Subscribe now

Hi friends,

AI has made a lot of progress, but even the best agents don’t feel quite like coworkers yet in the way that humans do.

Dwarkesh Patel made the case last year in a piece arguing AGI isn’t close, and the bottleneck he kept landing on was the idea of continual learning:

But the fundamental problem is that LLMs don’t get better over time the way a human would. The lack of continual learning is a huge huge problem. The LLM baseline at many tasks might be higher than an average human’s. But there’s no way to give a model high level feedback. You’re stuck with the abilities you get out of the box. You can keep messing around with the system prompt. In practice this just doesn’t produce anything even close to the kind of learning and improvement that human employees experience.

The reason humans are so useful is not mainly their raw intelligence. It’s their ability to build up context, interrogate their own failures, and pick up small improvements and efficiencies as they practice a task.

Models today learn during training, then the weights freeze, and from then on they don’t get better at your job no matter how many times they do it (absent skill files and memories).

Yash Patil, who runs Applied Compute, frames the same gap in ’s class in terms of how “cheaply” a human learns:

If you go and burn your hands on the stove, you just need to do that once and then you know the stove is hot and not to put your hand on the stove. These models today are not really like that.

A person updates from a single sparse signal. A model needs the lesson spelled out every session.

So the question then becomes how you let a system actually learn from doing the work. And there are basically two places you can do it: in the weights, or in everything around the weights.1

Weight space

The first place to potentially solve continual learning is the weights themselves. The most common version: take the signal from real usage at some interval and post-train on it with fine-tuning or RL, so that the model learns / improves over time based on actual usage.

Two research directions push further than this. One is test-time training, where the model updates its own weights while it runs, adapting to the specific input in front of it instead of staying frozen between training runs. The other somewhat adjacent idea is meta-learning, where instead of training the model on the right answers, you train it on how to find or discover them, so it learns to reason and search its way to an answer rather than memorize one. Technically, this doesn’t need a weight change since it’s a training time technique, but it could improve the ability of the model to feel like it adapts upon use better.

Getting to true continual learning through weight updates is challenging for multiple reasons:

One is catastrophic forgetting. Try to fine-tune a model on something new and it tends to quietly get worse at things it already knew. In the context of continual learning, the risk would be you teach it this month’s workflow and it loses something it was fine at last month. Keeping a model both stable and plastic at the same time is an open research problem which is why you can’t simply just fine-tune on every interaction.

Two is the practicality of learning continuously. Sure, you could maybe post-train a model every month or every week. Could you do it on every interaction? Would it be on a company basis or an organization basis, or could you do it on a per-person basis, per-person interaction? In an ideal sense, that is, in some ways, what continual learning might be: every person has a set of weights that are constantly changing and adapting based on how they are interacting with that model or the harness around that model.

Satya Nadella has a framing around this where perhaps all the company tacit knowledge lies in a LoRA layer in some way:

Clearly it resides in people’s heads and it’s the classic know-how that accrues and compounds. I think it’ll also reside and compound as weights in some LoRA layer that is unique to your company. I feel like the new intellectual property at Eli Lilly or at Microsoft or at Stripe at some point can be also, besides all the humans, besides all the other artifacts we have, I think we’ll also say, “Oh, they are in some embedding.”

One downside to continual learning in weight space is permissioning and governance. You can gate what context a given person sees pretty easily. But once something is trained into a model, it’s just in there, and “this person shouldn’t have learned that” is a much harder thing to enforce. Similarly, models could diverge in unintended ways and regress, and so being able to monitor and govern to do things like roll back to a stable checkpoint when an update makes things worse and what data changed which weights becomes important.

Continual learning in weight space is very much an ongoing research area, with Cursor’s online RL perhaps one good example in production, where the model is improving from the online production usage every 5 hours.

This is still a relatively basic form of continued learning, in that it’s still one model that every user and every company is using. It’s not being personalized in some ways, but at least it is continuously improving based on production traffic.

We serve model checkpoints to production, observe user responses, and aggregate those responses as reward signals. This approach lets us ship an improved version of Composer behind Auto as often as every five hours.

Let’s now talk about another “technique” to continually learn.

Token space

Token space essentially means everything around the model. You keep the weights frozen and you change what reaches them. On a practical level, you can think of this as having two layers, as Harrison Chase splits it into, the harness (the code, tools, and scaffolding that drive the agent) and the context (the instructions, skills, and memory that configure it), or combine both to think of it as token space or the stuff in and around the model.

Context is more straight forward. An agent keeps a running memory, rewrites its own instructions as it goes or creates new skills, and runs over yesterday’s logs overnight to fold what worked into its config for tomorrow. OpenClaw calls that overnight pass “dreaming”. The good thing about learning at this layer is that it can be easily scoped or personalized: one evolving context per company, per team, or per user, all riding on the same base model.

The idea of a continually improving harness is perhaps more interesting. There’s an idea called meta-harness that does exactly this. You keep a filesystem of every past run, and a coding agent reads it and proposes new harness code, new tools, new prompts, with the goal of maximizing an eval score. You run that harness around a frozen model against a set of tasks, score it, and store everything (the proposed code, the reasoning traces, the score) back to the filesystem so the next proposal is smarter. It’s basically RL, except the thing getting optimized is the code around the model, not the weights inside it.

In some ways, learning in token space can be easier and cheaper and have better governance. But it often relies on storing the right information in skills, memory files, etc and the problem somewhat shifts to making sure the model retrieves it at the right step.

Harrison Chase at Langchain summarizes the various tradeoffs to learning at each level. At a practical level, getting to some form of continual learning that feels solved will involve learning at all three layers.

Companies Tackling Continual Learning

A lot of companies are trying to tackle continual learning now other than of course the frontier labs themselves. They fall into three rough camps based on their approach.

The first camp works at the model level, with new architectures or new ways to update weights that avoid catastrophic forgetting. Some are big architectural bets, others are smaller changes to today’s transformers.

• Learning Machine is building models that keep learning during inference, so the model picks up the skill of “how to learn” and adapts to a user or company after it ships.

• Chronologies is a lab trying to build models that continuously update without forgetting past information.

• TTT-E2E (Stanford and Nvidia) trains a transformer to update its own weights at test time, folding the context it sees into the weights as it goes.

The second camp works with today’s frozen models. They help a company learn from its own production traces, both by post-training the model and by tuning the harness around it. In practice it looks like RL as a service: take the traces, define what good looks like, and use it to update the system. Enough improvements here could mimic a feeling of continual learning.

• Applied Compute trains custom agents on a single company’s data and calls it “Specific Intelligence,” a model that improves based on production traces over time.

• Trajectory turns production usage into the signal for post-training and bills itself as a complete platform for continual learning.

This is a bit more well understood approach and is what companies like Cursor are already doing. However, at least today, these largely involve one model for the company as opposed to learning on an individual user basis.

The third camp keeps the weights frozen and does the learning in the harness and context (“token space” as discussed earlier). It never touches the model, which is cheaper and easier to govern.

• LangChain‘s LangSmith Engine watches production traces, clusters the failures, and proposes harness fixes as pull requests, which is the meta-harness loop as a product.

• NeoSigma builds the feedback loop between how a product gets used and the agent running it, suggesting improvements and changes at the harness and context layers.

• Companies like Letta, Mem0, Pinecone and others which are around persistent memory that carries across runs help at least improve the system based on context improvements over time.

Closing thoughts

One question around continual learning is whether we’ll have one breakthrough that takes us closer, or if it will be an incremental iterative improvement across all the layers. Already, with techniques like better memory, dreaming, meta-harnesses, the experience has improved significantly. But there’s still a lot more to be done.

In the near term I expect most of the learning to happen in token space, a frozen model with a smarter and smarter wrapper, scoped per company and per user, until from the outside it’s hard to tell the model didn’t actually change. The weight-space version, where the model itself learns from the job and a company’s know-how ends up living in its weights may require more of a breakthrough, but we know all the labs are currently focused on that.

If you’re building in and around continual learning, I’d love to talk. You can reach me at tanay at wing.vc.

Subscribe now

Hi friends,

SpaceX filed its S-1 this week, ahead of what will likely be the largest IPO ever (~$75B raised at a reported $1.75T valuation).

SpaceX is a unique and inspiring company. It’s not often you see a mission statement on an S-1 include “to make life multiplanetary, to understand the true nature of the universe, and to extend the light of consciousness to the stars”.

SpaceX on the face of it is three companies combined together: a rocket launch business, a satellite internet business, and an AI lab (after the xAI merger). But the way I’ve come to think about SpaceX (or at least how SpaceX wants to be thought of) after reading the filing is less as a rocket company with an AI lab attached, and more as the physical infrastructure layer for the AI and space economy. I’ll walk through how the three businesses work today, and how they could be far more connected in the future than they look.

In this piece, I’ll cover:

• SpaceX’s three segments today

• The two deals that change the AI story

• What the financials look like played forward

• Orbital compute and the Space economy

• The Starship Bet

• Musk’s pay package, gated on a Mars colony and data centers in space

I. Understanding SpaceX today

SpaceX reports in three segments: Space, Connectivity, and AI. They’re wildly different businesses on one balance sheet, so it’s worth taking each on its own.

Connectivity is the star today, growing fast and very profitable, essentially helping fund the rest. Space looks like a loser on operating income but is actually EBITDA-positive once you add back the Starship R&D and depreciation. AI is a deep loss on every line but more on that later.

A. Space

SpaceX was founded in 2002. The early history may be familiar to some: Falcon 1 reached orbit in 2008 (the first privately built liquid-fueled rocket to do so), they landed a booster back on Earth in December 2015, and by 2017 they were routinely reusing Falcon 9 boosters. Reusability has been key to the rest.

It dropped SpaceX’s launch cost to about $2,700 per kilogram, roughly 85% below the historical industry average of $18,500, and that cost collapse is what made everything downstream possible.

SpaceX began as an external launch provider. The early customers were NASA (Commercial Resupply from 2008, then crew), commercial satellite operators, and the Pentagon, and those contracts funded the company for its first decade and a half. Starlink only started launching in 2019. So the business was built on outside demand, and then internal Starlink deployment grew on top of it until it became the majority of flights.

SpaceX flew 96 Falcon launches in 2023, 134 in 2024, and 165 in 2025. But most of those weren’t for paying customers. Customer launches were 33, 45, and 43. The rest (63, 89, then 122) were internal: mostly Starlink deployments and some development tests.

The chart below highlights this shift. The launch business is now overwhelmingly SpaceX flying for itself, and the internal share keeps climbing (74% of launches in 2025, up from 66% in 2023).

They built a launch business for other customers and used the same capability to deploy their own constellation. The launch business and capability is the railroad of sorts that made everything else possible.

Still, the Space business makes over $4B in revenue for external launches. While some of that is government development contracts that aren’t tied to specific launches, one rough proxy is that at 43 external launches in 2025, each launch is worth about $95M in revenue to them. But the thing to note is that for internal launches that deploy Starlink, SpaceX doesn’t book any Space-segment revenue. It capitalizes the launch cost inside the Connectivity segment and depreciates it over time. So Space-segment revenue only reflects the revenue from the ~1/4 of launches which are external.

The filing notes this: despite a rising launch cadence, Space has “relatively lower revenue scale and revenue growth” because its results don’t reflect the internal launches that are the foundation of the company.

Financially, Space is the smallest segment and currently a money-loser. Revenue was $4.1B in 2025, up only 7.6%. It swung to a $657M operating loss (from a $21M profit in 2024). The reason is Starship. R&D in the segment jumped 64% to $3B as they poured money into the next vehicle, and in Q1 2026 alone Starship R&D was $930M. So the cash from launching satellites is being plowed straight into the next generation rocket that’s supposed to make the future vision achievable.

B. Connectivity

Connectivity represents the Starlink business, and it’s the segment that actually prints money.

As I mentioned above, it is only possible with the launch flywheel. Cheap, reusable, high-cadence launch is what lets you put up (and constantly replenish) a constellation of ~9,600 satellites in low-Earth orbit. No one else can deploy mass to orbit at that cost, which is why no one else has a constellation at this scale. The S-1 notes SpaceX has launched more than 80% of all mass to orbit globally each year since 2023.

