat of an ugly duckling currently to a quality-growth asset.

The other part of the business is managing ads for game publishers, where they compete with dominant player Applovin. We also discussed the latter a number of months ago as it offered attractive value for the high growth they’re generating. Unity is nowhere close to Applovin in this space, or at least for the moment. Whereas Applovin’s ad business has been growing at rates of close to 100%, Unity’s has been flatlining.

There should be scope for Unity to become a much stronger player here over time, but it’s good to keep in mind that a large part of this business is lower quality in nature and won’t attract software-type multiples.

Let’s start with the better quality part of ads. This is basically a take rate business where a game publisher sells ads within their games and Applovin’s or Unity’s platform will make sure that these are sold to the highest bidders. So far so good. This business is also called mediation. The second part of the business is what is called discovery. A game studio can pay Applovin or Unity to attract say 1,000 new users for a fixed fee. What these discovery platforms subsequently have to do is find these users at a lower cost, so that the business can make a profit margin. The way they do this is by leveraging the data they have and then buy ads in other games to target gamers which are likely to be interested in this new game. So it is really sort of an algorithmic trading business, where you buy ads cheaply and make a margin on this trade. It reminds me a lot of Criteo.

To be clear, the latter is not a business I would invest in for the long term. However, in the case of Unity, they do have a lot of data and are now integrating acquired IronSource, so it could be an interesting improvement story over the coming years. This is Unity’s CEO updating us how they’re planning to compete more effectively with Applovin:

“We're operating with, frankly, two different data science departments. So we weren't fully integrated in what we were doing. This year, as part of completing the IronSource merger, we've brought those organizations together. There was frankly a lot of data we just weren't fully using. We are now training our models using substantially more data. We actually now have some initial results that look super positive. And so we will be rolling those out more at scale through the rest of this quarter and it will bleed into July as well, which is why we see the back half of the year, as we've projected, kind of sequential improvement in the back half of the year.

Some of these tests, as we then talk to partners, they have been extremely positive, and we've actually had some people moving some more to LevelPlay with the results. And these are large Max customers (Applovin’s mediation platform) and so that leads us to believe that the interventions actually close much of the gap versus AppLovin.

Are we equal to them or not in ROAS (return on ad spend)? It's hard to exactly say, but the customer feedback has been actually extremely positive, and we are seeing share shift. And we think when we roll these out, we will continue to see that. No one wants to have one partner and so people want competition there. And so we've had a number of customers just say, we're being patient. We're expecting you to catch up.”

Applovin currently not only has the momentum, but the company keeps improving their engine as well. This is the company’s CEO on the recent call:

“Two themes are important to understand as our business goes forward. First, a key driver of our growth will be the ongoing improvements to AXON. Our models are still in an early stage and will continue to improve themselves. But more importantly, our teams are still finding ways to materially improve these algorithms. While these gains may not be predictable, they may sometimes lead to quarters like Q1 where we far exceed expectations. Second, there is nothing that limits our models to just gaming. By expanding into web-based marketing and e-commerce, we expect our AI models to improve with added demand diversity.”

On gaming, Unity introduced a new runtime fee that should work out to be around 2.5% of game revenues above $1 million. Previous CEO Riccitiello really botched the introduction of this new pricing model as he wanted to introduce it retroactively i.e. including for games already published. Needless to say this caused a lot of resentment in the gaming community which led to Riccitiello’s exit last year. After this episode current interim CEO Whitehurst took over.

This is Whitehurst updating us on the runtime fee situation and the launch of their new gaming development platform, Unity 6:

“We are seeing customers who were a bit ruffled at the end of last year feeling much more confident with us. We've gone from the what and why of a runtime fee to a recognition that the industry wants to make sure that we invest in the runtime. So we're having super productive conversations with our large customers. On Unity 6, that doesn't release until later this year. But if we look at the leading metrics, both the downloads and usage, it's the highest quality release at this stage that we've ever had and the time lines have pulled in.”

Unity has now hired a new CEO, Matthew Bromberg, who previously held the positions of COO at mobile games developer Zynga and VP of Strategy at Electronic Arts. Whitehurst, who did a great job at Red Hat, will stay on as executive Chairman. This is Whitehurst on the transition and outlook for the business:

“I'm remaining as Executive Chairman and what I'm really excited about is working on the Industry side of the business. I'm probably biased, but I still think that's a bigger opportunity in the long run than the Gaming business. It is still a relatively nascent opportunity for us, but the more I'm out talking to customers, the more opportunity I see there. In the long run, I continue to be super optimistic about the business as we are essential to gaming.”

Industry is essentially using the Unity engine to create digital twins of workplace environments and then also use this to train the workforce. The engine can also be used to illustrate to customers what a final product will look like.

Unity also closed a partnership with CapGemini which is one of the largest system integrators in the world. In software, it is important to keep an eye on these types of deals as these IT services companies enjoy a large corporate customer base, to which they can now also sell Unity’s software.