The subscriber story of Starlink shows signs of a landgrab strategy. Starlink went from 2.3M subscribers (2023) to 4.4M (2024) to 8.9M (2025), and 10.3M by Q1 2026. Monthly ARPU fell the whole way: $99, then $91, then $81, and down to $66 in Q1 2026.

That ARPU decline is somewhat deliberate and represents both the push into lower-income countries and additional cheaper plans.

Overall, the financials of this segment are very strong. Connectivity did $11.4B in revenue in 2025, up 50%, with $4.4B in operating income. That’s a ~39% operating margin on a business growing 50%. Q1 2026 revenue was $3.3B, up 32%. Its a profitable, fast-growing, hard-to-replicate utility and has a big TAM ahead of it.

And Musk sums up the importance of Starlink to the overall company well:

C. AI

The AI segment is xAI, which was founded in 2023 and folded into SpaceX in early 2026 (the merger closed February 2). It houses Grok, the X platform, data licensing, and the compute infrastructure.

There are two things going on with this segment: the shift of the business from an ads business when X was bought to more of a subscriptions based business, and the additional revenue that comes from Grok (via API, etc) which is still relatively small compared to leading AI labs.

On the surface, it looks rough. “AI” revenue (which includes ads) was $3.2B in 2025, up only 22%, with a $6.4B operating loss as R&D rose 331% to $5B. In Q1 2026 it was worse on a growth basis: revenue of $818M, up just 12.5%, with a $2.5B operating loss, and ad revenue continued to fall in Q1. Below is a rough overview of the current state of the AI segment.

So if you stop at the income statement presented, the AI segment is a slow-growing, deeply unprofitable drag on the company. But that’s the picture before the recent deals which change things.

II. Recent developments in the AI segment

Two things happened in May 2026, right before the filing, that paints SpaceX in a somewhat different light.

The first is the Anthropic deal. SpaceX signed Cloud Services Agreements with Anthropic for compute capacity across its Colossus and Colossus II data centers. Anthropic agreed to pay $1.25B per month through May 2029, with capacity ramping at a reduced fee in May and June 2026. Either side can terminate on 90 days’ notice.

The entire AI segment did $3.2B in 2025. So this one contract, fully ramped at ~15B/yr, is roughly 5x the segment’s trailing revenue.

The second is the Cursor deal. In April 2026, SpaceX signed a compute and option agreement with Cursor. Under the compute agreement, SpaceX gives Cursor GPU capacity and they collaborate to improve Grok and potentially jointly develop new models. Under the option agreement, SpaceX has the right (not the obligation) to acquire Cursor outright at a $60B implied equity value, exercisable in a 30-day window after the IPO (or by September 30, 2026).

SpaceX frames coding as one of the best use cases for AI because it generates high-quality, verifiable data and constant inference demand. Cursor’s developer interaction data (prompts, iteration cycles, architecture decisions) feeds back into model training, including Grok. And if they exercise the option, Cursor becomes an owned application sitting on top of their own compute and own models.

If SpaceX terminates the option, Cursor is owed a $1.5B termination fee plus an $8.5B deferred services fee. So they’ve effectively pre-committed to either buying Cursor or paying a large breakup cost.

Put together, these deals give the AI segment a much clearer shape. Three layers that are compute, model/intelligence, and apps.

Compute is the base, and right now it’s the most valuable part. SpaceX built Colossus and is training Grok 5 on Colossus II. The interesting move is that they have shown willingness to lease the spare capacity. Anthropic is the first big tenant but there may be more.

Models is Grok, trained in-house, differentiated by real-time integration with X data. This is also where the Cursor collaboration sits, jointly developing models and feeding coding data back into Grok’s training.

Apps represents X and the Grok app today, and is the layer that the Cursor option can also bolster. If they exercise, Cursor becomes an owned coding app running on their own models and compute. They’re also building Macrohard with Tesla, an agentic platform meant to emulate digital workflows.

III. The financial picture

At the consolidated level, 2025 looked like this: $18.7B in revenue, a $2.6B loss from operations, and $6.6B in Adjusted EBITDA.

The capex split is quite stark, compared to the revenue split. Of the roughly $10B SpaceX spent in Q1 2026, about $1B went to Space, $1.3B to Connectivity, and $7.7B to AI. For the full year 2025, around 60% of capex went to AI.

The picture of spend and profitability relative to revenue growth isn’t great, which is why I think the Anthropic (and Cursor deals) are critical.

Annualize the Q1 2026 quarter and SpaceX is running at about $18.8B. Grow each segment off that base for the rest of the year and you conservatively get to roughly ~$22B run-rate by Q4.

Then layer in Anthropic and Cursor. Anthropic at the full $1.25B per month is $15B annualized. Cursor could generate another $3B-$4B in revenue (it’s at ~$3B today and growing). Stack those on the organic base and you land around a $40B to $41B run-rate exiting 2026, on a relatively conservative estimate of organic growth and not assuming more compute partners.

At least partially this helps ground the $1.75T valuation in some more reasonable sense. Compared to a ~20B run-rate business growing relatively slowly, it now looks like a ~40B+ run-rate business growing over 100% y/y albeit because of one compute deal. More importantly, it also shows Elon’s willingness to make sure that he monetizes his spare compute, even if internal products don’t take off which helps alleviate some of the concerns around the high Capex.

IV. Orbital compute and the space economy

Despite the improving AI segment in the near term, it still looks bolted on to the Space and connectivity business. Orbital AI compute is where that changes and the reason these are all in one company becomes more clear.

One way to think about SpaceX may be less a rocket company with an AI business attached, and more of a physical infrastructure company for the AI age. The launch business lowers the cost of accessing orbit. Starlink proves SpaceX can operate distributed infrastructure in orbit at global scale. The AI segment creates the demand for compute. Orbital AI compute are where those three pieces converge.

What is orbital AI compute you may wonder? Satellite constellations that act as data centers in space, using solar energy for power and the space environment for cooling. SpaceX expects to start deploying these as early as 2028.

The launch (Space, Starship) segment allows SpaceX to put that much mass in orbit cheaply. The satellite expertise from Connectivity gives them some of the know-how to do so (and frankly helps fund some of it), because orbital compute is basically the Starlink playbook pointed at a different payload.

And the AI segment is what monetizes that compute, either by selling that compute or using it in models or application.

The bottleneck for terrestrial AI is power and cooling. In orbit you get abundant solar and free cooling, if you can get the mass up cheaply and move data back down. SpaceX is arguably the only entity on Earth today that can pull this off, given their Space segment and know-how from Starlink.

It is interesting that SpaceX’s own TAM slide puts it at $28.5 trillion, and $26.5 trillion of that is AI, highlighting how important the orbital data centers may be into unlocking the AI TAM (thought SpaceX still needs to work on the AI demand piece).

There’s also a bigger vision in the filing, which is that Starship in the future could be capable of landing massive cargo on the Moon which then opens up factories on the Moon could manufacture millions of AI compute satellites using lunar resources and deploy them farther into space. The future markets they list include space manufacturing, lunar and Martian energy production, and asteroid mining.

V. The Bet on Starship

SpaceX’s future vision all hinges on Starship. The filing even says so directly: the Space growth strategy “depends on the successful development of Starship at scale.”

Falcon is proven. Falcon 9 carries about 23 tons to LEO, has flown roughly 620 times with 99%+ success rate. Falcon Heavy carries about 64 tons. But it won’t allow SpaceX’s full ambitions. That requires Starship

Starship V3 is designed to put 100 tons into orbit fully reusably, with aircraft-like turnaround, and future versions aim to double that. This is important for two reasons:

• Payload: you can’t deploy the next generation of hardware on Falcon. The new Starlink V3 satellites (1 Tbps of downlink each, versus the current generation) are designed to launch on Starship, starting in the second half of 2026, and the orbital compute satellites need that mass capacity too.

• Cost: full reusability with rapid turnaround is what collapses the cost per kilogram to orbit, which is the input to every other ambition.

So what’s the current status you may wonder? They’ve run 11 flight tests so far, with a 12th scheduled that debuts the next-generation vehicle and Super Heavy booster. They’ve already pulled off the “chopstick” booster catch, where the launch tower’s arms catch the returning booster, which is what enables rapid reuse and eventually multiple launches per day. They expect Starship to begin delivering payloads to orbit in the second half of 2026.

The investment behind this is large and concentrated. Space-segment R&D was $3B in 2025, most of it Starship, and another $930M in Q1 2026 alone, and arguably makes the financials of the core business (ex AI) look worse than it is.

Every ambitious plan of the company runs through Starship. Starlink’s next-gen satellites need it now. Orbital compute needs it for cheap mass. The Moon and Mars plans are impossible without it.

VI. Musk’s pay package

I’ve mentioned a few times that SpaceX is unique and inspirational. Nothing makes the scale of their ambition more clear than Musk’s pay package.

Musk has two performance awards in there, which tell a lot.

The first, granted in January 2026, is 1 billion performance-based shares. Both of two conditions must be met for any tranche to vest. First, market-cap milestones across 15 tranches, running from $500B up to $7.5 trillion. Second, and I’m quoting the filing:

The Company’s establishment of a permanent human colony on Mars with at least one million inhabitants.

So even if SpaceX becomes the most valuable company on Earth, the full award doesn’t vest until there’s a city of a million people on Mars.

The second, granted in March 2026 was tied to xAI and is called the “AI CEO Award” Same two-part structure: market-cap milestones (12 tranches, $1.065 trillion to $6.565 trillion), plus a second condition which I quote again:

The Company’s completion of non-Earth-based data centers capable of delivering 100 terawatts of compute per year.

For reference, total US electricity generation capacity is roughly 1 terawatt. So the target written into the filing is space-based compute at something like 100x the entire US power grid.

It gives you a sense of the scale of SpaceX’s ambition. Musk gets additional shares if and only if the company becomes multiplanetary and/or builds compute off Earth.

Closing Thoughts

This has been a long piece and there’s a lot more one can say about SpaceX. But to summarize, the launch business is a strategic asset, which is increasingly mostly being used internally to power a strong Connectivity business which is growing fast at 10B+ in scale and very profitable, but is being used to power investment in R&D for Starship and the AI business. The AI business has shown its ability to build infrastructure, which they are now open to leasing, and are trying to build better models to compete at the model and application layers as well.

Any forecasts or analysis based on current or the end of 2026 run-rate will make it difficult to justify the ~1.75T valuation, but SpaceX is arguably the only company today that can help build the Space economy and all that that may entail from AI compute to mining materials to making us an interplanetary species. So how much value do you give to that? And to Elon Musk being Elon Musk.

In closing, here is a particularly profound segment you don’t see in most S-1s:

For the entirety of its existence, human civilization has lived on a single celestial body: Earth. The current paradigm, in which human civilization is confined to one planet, exposes humanity to existential threats that are unpredictable and uncontrollable on a planetary scale. By moving beyond the only home we have ever known, we ensure species-level redundancy and that the light of consciousness will not be tied to a single planet subject to the inevitable hazards of a harsh and vast universe. We do not want humans to have the same fate as dinosaurs.

If you have any comments or thoughts, feel free to tweet at me.

Subscribe now

Hi friends,

Unitree Robotics recently filed for IPO on Shanghai’s STAR Market, looking to raise $620 million. The filing is quite interesting1 because it gives us a good sense of the current state of the robotics market.

Unitree is profitable, growing fast, and has shipped more humanoid robots than anyone else in the world.

In this piece, I’ll discuss:

• What Unitree makes

• The humanoid flip in revenue mix

• Who is buying robots today (and why)

• The vertically integrated approach

• The financial picture

• The model layer ambitions

I. What Unitree Makes

Unitree was founded in 2016 in Hangzhou by Wang Xingxing, a self-taught roboticist who famously built his first quadruped robot in his apartment. The company has 480 employees, roughly 175 in R&D.

It sells two product lines:

• Quadrupeds: the Go2 (consumer and research), B2 (industrial), and A2.

• Humanoids: the H1, H2, G1, and R1. The G1 is the one you’ve probably seen in viral videos, standing 1.32 meters tall and weighing 35 kilograms.

The company has been selling internationally since 2018. More than 35% of revenue comes from outside China, including a significant US academic customer base.

II. The Humanoid Flip

Two years ago, Unitree was basically a robot dog company and was primarily selling quadrupeds. Humanoids represented just 1.9% of its revenue in 2023.

By the first three quarters of 2025, humanoids accounted for over half of core revenue.

What drove it was a combination of product-market fit and aggressive marketing. The company’s humanoids performed at China’s CCTV Spring Festival Gala, one of the most-watched broadcasts in the world, for two consecutive years. Jensen Huang put a Unitree robot on stage at GTC in 2024.

The brand exposure converted into commercial and research demand in a way that most Chinese hardware companies have never really managed.

The shipment numbers for humanoids are particularly impressive when compared to others. Unitree shipped roughly 5,500 humanoid units in 2025, making it the largest humanoid maker by volume. AGIBot in China comes closest. As a point of comparison, the numbers for well-known US companies such as Figure AI, Agility Robotics are in the 100s if that

The 5-year target in the prospectus is 75,000 humanoids and 115,000 quadrupeds annually. That’s roughly ~14x the 2025 humanoid volume. It’s ambitious, but also highlights how early we are in the journey.

III. Who Is Actually Buying Robots Today

The prospectus breaks down buyers into three categories: research and education, commercial and consumer, and industrial applications.

The stark reality is most humanoid demand is for research and education use cases today.

1/ Research and education accounts for 74% of the revenue/shipments for humanoids. The academic buyer has been Unitree’s anchor since at least 2022 and remains the biggest source of aggregate revenue for the company. Researchers

2/ The commercial and consumer category accounts for 17% of shipments for humanoids. Non-academic consumers who buy these robots are mostly deploying them “for show”: as attractive promoters in retail settings, at tourist sites, in performances and exhibitions. Consumer revenue nearly quadrupled year-over-year in the first nine months of 2025, which sounds impressive until you realize the starting base was quite small. The real-world use case for the $25,000 humanoid robot, today, is apparently standing at the entrance of a store in Shenzhen to attract visitors.

3/ Industrial applications accounts for only 9% of shipments for humanoids. Unitree acknowledges that industrial deployment is more limited because the technology is less mature, highlighting the state of the technology today. Of this 9% of shipments, about 50-70% of it is for use cases like enterprise reception and tour guides, and so in aggregate only 3-4% of humanoid shipments are for things like enterprise reception and inspection.

On the quadrupeds side, things are a bit more promising: only about 1/3 of the revenue is from research, with over 40% from commercial use and the rest from industrial use. There, the productive use cases are more well established. Customers include State Grid, China Southern Power Grid, PetroChina, Sinopec, Baowu Group and JD.com (which is Unitree’s largest customer). These are companies that use the quadrupeds for real inspections of chemical plants, substations, coal mines, pipelines, etc.

If you don’t yet receive Tanay’s newsletter in your email inbox, please join the 10,000+ subscribers who do:

Subscribe now

IV. The Vertical Integration Approach

One of the unique aspects of Unitree is that it self-designs and manufactures most of its critical components: high-torque motors, precision reducers, encoders, joint modules, intelligent controllers, high-precision sensors, dexterous hands, LiDAR, and cameras. Actuation (the motors, reducers, and joint systems that actually move a robot) typically represents 40-60% of a humanoid robot’s total bill of materials, according to McKinsey.

Most companies in the space source these externally but Unitree builds them itself. Purchased components represent only about 14-18% of total costs. The only things it outsources are commodity parts such as the battery cells, flash storage and differentiated aspects such as the core compute board.

The per-unit manufacturing cost for a quadruped fell from roughly $3,300 in 2022 to about $1,800 by mid-2025, a 46% drop. Humanoid costs fell too, from about $10,800 to $9,200 over the same period.

Interestingly, the ASPs of quads and humanoids also fell significantly ever year as the graph below show. And yet gross margin expanded through this entire period, from the mid-40% range in 2022-2023 to nearly 60% in 2025, in big part because of their approach to vertically integrate.

V. The Financial Picture

Revenue grew from $58M in 2024 to expected ~$252M in 2025, a 335% jump owing to strength particularly on the humanoid side. International sales represented more than 55% of revenue for most of the company’s history. In 2025, domestic China overtook exports for the first time, though absolute export revenue still more than doubled year-over-year.

Gross margins are nearly 60% and have expanded over the years as below.

To put those margins in context: most hardware companies run 30-40% gross margins. Software companies often hit 70-80%. Unitree is relatively high for a company selling physical robots owing to their vertically integrated approach and the relatively differentiated products today.

The company turned profitable in 2024 on a GAAP basis and will have margins of around 18% or closer to 35% on an adjusted basis.

Unitree is targeting a ~$6-7B valuation at IPO.

VI. Model Layer Ambitions

Unitree is planning to spend nearly half the IPO proceeds on software. Of the $620M raise, roughly $300M is earmarked for AI model training over the next three years, about $100M per year devoted to what the company calls its “Embodied Large Model”.

The prospectus describes two parallel model architectures. The first is VLA (Vision-Language-Action): a model that maps directly from visual and language inputs to motor commands, letting robots generalize across unfamiliar tasks without hand-coded instructions. The second is WMA (World Model + Action), which they describe as their higher-conviction bet. A WMA model builds an internal simulation of physical reality. The robot predicts what will happen before it acts, rather than learning purely through trial and error.

They’ve shipped initial versions of both. In September 2025 they open-sourced UnifoLM-WMA-0; in January 2026, UnifoLM-VLA-0.

They also detailed the rough breakdown of spend towards the model, which is below:

Unitree’s hardware lead is real today, but the company understands that durable advantage in robotics probably requires owning the model layer too: the system that decides what the robot does and how it moves. The software ambition also makes sense as a hedge against commoditization. Unitree built its moat in hardware manufacturing.

But if actuators and joint modules eventually become standard parts, like batteries in EVs, the model layer is where the defensibility shifts to.

VII. Closing Thoughts

Unitree has a profitable hardware business, a real manufacturing moat, and more humanoid volume than anyone else at a price point others can’t touch. But the broad commercial adoption story is still in its early chapters as highlighted by how the humanoids are actually used. The “for show” use cases dominate consumer demand and Industrial deployment is narrow.

Unitree gives a glimpse of where the robotics market is currently with a lot more to come on the model side, hardware side and use cases side. If you’re building in the robotics and embodied AI space, feel free to reach out at tanay at wing.vc.

Subscribe now

Hi friends,

We’re past the thick of the Q1 2026 earnings season and like prior quarters, I’m going to round up some of the key themes I took away across earnings call in AI and software land.

I. Capex Keeps Going Up

The headline from every hyperscaler call this quarter: capex is accelerating. Not just for 2026 but for 2027 too. Just four companies combined will be spending over $700B: Amazon at $200B, Microsoft at $190B, Google at $180-190B, Meta at $125-145B.

Google CFO Anat Ashkenazi flagged that even that number keeps moving:

“We expect our 2027 CapEx to significantly increase compared to 2026” - Anat Ashkenazi, CFO, Google

Sundar Pichai shared the demand signal behind it:

“We are compute constrained in the near term. And as an example, our cloud revenue would have been higher if you were able to meet the demand.”

Microsoft guided to $190B noting that “strong customer demand across workloads, customer segments, and geographic regions continues to exceed available capacity.”

II. The Memory Chokepoint

There was a lot of discussion around Memory in particular as a key driver of constraint on the supply side, specifically HBM and high-bandwidth DRAM

Apple flagged it directly in its quarterly outlook:

“For the June quarter, we expect significantly higher memory costs...beyond the June quarter, we believe memory costs will drive an increasing impact on our business” - Tim Cook, CEO, Apple

Meta’s CFO called it out as the primary driver of their own raised capex number:

“Most of that CapEx increase is due to higher component costs, particularly memory pricing.” - Susan Li, CFO, Meta

Andy Jassy flagged the same thing:

“Everybody knows that the cost of components, particularly memory, has skyrocketed. We are in a stage where there is just not enough capacity for the amount of demand.” - Andy Jassy, CEO, Amazon

Intel CFO David Zinsner flagged the downstream risk:

“Constraints and rising prices around key components like memory, wafers, and substrates are driving higher costs that could impact demand for our product at some point in the year.” - David Zinsner, CFO, Intel

One underreported side effect: the memory shortage is accelerating enterprise cloud migration. Suppliers are prioritizing their largest customers, which means hyperscalers. Enterprises trying to procure on-prem can’t get equivalent access. Jassy on the call:

“A meaningful part of these suppliers are prioritizing their very largest customers, which cloud providers are. We have seen a number of conversations we have been having with enterprises for many months — where it has just been slower in getting the transformation plan to move to the cloud — accelerate rapidly just because we have a lot more supply than what others have.” - Andy Jassy, CEO, Amazon

III. The CPU is Back

For the past few years, every infrastructure conversation was a GPU conversation. That’s shifting. Intel, AMD, and Amazon all flagged that CPUs are re-entering the picture, specifically for agentic workloads.

Andy Jassy at Amazon discussed this trend:

“AI is commonly seen as a GPU story, but the rise of agentic workloads, real-time reasoning, code generation, learning, and multi-step task orchestration is driving massive CPU demand as well.”

Intel CEO Lip-Bu Tan also noted similarly:

“In recent months, we have seen clear signs that the CPU is reinserting itself as the indispensable foundation of the AI era. The CPU now serves as the orchestration layer and critical control plane for the entire AI stack.”

AMD updated its server CPU TAM forecast dramatically:

“We now expect the server CPU TAM to grow at greater than 35% annually, reaching over $120 billion by 2030” - Lisa Su, CEO, AMD

Both AMD and Intel shared similar perspectives on how the ratio may evolve. For training, you see roughly 1 CPU per 7-8 GPUs. For inference, 1 per 3-4. For agentic workloads: approaching 1 CPU per GPU.

IV. AI Improving Core Ads Businesses

One of the more under discussed stories is just how effective compute is at driving improvements in recommendation and ads systems, and its benefiting the core businesses of many companies.

Meta’s ads business is doing ~$56B in quarterly revenue, still growing 32% y/y and Google Search ads is doing ~$60B in quarterly revenue, still growing 19% y/y

For Google, search queries are at an all-time high, which is the opposite of what narratives from a year ago predicted. At the same time, the cost per click number is likely up significantly even though the exact percentage wasn’t shared. Google also noted that only about 20% of queries get monetized today and with AI mode and better ad recommendations, that percentage could move higher.

Meta saw a 10% lift in Reels engagement from ranking changes alone:

“On Instagram, the ranking improvements that we made in Q1 drove a 10% lift in Reels time spent. On Facebook, total video time increased more than 8% globally in Q1” - Susan Li, CFO, Meta

They also shared how their efforts around their Generative Ads Model for ranking is seeing ROI:

In Q1, enhancements we made to Lattice’s modeling and learning techniques along with advances in our GEM model architecture drove a more than 6% increase in conversion rate for landing page view ads

Pinterest and Reddit also posted strong quarters seeing strong growth in their ads businesses.

V. The AI-Generated Code Proxy

Multiple CEOs disclosed the share of their engineering output now coming from AI. I think it partly shows two things despite not being the best metric:

• Coding and Coding Agents has been the hero use case of AI that has been mass adopted across pretty much every enterprise

• Companies are using this metric at least externally to Wall Street as a proxy for where they stand in AI adoption

DoorDash’s Tony Xu gave the most direct quote:

“Well north of half of our code—probably closer to two-thirds of our code—is written by AI today.”

HubSpot’s Yamini Rangan shared both adoption and output data:

“100% of our engineers now use AI tools, and we have seen a 73% increase in lines of code updated per engineer.” - Yamini Rangan, CEO, HubSpot

Uber’s Dara Khosrowshahi specifically called out the rise of autonomous coding agents:

“About 10% of our code now that is committed is built by agents, autonomous agents out there.”

Spotify said they are “spending more compute per employee” and seeing “tremendous return in terms of productivity.” with internal metrics of output doubling.