Under Whitehurst, the company divested some non-core businesses which weren’t making any money, after the acquisitive legacy of Riccitiello. So this year we get a reset in revenues although the impact on profitability will be limited. And as from 2025, top line growth should return. These are the sell side’s numbers:

Note that SBC is high in this business at $750 million, which at the current market cap of $8.7 billion is very dilutive.

On the positive, if a new pricing plan is successfully introduced, there is a lot of upside to consensus EPS:

And then we also have the prospect of Unity’s gaming engine being further utilized in the so-called industries business, which is already contributing 30% of the engine’s revenues. Furthermore, Unity has a strong share in development for VR environments and the company is also integrating AI tools into their engine, which can speed up development for studios and which is something they can start monetizing as well. Lastly, there is potential for the company’s ad business to become a more effective competitor to Applovin. So there are really a lot of levers here. Overall, I still like this story although the high SBC is obviously the main negative.

Applovin’s valuation remains undemanding at 15x next-twelve-months’ EPS. The company also has an attractive FCF yield of above 5% while consensus is projecting double digit annual top line growth with attractive operating leverage. SBC is much better under control, with only 1.3% of the market cap being paid out to employees over the last twelve months.

I actually invested in both names in January and while Unity has been a dud, or at least so far, Applovin nicely more than doubled. So net-net this space has still been a good place. Unity is a name I can see myself staying in for the long term, but with Applovin the plan is to purely ride this wave of momentum the company is currently enjoying.

Shopify’s sell-off

We reviewed Shopify last December following the company’s investor day, and while the SaaS platform for e-commerce stores remains a compelling growth story, valuation was looking a bit steep at the time with a forward PE of 75x. While the company remained performing solidly with gross profit growth rates currently at 33% year-over-year, the company gave a tepid outlook for Q2, causing a sharp decline in the share price of around 25%. This stock basically had the same problem as Cloudflare, with such a highly demanding valuation, results and outlook need to be impeccable.

Shopify has two revenue streams, software subscriptions where customers can run their e-commerce operations from and merchant solutions which is mostly payments processing:

This is Shopify’s President discussing the business and outlook, note that the company recently divested their logistics business which is impacting results:

“Revenue for the first quarter was $1.9 billion, up 23% year-over-year, which equates to 29% year-over-year growth when excluding the logistics businesses. This represents the fourth consecutive quarter that our revenue growth has been greater than 25% on an organic basis. The key drivers of this growth were: the GMV strength just discussed; growth in Subscription Solutions revenue from both new margin growth and the pricing increases on standard plans; and lastly, increased payments penetration, which hit 60% for Q1.

Our expectations for the second quarter of 2024 are as follows. We expect Q2 year-over-year revenue growth to be in the high teens on a GAAP basis, which equates to a year-over-year growth rate in the low to mid-20s when excluding the 300 to 400 basis-point impact from the sale of our logistics business. An important dynamic to highlight is the impact of the standard and plus pricing changes and how they affect our growth rate for Q2 versus Q1. The impact of the pricing changes in standard and plus will have a smaller combined benefit in Q2 versus Q1.”

The company also has been introducing its Shop Pay button which has a very smooth user experience, or at least whenever I’ve used it, in the form of low latencies to complete a payment. This compares to Paypal’s current button which gives a rather sluggish impression, probably that codebase was written 20 years ago or so and has only been expanding ever since. Sometimes it’s best to just rewrite the entire thing from scratch and speed things up.

This is Shopify’s President on Shop Pay:

“Shop Pay is the highest converting checkout on the Internet. It converts up to 36% better than competition and 15% more on average. There's now 150 million buyers that have opted into Shop Pay. And for Q1 alone, we facilitated $14 billion of GMV, that's a 56% increase year-on-year. Customers want to use Shop Pay because the mere presence on the checkout results in a 5% higher conversion. It's becoming a really bad idea for any brand or any retailer on the planet to not use it.”

This is a company with limited competition and Shopify is also successfully moving up-market. Traditionally Shopify was a SaaS platform for mom-and-pop ecommerce stores, but more and more larger brands are starting to run their operations from the platform. Both for direct-to-consumer and wholesale activities. Shopify’s President highlighted an independent study that concluded the Shopify platform to have a lower cost of ownership of up to 36% in the enterprise space.

With the still attractive growth rates and low penetration rates in e-commerce, there is a lot to like in this company. Near term growth will be a bit weaker as the tailwind of past price increases rolls off, however, investors have the prospect of new price increases in the coming years. Typically I like adding to good quality, high-growth assets when they sell off sharply for short term reasons and historically, this has been a strategy which has been working extremely well. So I’ve taken again a position in Shopify and will probably stay in it for the long term.

If you enjoy research like this, hit the like and restack buttons, and subscribe if you haven’t done so yet. Also, please share a link to this post on social media or with others who might be interested, it will help the newsletter to grow which is a good incentive to publish more.

Subscribe now

Share

I’m also regularly discussing tech stocks on my Twitter.

Disclaimer - This article is not a recommendation to buy or sell the mentioned securities, it is purely for informational purposes. While I’ve aimed to use accurate and reliable information in writing this, it can not be guaranteed that all information used is of this nature. Before making any investment, it is recommended to do your own due diligence.