For those curious, here’s a quick summary of some of the disclosures around what percent of new code committed is AI-generated:

• Chime: 84% (as of May '26)

• Google: 75% (Apr '26)

• Uber: ~70% (Apr '26)

• DoorDash: ~65% (May '26)

• Atlassian: >50% (Apr '26)

• Shopify: >50% (May '26)

• Snap: ~40% (Feb '26)

VI. SaaS Cos Layering on Agents

Among SaaS companies, the questions around a SaaSpocalypse and their ability to adapt to AI are a few things: with the shifts to agents, what happens to their core businesses built on seat-based models and will they or the labs / other upstarts give customers the agentic version of their offering that customers will want over time, which often looks like headless versions / MCP servers to be compatible with existing AI agents.

SaaS companies touched on their agentic offerings and also the seat compression piece. Atlassian discussed how their Rovo AI credit usage grew 20%+ month over month and that they aren’t seeing seat compression yet:

“We are not seeing any signal of seat compression from customers. If anything, we are seeing the opposite. We are seeing strong expansion numbers, strong cross-sell numbers between Collections, strong usage of AI.” - Michael Cannon-Brookes, CEO, Atlassian

HubSpot’s Dharmesh Shah touched on the agentic experience and the importance of going headless:

“We are big believers in the idea of headless - not big believers in the notion of humanless. The right platform for go-to-market for our customer base is going to be a combination of serving humans with a very personalized modern experience, and supplementing that with a really good agentic experience.” - Dharmesh Shah, CTO, HubSpot

Datadog framed their AI strategy in two explicit buckets: “AI for Datadog” (AI making the Datadog platform better through various agents) and “Datadog for AI” (Datadog providing observability and security for AI workloads).

In March, we launched our MCP server for general availability. With MCP Server, developers access live production data to debug their applications directly in their AI coding agent or IDE. We delivered this AI security agent, which autonomously triages Datadog Cloud SIEM signals, conduct in-depth investigations of potential threats, and delivers actionable recommendations. We’ve seen Bits AI security agent reduce investigations that could take hours to as little as 30 seconds - Olivier Pomel, CEO, Datadog

VII. SaaS Leaning on Context

SaaS companies are also leaning on their data and context as differentiation and a big part of their value add vs AI labs and other products.

Atlassian’s CEO:

“In a world where humans will run teams of agents, context is the only anchor to avoid chaos, and we believe that companies who prioritize context will become truly AI-native.” - Michael Cannon-Brookes, CEO, Atlassian

HubSpot’s Yamini Rangan and Dharmesh Shah talked about context as their common foundation:

“AI without the right context produces output; AI with the right context produces outcome.” - Yamini Rangan, HubSpot

“We are essentially ambivalent as usage shifts from human usage to agentic usage — whether it runs on our runtime agents that we have built or if there are third-party apps and agents that have been built. All of those agents are going to need a common foundation.” - Dharmesh Shah, CTO, HubSpot

VIII. The Roadblocks in Agentic Commerce

There was also a lot of discussion about agentic commerce and the changing nature of distribution, which is still very early and arguably has been a let down so far.

Amazon’s Andy Jassy gave the clearest explanation of why horizontal agents have struggled as shopping tools:

“The experience just has not gotten great with these third-party horizontal agents yet. They are not often able to get the pricing right or the product information right. They do not have any personalization data or any shopping history... I happen to think that if you are going to a particular retailer that you would like to do business with, you would like to shop from, if they have a great agentic shopping assistant, you are going to often start there because it is where you are doing your shopping. That is what we are aiming to make Rufus be — we are aiming to have it be the best shopping assistant anywhere.” - Andy Jassy, CEO, Amazon

Pinterest CEO William Ready also had a view on why third-party agentic commerce efforts have stalled:

“On agentic commerce more broadly, you have also seen meaningful strategic pivots from some of the platforms that were most aggressively pursuing that space. That validates our view that the barriers to progress in agentic were likely not technical, but around user behavior and ecosystem incentives.”

At the same time, it’s clear that AI is being used for top of funnel and mid funnel search and driving traffic to merchants (in commerce and beyond). Shopify’s view shared was the more optimistic one:

"The early signals on AI channels are really compelling. In the first quarter, AI-driven traffic to Shopify stores has grown 8x year-over-year, while orders from AI-powered searches have increased nearly 13x. And within this, new buyer orders are occurring at nearly twice the rate of other channels." - Harley Finkelstein, President, Shopify

Subscribe now

Hi friends,

AI has a public perception problem and increasingly the data on this is becoming somewhat embarrassing.

Half of Americans are more concerned than excited about AI in daily life, per Pew’s June 2025 survey, up from 37% in 2021. Only about 10% are more excited than concerned. An NBC poll last year clocked AI’s net positive view at 26%, which makes the technology less popular than ICE. Over $50B of data center projects have been blocked or delayed by local opposition over the last 18 months. There are over 100+ activist groups across the country specifically to fight data center construction.

I know this topic is a bit different than I usually cover but is important enough that it’s worth going into.

In this piece, I’ll cover:

• The shape of the backlash

• The issues with the current messaging

• The case the industry should actually be making

I. The shape of the backlash

The fear around AI is broadly around three things.

The first is jobs. Dario’s “AI could eliminate half of all entry-level white-collar jobs in 1-5 years and push unemployment to 10-20%” line has unfortunately becoming perhaps the single most-cited AI soundbite of the past year. In a Mercer survey of 12,000 workers globally, the share worried about losing their job to AI jumped from 28% in 2024 to 40% in 2026. A Resume Now survey found 60% of workers expect AI to eliminate more jobs than it creates this year, and 51% are personally worried about their own job.

The second is data centers and the environmental story. People associate AI with rising electricity bills, water use, and noisy industrial sites near their homes as the graph below shows. A Heatmap poll found only 44% of Americans would welcome a data center in their community, which makes them less popular than gas plants and nuclear.

The third is a more diffuse fear about AI itself. Safety, kids using chatbots, deepfakes, AI psychosis, autonomy and more generally “the unknown.” AI’s favorability rating is lower than ICE.

What’s striking is that this is the case despite ChatGPT being the fastest-adopted consumer product in history. People are using AI more but they just don’t trust it. And they don’t fully trust the people building it.

II. The messaging today

There’s a much more optimistic story to tell here but the story making rounds is a quite different one.

Dario’s pitch (and to be fair, parts of Sam’s) reads roughly like: AI will be the most powerful technology ever, half your jobs are gone, unemployment goes to 20%, and oh by the way, in the same breath, cancer gets cured and the economy grows at 10%. The optimistic part is real and discussed in Machines of Loving Grace (probably the best optimistic case any AI CEO has put in writing). But the TV soundbite that goes viral “half the white-collar jobs are gone.”

Axios called this “fear-profiting” in March. Namely, framing AI as immensely powerful and dangerous reinforces that only a few well-resourced labs can build it safely. That’s an effective fundraising message and a terrible consumer message.

The industry has spent two years training the public to believe two things at once:

1. AI will replace your job, soon, and there’s nothing you can do

2. AI is a generational technology that everyone benefits from

These are not inherently contradictory (they can both be true in the long arc) but as a pair of talking points delivered by the same people in the same week, they read as a threat. If you’re a 45-year-old radiologist or paralegal or junior accountant watching the clips, you don’t hear “the economy grows at 10%.” You hear “you’re getting fired.”

The data centers piece adds to it. The local political story for the average voter is my power bill went up, my water table is at risk, my kid’s commute got worse, and the people benefitting are tech billionaires who are also publicly bragging about putting my neighbors out of work.

The other aspect is that while AI is indeed a reason for some job displacement, many companies are blanket using that as an excuse for large layoffs stemming from periods of excess and overhiring.

Marc Andreessen noted this on a recent podcast with Harry Stebbings:

“The hiring binge that companies went on in COVID was just wild. Essentially every large company is overstaffed … Now they all have the silver bullet excuse, ‘Ah, it’s AI.’”

III. The more optimistic messaging

There’s a much better story sitting right there and it often is taking a backseat if not being ignored completely, including:

AI as the end of drudgery. Every job has a part of it that nobody actually wants to do. Filing claims, drafting boilerplate, parsing PDFs, answering the same email for the 200th time. The honest pitch is that AI eats the part of work people hate. This is also closer to what’s actually happening in most enterprise deployments. today.

AI as the end of dangerous work / human error. Mining, long-haul trucking, hazardous materials handling, certain kinds of agricultural and industrial work where injuries and deaths are still routine. About 5,000 Americans die at work every year, and most of them are not white-collar workers. Robotics plus AI is one of the few things that actually moves that number. It’s still relatively early in Robotics so this isn’t quite a major talking point yet, but can be one as Robotics improves. 40,000 people in America die every year from automobile accidents, and Waymo’s initial data shows ~90%+ decreases in accidents on a miles traveled basis. Stories of accidents prevented is a better one than a junior lawyer let go.

AI as actual democratization of expertise. Products like ChatGPT, initially at $20/mo and now basically for free have actually democratized access to infinitely patient tutors, a second opinion on any medical issue, a therapist and a life coach. That is genuinely new in the sense that these versions of expertise have always been gated and expensive. These products and use cases largely work today, but haven’t been the major messaging points compared to AGI timelines.

AI for discovery. AI is going to enable us to discover novel drugs, novel materials and cure diseases we didn’t think were curable. Three of the last two years’ Nobel Prizes went to AI work. The “AI for discovery” story is perhaps the most optimistic and most abundant view of the benefit of AI and has been largely underplayed compared to labor displacement. Yes, we’re still relatively early in terms of progress on AI scientists and AI drug and material discovery but is a story worth amplifying.

Closing Thoughts

Some of you may rightfully retort that people like Dario do talk about various things I’ve mentioned. But that’s primarily in his long-form essays, and not the most repeated points relative to existential risk and labor displacement. Another valid counterargument is that they are just doing their part to warn people about what’s coming, which is fair. But at the same time, it’s important for these leaders to paint the optimistic case of the age of abundance or risk us never getting to that outcome because of bans driven by the backlash.

The talking points in every interview should include the most optimistic messages of AI as well - discovery, democratization and no more drudgery and dangerous work.

AI is the most powerful technology of our generation and could be the most popular. But today only 26% of voters view it positively. The gap between those two is at least partially a messaging problem, and it’s fixable.

If you have any comments or thoughts, feel free to tweet at me or email me on tanay at wing.vc

Subscribe now

Hi friends,

A few weeks ago, I wrote about how AI application companies are increasingly going full-stack, integrating down into the model layer or up into the service layer. Since then, there’s been a lot of discussion around the pros and cons of going into the model layer and when the right time is, which will be the focus of today’s piece.

To be clear, this discussion is in the context of application-layer companies, not frontier labs. Very few of these companies are pre-training from scratch. Instead, most are post-training and RL on strong open-weights bases.

In this piece, I’ll cover:

• The training spectrum

• The case for doing it

• The case against doing it

• When it makes sense, and the new infrastructure that’s lowering the bar

I. The training spectrum

“Training your own model” gets used to describe wildly different commitments.

At the extreme end, there’s prompt engineering, RAG and harness engineering, which isn’t training in any form. A step up, fine-tuning a small model. Further up, supervised fine-tuning or RL on a strong open-weights base as the primary model in the system. Further still, continued pre-training on top of an open-weights model. At the far end, pre-training from scratch.

For application companies, basically nobody is at the far end but as you go right from there, you see examples at the different points. Cursor’s Composer 2 builds on Kimi K2.5. Intercom’s Fin Apex 1.0 sits on an undisclosed open-weights base. Cognition’s SWE-1.5 is, in their words, “end-to-end RL on real task environments using our custom Cascade agent harness on top of a leading open-source base model.”

As Intercom’s CEO put it, pre-training has become “kind of a commodity” and the action is in post-training. And so most of this piece will be in the context of post-training rather than pre-training.

II. The case for training

In my view there are three real reasons to invest in post-training.

1. Unit economics and latency

Once you’re at scale, API calls add up. A smaller specialized model that runs cheaper and returns in 200ms can beat a frontier call that takes 2 seconds and costs 10x as much.

Intercom’s Fin Apex 1.0 reportedly runs at roughly one-fifth the cost of frontier models, responds 0.6 seconds faster than the next-fastest competitor, and resolves customer issues at a higher rate. At the scale of a company like Intercome doing ~2M conversations per week, that gap in latency and unit economics is meaningful.

There’s a related destiny argument: if your unit economics depend on one frontier API, you’re exposed to pricing changes, rate limits, and the provider showing up in your category. No one has felt this more than Cursor, which despite tremendous success and usage, reportedly has -21% gross margins, and has found it difficult to compete with Claude Code when Anthropic’s limits supports thousands of dollars of equivalent compute in their $200 max plan.

2. Differentiation through proprietary data

If everyone is calling the same frontier API, where’s your edge? Increasingly, it has to come from the traces you’ve accumulated.

Cursor sees which completions get accepted and rejected. Intercom has billions of customer-service interactions. OpenEvidence has the queries and citations of 40% of US physicians and trained a domain-specialized model which captured data from peer‑reviewed medical literature via their partnerships. That’s proprietary training data and the best way to put it to work is through training or post-training.

In some ways, the real vaue of the application layer is that they get the best telemetry on the real use cases of their customer base, far better than any off-the-shelf evals can measure.

And by leveraging those to post-train models allows them to actually improve performance for their users while also getting the latency and unit economics benefits outlined above.

For example, Cursor’s Composer 2 was in part optimized on their internal eval Cursor-bench which they created based on their real traces. They felt this benchmark was more representative of the real work developers were doing within their platform than the publicly available benchmarks.

3. Specialized models for the parts frontier labs don’t prioritize

Most application companies aren’t just training one big custom model to replace the frontier. They’re running systems of small specialized models, each fine-tuned for one part of the pipeline that the frontier labs don’t optimize for.

Decagon talks about this: smaller fine-tuned models for query rewriting, routing, and intent classification, with a frontier model only where it’s actually needed. Sierra trained custom search models (Linnaeus and Darwin) inside a constellation that still uses OpenAI, Anthropic as core models. Cognition has SWE-grep for context, SWE-check for bug detection, SWE-1.5 for the main agent.

Most of the value is in the boring parts of the pipeline (voice activity detection, query reformulation, retrieval ranking, tool selection). None of these need a frontier-grade reasoner. All of them benefit from being faster, cheaper, and tuned to your specific data.

This is a great way to step into fine-tuning and post-training for many companies: train where the frontier underserves you and keep using the frontier where it doesn’t.

III. The case against

The biggest reason to be careful: your post-trained model may not survive the next base-model release from the lab. The labs are now releasing new models faster than ever, because the labs themselves are using their own models to build the next ones.

Anthropic’s Dario Amodei has said that 70-90% of the code for new Claude models is now written by Claude itself. OpenAI was even more direct in their GPT-5.3-Codex announcement (now 0.2 generations behind :)):

“GPT-5.3-Codex is our first model that was instrumental in creating itself. The Codex team used early versions to debug its own training, manage its own deployment, and diagnose test results.”

What this means: model releases that used to take months are now arriving weeks apart. OpenAI shipped GPT-5, 5.2, 5.3, 5.4, and 5.5 within months.

For an app company, that’s the biggest risk. A lot of fine-tuning wins from 2022-2024 dissolved when GPT-4 and Claude 3.5 came out, and the cycle is faster now.

That’s why in general, it’s much safer to post-train or fine-tune specialized models that work in a system of models rather than the core reasoning model for frontier tasks since the cost and latency benefits of those are more likely to survive even as base models improve.

There are other costs to consider as well namely that post-training talent is scarce and expensive. The opportunity cost of the talent and capital used to post-train models could be better deployed or leveraged in other aspects of the product and company.

IV. When to do it

One useful proxy is to train when you have enough proprietary traces to make a small specialized model meaningfully better than the frontier on a specific part of your pipeline.

I should note that the bar to start is also lower than it was a year ago, because a new infrastructure layer has shown up to support post-training in different forms.

• Tinker from Thinking Machine Labs is a managed post-training API. They handle distributed training and LoRA infrastructure while you bring data, algorithms, and environments. As Andrej Karpathy’s noted: it lets users keep about 90% of the algorithmic control while removing about 90% of the infrastructure pain.

• Prime Intellect’s Lab is similar, hosted RL training plus an open Environments Hub with hundreds of community-built RL environments.

• Applied Compute, founded by ex-OpenAI researchers, is a more white-glove version, post-training on enterprise proprietary data using RL environments and is one of many companies offering some version of RL-as-a-service.

• Vendors such as Mercor, Surge AI, Fleet and others sell custom expert-authored RL environments

• Lastly, the Chinese labs and their open-source base models that are competitive with the frontier closed source models serve as a starting-off point for many.

I do think the improved infrastructure has meant that even small teams of 10-20 can now post-train if they want to. But one mantra I like that still holds true for most application layer companies is “no GPUs before PMF.” If you don’t have the product or the traces yet, training a model to do that product is premature.

Post-training should be part of the conversation once a company is either rapidly scaling and has collected enough traces or at least has PMF and feels underserved by the frontier in some specialized aspects of their product which they feel the base model can fix.

Closing Thoughts

The companies integrating down into the model layer (Cursor, Intercom, Sierra, Decagon, Cognition, OpenEvidence) aren’t doing it because they like training models. They’re doing it because at their scale, with their traces, the economics and differentiation arguments finally pencil out. And almost all of them are doing it as post-training, not pretraining from scratch.

For most app companies earlier in their lifecycle, the honest answer in 2026 is: not yet, but start setting up to. Build the data collection now (traces, evals). Start with one small specialized model in a boring part of your pipeline rather than trying to replace the frontier on the main reasoning call.

The durable training investment is the data and environments you accumulate, which let you keep producing better models as the bases under you keep improving. And remember, those base models are improving faster than ever.

If you’re working on an AI application or thinking about this tradeoff, feel free to reach out at tanay at wing.vc.

Subscribe now

Hi friends,

I’ve been thinking about a pattern that we’re increasingly starting to see across AI application companies. Over time, I think most of them become “full-stack”.

At a high level, you can think about an AI product that achieves outcomes as having three layers:

1. At the bottom, the model

2. In the middle, the application or agent which includes the data/context, etc

3. At the top, the human or service layer needed to review/prompt/do the last mile to actually get to an outcome

This is obviously a simplified way to describe the world. Models sit on top of chips, etc and there’s a lot of infrastructure, data systems, orchestration, evals in building agents. But I still think the simplification is useful.

Traditional application layer companies would sit just in the middle layer. But these companies are increasingly beginning to (or starting off) vertically integrate in one of two directions. Some move down into the model layer. Others start or move up into the human or service layer. Both end up looking “full-stack1,” just in very different ways.

1. Full stack down

The first version of full-stack is when the application company integrates down into the model layer.

This is the pattern we are seeing in coding, customer service and a few other categories where the company has enough usage, enough proprietary traces, and enough economic incentive to start pulling intelligence in-house.

Cursor is a very recent example. Last week it introduced Composer 2, positioning it as a frontier-level coding model with meaningfully improved benchmark performance and lower pricing than many alternatives. In its technical report, Cursor says Composer 2 uses Kimi K2.5 as a base model, then extends it with continued pretraining plus reinforcement learning on long-horizon coding tasks.

Intercom made a similar move this week with Fin Apex. Intercom CEO Eoghan McCabe described it as “the age of vertical models” and Intercom says Apex now powers essentially all of its English-language chat and email customer conversations.

Cursor, Intercom, Cognition, Harvey, Sierra, and are increasingly trying to shape, tune, route, and in some cases train the intelligence itself around a specific domain.

Why does this happen?

The biggest reason is that for AI companies, this is the most important flywheel they have to drive increased performance. They have the traces of the agents. They see the prompt, the outputs, the edits/acceptances/rejections. Over time, that becomes extremely valuable proprietary training data. Better product creates more usage. More usage creates more traces. More traces make the model better. Then the better model improves the product again.

Other reasons include:

• Cost and Speed: Once you have enough scale, the COGS can add up, and smaller, fine-tuned models can end up giving you enough performance for your use case at much lower cost and at much faster speed.

• Differentiation: If everyone in a category is calling the same handful of models, it becomes harder to build real product distance from the intelligence layer alone.

2. Full stack up

The second version of full-stack goes the other way. Instead of integrating into the model, the company integrates upward into the human or service layer. It sells true outcomes instead of software.

This is the more classical full-stack startup idea that Chris Dixon wrote about in 2015. You do not just sell software into a workflow. You own the end-to-end process and provide an outcome or service to the end customer.

Historically, that often meant ugly services economics and challenges with scaling. But AI changes the shape of that equation and opens up the possibility to take this route in many end markets that weren’t attractive.

This idea of “sell outcomes, not software”, “services-as-software” and AI-native services has been growing in popularity over the last few years, but as models increasingly get better at agentic tasks over longer horizons, we’re now starting to see the first wave of these companies truly break out.

There are a number of examples of these AI-native Services business across various services categories such as legal, insurance, accounting, customer support, recruiting, and IT modernization with the extreme version of these the “roll-up and layer on AI approaches”. On the de-novo AI-native services businesses side, we see companies such as:

• Crosby AI combines software, AI, and attorneys to own more of the delivered result and outcome in what co-founder and CEO Ryan calls a “Neofirm”.

• WithCoverage and Harper building AI-native insurance brokerages

• Mechanical Orchard is an AI-native software modernization services firm

AI is great but it isn’t 100% there, and a company that owns the end result can both plug the gaps in all the missing places is both incentivized to contuosly improve and allows the customer to not have to think about the last mile.

This path is especially interesting because it may eventually absorb even more of the stack. What starts as application plus service can, over time likely post-train specialized models too. As AI does more of the work, some of these companies may end up owning two layers at first, then increasingly all three.

3. Closing thoughts

I think AI application companies will not tend to stay just AI application companies. Over time, they will vertically integrate. Many will use their traces to own the intelligence as well to increase differentiation, performance and reduce costs, particularly given that the model companies themselves are encroaching on some of their applications. The others will either start out delivering full outcomes or over time as they can do an increasing amount of the work, decide they can capture more value by selling the full service rather than an agent that does 90% of it.

For more related to this topic, here are some good pieces:

• Ryan Daniels (Crosby AI) on the Rise of the Neofirm

• Cursor’s Composer 2 Technical Report

• Eoghan McCabe (Intercom) on the Age of Verticalized Models

• Chris Dixon’s original piece on Full-stack Startups

Subscribe now

Hi friends,

We’re currently in the SaaSpocalypse. People believe software is dead and margins will compress to zero. Some are even saying that companies like Visa get bypassed and DoorDash gets aggregated away in the age of AI. Everything that looks like software becomes a commodity and no moats remain.

Before we declare the end of defensibility of all businesses, I think it’s worth grounding ourselves in the actual sources of defensibility that exist. My favourite book around defensibility and moats is Hamilton Helmer’s 7 Powers which outlines the common ways companies build defensibility.

The question is: In an AI world, which sources of power weaken, and which survive? Let’s walk through all seven, particularly in the context of software and technology companies.

1. Scale Economies

What it is

Scale Economies exist when larger volume reduces unit cost in a way smaller competitors simply cannot match, creating a cost advantage.

In traditional software and digital businesses, scale allowed for spreading investments in engineering, infrastructure, and sales across a large customer base, creating structurally lower cost per unit.

For example:

• Amazon Web Services spreads massive capital investments in data centers and infrastructure across millions of customers

• Netflix amortizes its 10B+ content budgets across over 200+ million subscribers, allowing it to have more volume of content and more niche content that competitors.

• Large SaaS vendors like Salesforce historically spread R&D and support costs across thousands of enterprise customers, reinforcing margin and reinvestment advantages.

What AI does

AI compresses labor-based scale advantages in software / digital work. A 20-person team equipped with agents can now build features, handle support, write documentation, and run experiments at a velocity that previously required much larger organizations.

However, at the infrastructure and model layer, scale continues to be important and a source of power.

Net result

Application-layer scale advantages that were based on spreading R&D and similar costs that across large user bases weakens.

Infrastructure-layer scale advantages still remain. In fact, Scale economies provide a source of power for model layer companies like OpenAI and Anthropic.