The chart below illustrates how large pharma R&D has been growing at a 7% rate. However, biotech is much more volatile, with booms and busts of R&D funding. According to Simulations Plus (SLP), the overall spend on drug development from large pharma and biotech combined would be around $200 billion per annum. Within that, biosimulation is a market of only $3 billion and is growing at an attractive annual rate of around 17%.

This is SLP’s CEO at the BoA conference in ‘21 discussing the challenges in drug development and how their software tools can improve outcomes:

“A number of 5,000 target drug molecules go into the development cycle for every 1 drug that is brought to market. The challenges are how does the drug work, is it safe, what's the right regimen, what is the biological effect and what are the risks or toxicity? Our products and services are focused on answering those questions, reducing the trial-and-error nature of the historical process. Modeling and simulation is a means by which you can get to an efficacy, safety and regulatory approval more efficiently.

We deliver very accurate drug and biological models to our clients, in the analysis from early discovery through to regulatory approval. We have a portfolio of software products, all of which are recognized as leading tools for model-informed and data-driven drug development. And we have a consulting practice that supports our clients from basic outsourcing needs to addressing more tricky issues in terms of simulation techniques.

Having started in the early days of adoption 25 years ago, we now have 250 clients and a high retention. Modeling and simulation is a combination of skill sets that runs the gamut from physics, chemistry, statistics and pharmacology. The number of scientists in this space that could run the engine and use the software was relatively low in the early days, it was a unique skill set but today, there's more than a dozen PhD programs in computational biology out there.

Scientific acceptance in the regulatory environment has been another factor. Over time, the FDA and their equivalents around the world began to utilize simulation internally in their evaluation of drug submissions. That built up to a FDA that today issues guidelines to the industry in support of modeling. When COVID hit, one of the first FDA guidelines to the industry was how modeling and simulation could be used to still maintain the value of clinical trials that were disrupted.

When the FDA issues for example a bioequivalence waiver, basically it waives the requirement to go into the clinic and perform a clinical trial to prove an efficacy or toxicity issue. That can be based on data and analytics brought to them through simulation. Obviously, the time and the cost of the clinical trial can be high, but the real big impact is getting that drug to market 6 months or a year earlier. Initiating the revenue stream for a drug can be quite dramatic for our clients.”

At the recent capital markets day, the company disclosed that the number of clients had risen to 400. Around 40% of SLP’s revenue mix stems from services which usually I’m less interested in as barriers to compete tend to be lower here. However, as these services are mostly helping clients with modelling and simulation in SLP’s tools, it’s really tied to the software business and there probably aren’t many competitors around which will have this know-how.

The number of scientific publications making use of biosimulation has been accelerating over the covid years:

SLP mentioned that over 1,300 publications cite their products and services contributions, so comparing these numbers to the data above would give them a market share of 16% in the overall space. Later on we’ll dive deeper into the competitive landscape.

Also the FDA has been expanding guidance on where biosimulation tools can be incorporated:

Another new driver for adoption is a regulatory push to reduce testing of chemicals on animals, from the recent capital markets day:

“About 10% of our customer base today is in the non-pharmaceutical space, chemicals, cosmetics and consumer goods companies that have already been told by regulators around the world, especially Europe, that they need to greatly reduce or completely eliminate animal testing on some products. With the offerings that we have, you can start predicting without any animal data and with reasonable accuracy, what the absorption and exposure of chemicals systemically would be.”

Similar to the industrial CAD and simulation software markets, the biosimulation market utilizes a series of software tools specialized for each task. Gradually tools are now being integrated into software platforms, although the focus is also still on building out these tools with more advanced methodologies so that they can be cross-sold to clients. These then create a workflow from this series of tools. SLP’s CEO discussing this topic:

“We have several large entities that utilize the full suite of products and services but there are gaps. The industry itself is a very siloed process, with discovery and clinical departments. Modeling and simulation developed in its early days to provide solutions to these various departments in a sort of point effort. As modeling has become more strategic for our clients, they look more broadly at its use from start to finish. And that's been a focus on our part to develop a suite of platforms that can support the full scope of modeling needs. We’re very focused in terms of making these tools easier to use and more efficient, and that brings new companies into the fold as well.”

A number of these tools have been added to the portfolio by acquisition. This is SLP’s CEO discussing their M&A strategy:

“The strategy is broad and covers both the software as well as the service space, albeit our focus leans in the direction of the software side. This is evidenced by 2 of the 3 acquisitions to date having been software oriented and I would anticipate that to be the case going forward. Our focus on the software side is both in terms of expanding functionality that could be integrated into our existing platforms as well as acquiring other stand-alone solutions that address needs between what we provide today. There is a good population of companies out there, small privately funded entities that have developed solutions that would be great additions. We've been pretty strict in terms of our criteria for acquisitions, looking for product and culture fit, but also focusing on accretive acquisitions at a good valuation point.”

For premium subscribers, we’ll review in-depth:

• Simulations Plus’ software and services portfolio

• How biotech should be moving into a new upcycle

• A detailed analysis of the financials and valuation of the company

Simulations Plus’ software and services portfolio

Read more