2. Network Economies

What it is

Network Economies arise when a product becomes more valuable as more participants join. There are two primary types:

• Same-sided network effects, where users benefit directly from other users being present as is the case on social networks and messaging apps.

• Cross-sided networks, where two distinct user groups create mutual value as is common in marketplaces like Uber and Doordash.

For example:

• WhatsApp became indispensable because every additional user increased the utility of the network for everyone else.

• DoorDash strengthened as more restaurants joined, which attracted more consumers, which in turn attracted more restaurants, reinforcing marketplace liquidity.

In each case, value scaled with growth of the network.

What AI does

Agents introduce frictionless multi-homing and can make it easier to simulate aggregating one side of a marketplace even if they didn’t exist. For example, a new food ordering service could use voice agents to call a restaurant not on their platform to order from them. Or a consumer-facing agent can arbitrage across marketplaces to find the best price.

That can weaken shallow exclusivity.

But AI cannot fabricate real-time liquidity, courier density, reputation history, or a canonical identity graph. Marketplace density and trust are structural, not labor-based.

Net result

Some surface-level network effects can weaken but deep liquidity networks with trust, reputation, and coordination density can still remain durable.

3. Counter-Positioning

What it is

Counter-Positioning occurs when a new entrant adopts a business model that incumbents cannot replicate without damaging their existing business. This often happens during technological shifts where new entrants “turn” the incumbents existing assets and business into a weakness.

For instance:

• Netflix’s transition from DVDs to streaming undercut traditional rental economics and forced incumbents into painful strategic trade-offs.

• SaaS companies offering subscription pricing challenged on-prem vendors reliant on upfront license revenue.

What AI does

AI creates new waves of counter-positioning. Startups can offer forms of usage outcome-based pricing instead of per-seat pricing. They can replace workflows with agents rather than augment users.

Incumbents may struggle to adopt these models because doing so cannibalizes existing revenue streams or disrupts organizational incentives.

Net result

Counter-Positioning is a powerful form of power for startups to get going in this current market. However, incumbents are more savvy than ever before and recognize that cannibalization remains needed.

In addition, startups may not be able to counterposition against all players vying to capture the new market such as AI labs, etc who are not tied to prior business models like incumbents.

In aggregate, counter-positioning remains a good source of power early on to get going.

4. Switching Costs

What it is

Switching Costs arise when customers face meaningful pain in leaving a product. In enterprise software, this historically included:

• Complex data migrations

• Rebuilding custom integrations

• Retraining teams on new workflows and systems/UI

It was arguably the primary source of power many software companies seeked to create.

For example:

• Salesforce embeds itself deeply into sales processes, capturing years and years of data, making migration costly and risky.

• Workday runs payroll and HR compliance systems that are tightly integrated into regulatory reporting.

What AI does

AI directly attacks labor-based switching costs. Agents can map schemas, rewrite integrations, generate training materials, and even run systems in parallel to reduce migration risk. What once required months of consultants may compress to weeks of automated orchestration.

In addition, if agents do the workflows, human workflows may not be as relevant and human AIs may evolve to simpler natural language or change regardless, meaning that workflow as a switching cost may also drastically reduce.

Net result

Switching costs for software businesses weaken meaningfully. Companies can more easily adopt new software, get their data in, customize UIs bespoke to their workflows quickly and relatively cheaply.

Some levels of risk remain as a source of friction that retains some level of switching costs, but most companies will no longer feel held hostage by vendors.

If you don’t yet receive Tanay’s newsletter in your email inbox, please join the 10,000+ subscribers who do:

Subscribe now

5. Branding

What it is

Brand reduces evaluation cost. It acts as a shortcut for trust, quality, and reliability.

In enterprise markets, brand often signaled safety. No one got fired for choosing IBM or Microsoft. In consumer markets, brand shaped preference and habit in a noisy world full of unclear claims.

Examples include:

• Microsoft, whose enterprise brand reduces procurement friction across software categories.

• Consumer brands like Nike, where emotional attachment drives durable preference.

What AI does

If agents continuously benchmark products on performance and cost and can do deep research based on individual needs of users/businesses, brand as a heuristic shortcut weakens. Evaluation can become systematic rather than reputational. This may be particularly true for consumer brands where performance and functionality matter more and brand was a shortcut for that than say brands perceived as luxury.

At the same time, AI also introduces new forms of risk: model unpredictability, hallucinations, security concerns. In high-liability contexts, institutional trust becomes more valuable.

This leads to the notion of brand splitting:

• The strength of marketing-driven brand compresses in areas where price/performance tradeoffs matters and agents can thoroughly evaluate that

• Institutional reliability brand persists and may even strengthen in a world where stakes remain high and risks associated with AI could even increase.

Net result

Some forms of brand weaken significantly since deep and thorough evaluations become possible, even for purchases where they didn’t make sense prior.

Other forms of brands persist or even strengthen as trust and accountability become even more important.

6. Cornered Resource

What it is

Cornered Resource exists when a company controls an asset that competitors cannot access.

This may include exclusive data, regulatory licenses, distribution control, or proprietary intellectual property such as patents in biotech/pharma. In software, the most common form of cornered resource has been proprietary data often collected and aggregated via customers directly over long periods of time.

Examples include:

• S&P Global / Moody’s / Fitch have cornered resources in the form of credit ratings, indices, and reference data that are written directly into regulations, covenants, and mandates

• CoStar in commercial real estate has a hard‑won proprietary dataset of listings, comps, historical deals, building attributes, and contacts that brokers and landlords rely on

• Helmer uses the example of Pixar which had a once-in-a-generation cluster of creative talent as a cornered resource

What AI does

AI increases the value of proprietary data because models improve with exclusive signal. At the same time, especially for datasets that were public in some form, AI means that they are no longer “cornered” since they can now be acquired much more cheaply than in the past by leveraging LLMs to scrape / structure those datasets.

So if exclusivity is real and structural, AI strengthens this power. But for some companies that thought they had proprietary data (e.g., data businesses where data was available online) which were not truly proprietary, it shines a light on them not having as much defensibility as they thought.

Net result

True forms of Cornered Resource becomes even more important in an AI-native world. Proprietary data (or rails) and what is enabled with that increases.

7. Process Power

What it is

Process Power arises from deeply embedded organizational routines that compound over time giving companies a durable advantage over in areas such as speed, product quality, cost advantages, etc.

For example:

• Toyota’s Production System is the prototypical examples that allowed allowed them to have manufacturing lines that improved continuously and produced cars at extremely low defect rates.

  

• Netflix’s data-driven commissioning process allows it to greenlight content with confidence.

• Meta’s large scale product experimentation and growth process that allows for continuous improvement of their products for the metrics they care about via large scale experimentation and A/B testing that was iterated on over a decade.

These routines create compounding advantage and provide a source of differentiation over others.

What AI does

AI commoditizes some forms of process advantages that happen in the digital world (product development, etc) because it allows companies to iterate faster than before and allow agents to iterate on their behalf (which arguably have best practices and experience in them via training).

However, data that is proprietary to a company does not become available to competitors, so any source of process power that depends on using that data remains.

At a baseline, the core processes that are digital oriented in any business likely improve because of agents, making it more difficult for competitors to have a delta over them. But TSMC’s process power for example still remains.

Net result

For the companies that had true process power based on 1p data / institutional knowledge / compounding intelligence, their process power will remain but potentially weaken as the competitors may improve. At the same time, by integrating AI into their feedback loops, they could drive even further improvements in their process, and so can stay ahead.

Generic process advantages will weaken.

Closing Thoughts

The powers and what happens to them are summarized above. If you’re building an AI-native company, the reality is that default software moat over the last few decades, switching costs, is likely not sustainable. This may actually help you replace incumbents in that it may be easier to migrate customers over, etc, but it hurts your ability to remain sticky.

In this world, defensibility shifts toward cornered resources (typically proprietary data), real network effects (when your line of business supports them), and compounding process power (being extremely AI-native is a good one to start with). Over time you can also built a trust-based brand which also will remain a source of differentiation.

If you’re interested in discussing further or have thoughts or feedback, I’d love to hear from you!

Subscribe now

Hi friends,

Over the last year, we’ve clearly entered a new era where models improve less from static human data and more from real experience. By “experience” I mean these models taking actions in environments, observing outcomes, getting feedback, and then getting better. When you look at what the frontier labs are doing, it increasingly seems like scaling reinforcement learning is how that experience gets converted into improvement.

In this post, I’ll discuss this new paradigm, the emerging stack around it and the open questions at play.

1. The Era of Experience

We were in the “human data” era where models learned from tokens from both the internet and expert data specialists. We trained huge models on massive datasets, and a lot of capabilities emerged from learning over large amounts of language and code.

Now, in the “experience” era, models learn from attempting tasks and receiving feedback in environments that mimic real tasks that we as humans may face. A model is taking a sequence of actions across tools and time.

Silver and Sutton of DeepMind, in their “Era of Experience” paper describe it as: the next generation of agents will learn from streams of interaction, take grounded actions in real action spaces, receive grounded rewards from consequences, and increasingly plan in the currency of experience.

Perhaps the clearest evidence of this is the shift in how labs allocate compute and spend on data. On the compute side, Dan Roberts, an OpenAI researcher, talks about compute going towards RL scaling and talks about a potential world where compute on RL scaling continues to grow to even beyond pre-training.

We added some reinforcement learning compute for o1 … o3 maybe had a little bit more RL compute. At some point in the future maybe we’ll have a lot of RL compute and then at some far point in the future maybe we’ll have just like be totally dominated and crushed by RL compute - Dan Roberts

Similarly, on the data side, the labs have put over dozens of vendors in business in the last year who sell environments to the labs that mimic real world tasks of some form that can be used to improve model performance on these tasks through RL. More on that later in the piece.

Additionally, Brendan Foody, CEO of Mercor notes that RL is getting so good that it can saturate any eval.

RL is becoming so effective that models will be able to saturate any evaluation. This means that the primary barrier to applying agents to the entire economy is building evals for everything. However, AI labs are facing a dire shortage of relevant evaluations. Academic evaluations that labs goal on don’t reflect what consumers and enterprises demand in the economy.

Continued progress through this then hinges on continuing to have new sets of tasks packaged in the right environments that can be used for the models to “play” in and learn from experience in. An environment is the sandbox where the agent can act (100s of thousands of times) and receive feedback (in the form of rewards). In practice, it’s an instrumented version of a real workflow: a UI/application, an API surface or a codebase.

2. The RL “stack”

As RL continues to scale, first inside frontier labs and over time across enterprises and ISVs, a startup ecosystem has formed in a few concrete areas. At a high level, there are three categories of companies: Environment builders (who supply the “data”), RL infrastructure (infrastructure needed to run different RL jobs) and RL-as-a-service (serving those that want more reliable agents but lack the expertise). Let’s go deeper into each!

RL environment builders

These companies supply the form of data that matters for RL, particularly to the labs: environments where agents can take actions, observe outcomes, and get scored.

In practice, an “environment company” usually sells environments packaged as: a workflow surface the agent can act in (UI, API, codebase, toolchain), a task library so you can generate many attempts, an eval that scores the attempts, and logging so you keep full trajectories.

Most of these companies primarily sell to labs today, and tend to have an area of focus although over time will likely expand areas and also potentially sell to enterprises. Common focuses and examples include:

• Coding: Preference Model, Proximal, Mechanize

• Computer Use: Fleet, Deeptune, Habitat, Matrices

• Other forms of specialized knowledge work: Phinity (Chip design), Isidor (finance)

Many of these companies are scaling revenue rapidly, though it’ll be interesting to see which ones can continue to scale and remain flexible and adaptable over time and needs from labs evolve as well as actually provide environments that mimic real-world scenarios at the right level of difficult to drive model performance.

The “incumbent” data vendors to labs such as Turing, Mercor and Surge are also supplying environments, and indeed arguably have the highest share of the market, in addition to traditional forms of human expert data.

RL infrastructure

While the labs may have their own infrastructure built to manage RL, there are a number of companies that have built various forms of infrastructure to make aspects of RL simpler.

Teams that want to run their own RL loops need tooling for rollouts, data collection, evaluation, policy updates, gating, and iteration. They also need the underlying compute and serving stack to keep throughput high.

Startups building in this area include companies building tooling to simplify RL post-train models such as Miles from RadixArk and Tinker from Thinking Machines to all-in-one compute platforms like Prime Intellect or tooling to make building environments themselves easier such as Hud.

RL-as-a-service

There are many enterprises (and software vendors building agents) that have repeatable workflows for which models and agents don’t work out of the box but could benefit use RL.

To serve this group of companies, a number of startups have emerged that are essentially providing RL expertise and outcomes as a service. The pitch is usually: tell us your workflow / success metric, give us your data (agent traces, workflow data) and we will help do the work needed to post-train the model to give you better outcomes.

Most of these startups have built some level of infrastructure and tooling in-house which they pair with FDEs with RL expertise.

Early customers of these companies tend to be AI-native companies that have a core agent in production that they would like to improve. Over time, the same approach can extend to traditional enterprises for all workflows.

Examples of these companies include Applied Compute, Osmosis, and CGFT.

Applications

So far, most RL scaling has happened inside frontier labs. But given the improving infrastructure, I expect that number of companies to grow, and many AI applications that are sitting on real-world trajectories and outcome feedback in a valuable workflow will likely leverage that.

Coding is a great example of an application category with lots of these trajectories given the adoption. Cursor was able to use the feedback from their tab model to train an improved one using online RL, as they describe in this piece. This is how many application companies will actually see the benefit from the data they collect.

3. Closing thoughts

While we have a new vector of scaling that should continue to result in progress over the short to medium term, there are a number of open questions that remain. For one, how well will the models generalize to pick up new “skills” they are explicitly RL’ed on. Ex-OpenAI research lead Jerry Dworek had this to say:

"How do those models generalize? How do those models perform outside of what they've been trained for? ... How do models do tasks that you don't reinforcement learning on? And probably not that great. And this is... basically the remaining questions in the world of AI because what we train for... we are getting really really good at."

The other is around whether RL scaling enough, or do we need more beyond that, with continual learning being the most obvious gap today. Ilya notes that humans are continual learners and whether we need models to do the same:

"Where on the curve of continual learning is it going to be? ...You produce like a superintelligent 15-year-old that’s very eager to go... they don’t know very much at all... and you say 'go and be a programmer, go and be a doctor, go and learn.' So you could imagine that the deployment itself will involve some kind of a learning trial and error period. It’s a process as opposed to you drop the finished thing."

Regardless of whether RL takes us to AGI however, it’s clearly driving model improvement, and a fruitful area for companies to both build in in the infrastructure area and application companies to leverage over time. If you’re building anywhere in this stack, I’d love to hear from you.

Subscribe now

Hi friends,

ChatGPT is finally starting to look like a real consumer internet product, in the most obvious way: ads. If you’re going to offer a free plan at scale, you can’t subsidize inference forever absent another cash cow business. The question was always when OpenAI would do it, and what the first ad products would look like.

I want to walk through (1) what other AI surfaces have done with ads (2) what we know so far about OpenAI’s approach (3) what the initial opportunity might look like and (4) where this goes once “the user” becomes “the agent.”

Let’s get started.

I. Ads in AI today

OpenAI is by far the most prominent and at-scale surface to start AI surface starts out in ads, but it’s not the first one. A bunch of AI products already monetize various forms of AI surfaces via ads, including:

• Perplexity which has been explicit that advertisers shouldn’t influence answers, and the ad unit they shipped is basically “sponsored follow-up questions” plus paid media alongside an answer.

  

• OpenEvidence in medicine where intent is unusually high and budgets are concentrated (pharma, devices). They’re free to doctors and ad-supported, and they’ve been public about ads being a meaningful revenue driver, reportedly close to / in the 9 figures of revenue.

• Amp Free from Amp Code which offers a free agentic coding experience that is ad supported.

• Talkie (owned by Chinese AI lab Minimax) is a Character.AI style app where people can chat with different characters and generates revenue primarily through advertising (on the free tier) in the form of banners and interstitials and in-app purchases. Talkie notes that their eCPMs range from $2-$10.

The way to think about these early AI ad systems is simple: they’re still advertising to humans. The ad is shown to a person, the person clicks (or doesn’t), and measurement is ultimately downstream conversion and lift. The “AI” part mostly changes targeting (more intent signals because of clear queries/intent) and the canvas (the answer box can absorb a lot of context).

II. OpenAI’s Ads Plans

From what’s been reported so far, OpenAI is taking two approaches to monetize their free tier and capture more of the value they create:

1) Classic intent-based ad formats

The most obvious format OpenAI is starting with is intent-based ads within select conversation as part of the response but clearly separated from the LLM response). OpenAI has also been careful to communicate this fact that ads shouldn’t change the actual answers of the LLM and that they’ll be clearly labeled.

These ads will look more similar to Google’s ads in that targeting will be based on the intent of the queries themselves, but interestingly seem to have a component where you can click through to chat with the Business AI (the action upon click is more similar to “click to chat” ads on Meta). It remains to be seen if users and advertisers will prefer that to clicking through to the website.

The early chatter suggests high CPMs in the $60 range which is shown below (and pricing based on impressions rather than clicks or other forms of downstream performance), limited inventory, and non-self-serve buying to start, which is exactly how you’d expect this to launch. If you’re an advertiser, you’re paying for novelty, perceived intent, and scarcity, and in being early to shape the advertising product to what the ideal is over time (various ad formats, measurement, targeting, brand safety, etc). My guess is these premium CPM driven ads don’t scale and ChatGPT adopts a more classic performance-based model driving real results (leads, conversions, clicks) and charging based on those outcomes (on a per click basis).

2) Affiliate fees on native checkout

The other announcement was related to affiliate commerce. ChatGPT will start supporting direct native checkout of products across Shopify, Walmart and Etsy among other platforms. As part of that, OpenAI will take a 4% cut of any transaction for merchants that choose to allow native checkout (for now, just from Shopify merchants).

This method is aligned with merchants and doesn’t require adapting LLM responses, and only charges them for native checkout which is optional (they can still show up as they would in the response for no cost if native checkout is not enabled).

The importance of this is two-fold:

• Another monetization avenue: It provides another monetization path around the commerce-related queries on ChatGPT for purchases that happen directly on the platform, without having to directly sign up every brand as an advertiser. Brands are used to paying affiliates across channels and this allows them to view OpenAI as a paid channel for native checkout as they treat some other marketplaces such as Amazon, etc.

• Higher value capture of Commerce Ads budgets in future: Commerce is a very important vertical (typically 35%+ of ad revenue from Meta and Google) and so if native purchasing behavior on ChatGPT happens and improved conversion rates, that makes it easier to measure impact/outcomes and incrementally and later capture larger budgets but also more value per transaction from this universe of advertisers.

III. The OpenAI Ads Opportunity

Long-term, the Ads opportunity for OpenAI is massive and could even be a Google, Meta or TikTok scale opportunity ($50-200B). But to help contextualize the shorter term path, below is what the opportunity might look like in the short term across both avenues.

Affiliate commerce and native checkout is a relatively new phenomenon and so may take a while to get going and as you can see need aggressive adoption and annual spend to get to even a $1B business.

Intent ads in chat are relatively familiar to users, but may take OpenAI a year or two to build the required infrastructure to support self-serve advertising, measurement, the right targeting and so on to really command large budgets. As you can see below, relatively conservative assumptions based on the current scale can see this being a $2-4B business in the coming few yeas, depending on how many users stay on and continue to use the free plans. Long-term, the potential for it to be a $10B+ business even just in the US is quite clear, if they can get to high levels of ARPU.

IV. Ads for Agents

So far, we’ve been talking about ads shown to humans. I think a larger shift will happen when the interface isn’t “chat” but rather “an agent completing a task.”

If an agent is buying running shoes for you, it’s going to do the research anyway. You can’t really “persuade” it with a banner. Doing so would mean it chooses the advertisers interests over the user and users will not like that.

The way you could influence the agent is by changing the utility function in a way that is positive for the consumer.

That’s why I find Google’s direction around “Direct Offers” interesting. Direct offers allows merchants present exclusive discounts directly inside AI Mode when Google determines a shopper has high purchase intent. It’s essentially a merchant trying to convince or negotiate the agent to go with a given brand’s product to help drive the sale.

This is how they position Direct Offers:

“Advertisers should think of Direct Offers less like a standard ad and more like a salesperson negotiating a deal on your behalf. It allows you to offer special discounts to high-intent shoppers to turn browsers into buyers.”

If we stop using chat interfaces and start to delegating more decision and purchases fully to agents without being in the loop as much, my guess is ads will have to evolve on these surfaces to look like the above. Something that influences the agent, while still driving towards better outcomes for the user.

V. Opportunities for Startups

This could be a post in and of itself, but I figured it’s worth touching on what opportunities for startups this ecosystem presents1. We’ve already seen the rise of AEO/GEO companies like Profound and AirOps to support brands in AI channels organically, and the advertising/affiliate ecosystem brings forth new opportunities for founders:

• Monetization rails for AI surfaces: OpenAI will obviously build their own ad platform, but the long tail of AI services and chatbots may want to plug into an existing network that brings both advertiser demand and the SDKs for targeting, brand safety, formats, and measurement. Companies like Koah, Kontext, and Gravity are building in this category.

• An “AI-channel DSP” for brands: A buying and measurement layer that helps advertisers manage spend across AI surfaces as more of them support ads, similar to how DSPs helped coordinate spend across social platforms and CTV. The core product is cross‑channel budget allocation, creative management, and measurement and reporting built for chat outcomes (leads, conversions, purchases).

• Agentic commerce enablement for merchants: Infrastructure that helps brands support native checkout and “offers” across emerging agentic commerce standards, especially for merchants outside Shopify’s ecosystem. Think products that let merchants quickly support agentic commerce and native checkout with connectors for catalog, inventory, pricing, and payments, without forcing a full eCommerce replatform. I’ve previously written about other opportunities in agentic commerce.

V. Closing Thoughts

I’m excited to watch OpenAI’s execution over the next year around both these monetization avenues: classic intent ads and affiliate checkout. The first one is about building a real ads product, pacing supply, proving measurement, and doing it without users feeling like answers are getting “shaped” by budgets. The second one is about behavior change and whether ChatGPT actually become a place where purchases happen.

I’m also curious how quickly the rest of the ecosystem follows. Some players will copy the search-like ad unit, some like Anthropic may ignore all this completely and some will outsource monetization to companies like Koah and Kontext to get going faster. And then longer term, the agent shift could change the fundamental nature of what ads look like on these channels2. Once models start completing tasks end-to-end, the winning “ads” probably look less like banners / units to persuade humans and more like offers that make the user better off (price/discounts, guarantees).

If you’re building in and around this space, I’d love to hear from you at tanay@wing.vc

Subscribe now

Hi friends,

Two Chinese foundation model labs, Zhipu and MiniMax, went public over the past few weeks, with Zhipu becoming the first foundation model company to IPO globally. What stood out to me reading both filings is the different approaches they took to turn their models and capabilities into a real business.

Zhipu focuses on enterprise more, with a bulk of their revenue coming from on-prem enterprise deployments, while MiniMax is more product and consumer focused with multiple apps that leverage the models they build driving the majority of their revenue.

I’m going to break down both companies across product, business model, financials, and IPO reception1. Let’s dive in.

1) Product

Zhipu

Zhipu is a deployment-first model company. The core asset is the GLM model family, and the filing emphasizes a set of newer releases like GLM-4 and its variants, including lighter versions and vision-capable versions, plus newer iterations in the GLM-4.x line which are most popular for their coding abilities. Zhipu is very explicit that it delivers capability through public cloud, private cloud, and on-prem deployments, and it pairs that with professional services. In other words, the “product” is often the model deployment plus a set of services that make it usable inside a real enterprise environment. They talk about their “Model as a Service” platform, which is highlighted below.

MiniMax

MiniMax is a product-first model company that was founded in December 2021. They still talk about their core model suite as below, but what jumps off the page is the multi-modality and the consumer-facing surfaces they have built. The filing discusses their lineup of models that spans text models and models for video, speech, and music generation. But the way these are primarily delivered is via consumer apps - Hailuo AI, Talkie (and Xingye), MiniMax, and MiniMax Audio. MiniMax is trying to express model capability through native products that can acquire users and monetize directly, while also running an Open Platform that can serve developers and enterprises.

2) Business Model and Monetization

Zhipu

Zhipu monetizes primarily through enterprise solutions delivered via deployments and support.

They generated $8m in 2022, $17m in 2023, $44m in 2024, and $26m in H1 of 2025, and so went public quite early in their monetization journey, although it should be noted that many of their key model launches that put them on the map outside of China came in the back half of 2025.

Zhipu has two main offerings: An on-premise deployment where they can help either fine-tune models and host them locally within a customer’s own environment and a cloud offering where anybody, including individual developers, can use their models as a service via an API.

The on-premise deployment accounted for 84.5% of revenue. The economics of the two offerings are quite different. The on-premise deployment has only 95 customers and so much higher ACVs while the cloud offering is used by individuals and teams across thousands of enterprises. Additionally, their on-prem deployments have a gross margin of about 60%-65% depending on the quarter. Their cloud deployment is running on a gross margin of 0% to 5% depending on the quarter.

Given the on-prem model driving majority of their revenue today, it’s perhaps not surprising that their revenue is also quite concentrated. The top 5 customers represent about 40% of the revenue in H1 2025, with the largest being 11%.

MiniMax

MiniMax monetizes through a two-pronged strategy that balances AI-native consumer applications (B2C) with a B2B Open Platform for developers and enterprises. The company generated no revenue in 2022 during its initial R&D phase, but revenue grew to $3.5m in 2023, jumped to $31m in 2024, and reached $53m in the first nine months of 2025. Similar to Zhipu, MiniMax is entering the public market at an early stage of commercialisation.

While Mainland China accounted for 81% of revenue in 2023, the business has moved globally; by the first nine months of 2025, Mainland China represented only 27% of revenue, with Singapore and the United States contributing 24% and 20% respectively.

On the B2C side, the company operates four core applications that served over 212M users as of September 2025. Its largest revenue driver is the entertainment platform Talkie (Xingye), which features emotionally responsive virtual characters and recorded 20M MAUs and $19M in revenue for the 2025 period. The fastest-growing app, Hailuo AI, focuses on cinematic visual and video generation, contributing $17M. The suite is rounded out by MiniMax Audio, which provides speech and music synthesis and a specialized intelligent agent for professional tasks like coding and research. Interestingly, given the discussion about ChatGPT ads, it should be noted that their core app Talkie primarily monetizes via advertising, with Hailuo and others primarily monetizing via subscriptions and in-app purchases.

The economics of its two segments differ significantly. The AI-native products (B2C) segment accounts for ~71% of revenue but operates on thin gross margins of ~5%, having only recently recovered from deep negative margins as inference efficiency improved. In contrast, the Open Platform (B2B) segment, which represents ~29% of revenue, has ~70% gross margins. This platform provides a mix of usage-based cloud APIs and licensed deliverables that allow enterprises to deploy and operate models locally within their own internal systems.

3) Financial Profiles

It’s worth touching a bit on some of the financials in aggregate for the businesses. If you take the reported numbers given and convert them to annnualized numbers, the shape of the two businesses looks roughly as follows.

Revenue: MiniMax is a bit larger and growing faster, though Zhipu’s 2025 numbers may meaningfully accelerate in H2 ‘25 with their model launches in the back half of 2025.

• Zhipu went from $45M in 2024 revenue to ~$54M in annualized revenue

• MiniMax went from $30M in revenue to ~$71.2M in annualized revenue

Gross Margins: Zhipu has strong gross margins given it’s more enterprise / on-deployment focus.

• Zhipu’s aggregate gross margins were around 50% in H1 2025, down a bit from ~56% in 2024

• MiniMax’s gross margins were 23% in Q1-Q3 2025, up from 12% in 2024.

Losses: Both companies are burning a lot of money as one might imagine, with R&D and more specifically cloud services for model inference and training being the primary drivers of spend, at around 70% of R&D spend and R&D itself being about 65-80% of revenue for each company.

• Zhipu lost 415M in 2024, with true cash burn of ~$300M, and plans to burn ~550M in 2025

• MiniMax lost ~465M in 2024, with true cash burn of ~$260M and plans to burn ~340M in 2025

Cash: Both companies had raised over $1B prior to IPO and raised ~500M each in the IPO itself. Interestingly, at their high levels of cash burn, Zhipu in particular was down to <1 year of runway on their liquid cash and equivalents prior to the IPO.

4) IPO Reception

Both companies raised large IPO proceeds by any normal standard, and a small round by the standards of AI labs in the US that we’re routinely seeing these days.

Zhipu raised over $550M and MiniMax raised over $600M in their IPO. Both IPOs have fared really well, with Zhipu up almost 80% and MiniMax popping over 100% on its IPO day, with both pricing their IPOs in the ~$6.5B range.

Currently, MiniMax trades at a valuation of $15.6B, and Zhipu trades at a valuation of $11.7B. At these valuations, MiniMax and Zhipu trade at 2025 annualized revenue multiples of ~220.4x and ~215x respectively!

Seems like a good sign for all the various AI labs globally that perhaps want more capital or go public. In closing, here’s an infographic based on this post.

Hi friends,

I hope everyone is enjoying a restful holiday period. As we head into 2026, I’ve been jotting down a few areas in AI and technology that I’m paying closer attention to with some predictions as well. Let’s dive in!

I. Monetizing free AI surfaces

A lot of the most widely used AI products today still have an unresolved question sitting at the center: what do you do with massive free usage at scale?

Subscriptions work well for power users. Enterprise/APIs works well for teams. But neither fully addresses the reality that tens or hundreds of millions of people are using chatbots regularly, often for expensive queries.

I think 2026 is when we start seeing experiments of ads in the major chatbots such as ChatGPT, Gemini and Claude. They probably don’t take the form of ads inside answers, which would immediately undermine trust, but through more separated, carefully designed monetization. Sidebar ads. Related sponsored results. Affiliate-style ads on commerce.

OpenEvidence is reportedly already on a 9 figure run-rate from ads, and highlights that in many cases the commercial intent of these queries is extremely real and that you can’t just rely on monetizing the small fraction that subscribe to paid plans forever.

II. Proactive AI

Most AI today still waits for you to ask.

One of the trends I’m most excited about is the move toward more proactive behavior. Systems that monitor things on your behalf and surface updates, summaries, or drafts without you having to prompt them.

This doesn’t look like a general life assistant. It shows up in narrow, permissioned contexts where the AI already has the right signals. Daily briefs. Alerts when something changes. Work that starts before you ask for it.

We’re already seeing early versions of this with ChatGPT’s Pulse and Yutori’s Monitor, and I expect more products to lean into this over the next year.

III. Claude Code for knowledge work

Outside of software engineering, the dominant interface to AI is still chat. That’s fine for quick tasks. It’s a bad fit for work that is multi-step, long-running, or closer to delegation than prompting.

I think we’ll see something like the equivalent of Claude Code emerge for the typical knowledge worker in 2026. An agentic interface to manage tasks that can take a task, operate for a while, use tools, and come back with completed output rather than suggestions.

Manus is probably the closest thing in spirit today. But I think the category becomes more legible in 2026, even if it doesn’t converge into a single product. It may emerge role by role, the same way people talked about “Cursor for X,” but now in “Claude Code for X” form.

IV. Computer use as a first-class capability

Agents get much more interesting once they can operate the world as it exists today.

That means using browsers, navigating apps, filling out forms, clicking through workflows. Computer use has been an area of focus for a while and we have started to see improvements in benchmarks, but today, at least on true taking action, browser use and computer use are not still reliable enough to do most core consumer or enterprise workflows.

I think we start to see the payoff in 2026, with models that are meaningfully better at doing this reliably - operating applications, filling 10 page forms, etc.

I expect to see meaningful jump in computer use benchmarks, but more importantly in products that can make use of the capability to actually carry out work (beyond the toy book a flight example).

V. Continual learning breakthrough

One research topic that keeps coming up more in conversations is continual learning which is the capability of models to learn new tasks / memory / feedback without forgetting past ones.

Systems that can actually improve over time from experience and feedback aren’t quite there yet. Dwarkesh wrote recently about this being a meaningful blocker for more general intelligence which I agree with.

I don’t think this gets “solved” in 2026. But I do think we see a real leap from one of the frontier labs which also becomes available in the products we use and makes AI feel even more useful. An agent that gets better as it runs is fundamentally different from one that stays static forever.

VI. A blockbuster IPO

Finally, I’m watching the public markets. I think this year brings even more liquidity, in the form of all three, pseudo acquisitions, actual acquisitions and IPOs.

And maybe most importantly on the IPO front, my expectation is that we see at least one (if not more) of the $100B caliber companies file to go public in 2026. Yes, the ones that have been perma-private so far and raised boatloads in the private markets which we questioned whether they’ll ever go public (but not Stripe).

Closing Thoughts

Given how far we’ve come in AI over the last year, when reasoning models were barely getting going, 2026 promises to be another eventful and action-packed year. What are you keeping an eye on this year? As always, I’d love to hear your thoughts, and if you’re building in the first few areas discussed above, feel free to email me at tanay at wing.vc.

Subscribe now

Hi friends,

Most of us today use AI in a chat box. You type, it replies, everything happens in that one thread. That was the right starting point. It is the wrong place for a lot of the work we are now trying to hand off to agents.

Refactoring a codebase, watching the web for changes or building a briefing for tomorrow are not five second tasks. If you force those into chat and let it run for 10s of minutes, people inevitably get bored and leave.

For these tasks, The model should not be a slightly faster colleague who types back at you. It should feel more like a background process that you brief, that disappears for a while, and that hands you something useful where you actually work.

That is the shift I want to talk about: the rise of background agents, how they run, and where we already see them showing up.

What are background agents

A background agent is an agent you do not have to sit and watch. You might talk to it in chat, you might tag it in Slack, you might forward it an email, but once it understands the task, it runs elsewhere and comes back only when it has something worth your attention.

That already makes it different from a classic chatbot. In the chatbot world, every interaction is “you ask, it answers.” In the background world, the unit of interaction is a task. You brief the agent once, it may run for minutes or hours, and the result shows up as a pull request, a digest, a comment, or a notification.

On top of that, background agents can be:

• Ambient: They are always running in some sense, and responding to some changes in some input. They wake up because the world changed, not just because you typed a prompt in. A new email arrives, a web page changes, a metric moves, a calendar event appears.

• Proactive: They are allowed to tap you on the shoulder when there is something important, rather than waiting for you to ask at all. Think alerts, daily briefings, suggested decisions or actions (or at some point it may have taken these actions for you as well).

Not all background agents have to be proactive or ambient. You can have a very simple background agent that only acts when you tell it to. But once you accept that the work happens away from the chat window, it becomes natural to let agents respond to time and events as well as direct commands.

Chat is still useful. It is a great place to negotiate scope, explain edge cases, and set guardrails. It just does not have to be the only interface or the place where the heavy lifting happens.

What these agents actually look like

Broadly, background agents show up in two ways.

The first are the ones you explicitly create. Early on, that mostly happened in chat. You open something like ChatGPT and kick off a Deep Research request or a coding assistant, describe the task, and a background job spins up from that thread. Cursor or Codex style agents that take a spec, go off, and eventually return a pull request are good examples here.

What is changing is that these agents are now being created in the places where the work already lives. Instead of only briefing them in chat, you:

• Assign an issue in Linear or Jira to an agent instead of a person

• Tag an agent in a Slack or Teams thread and ask it to own that conversation

• Forward an email to an agent address which can ingest the context and get started

You are still creating a discrete job, you are just doing it from tickets, threads, and emails rather than only inside a chatbot.

The second style is more ambient and always on. Here you create the concept once, not each individual run.

That might look like:

• An agent that runs whenever new emails come into your inbox, monitoring your inbox and turning a pile of messages into a single summary and task list or drafting replies (e.g. Fyxer)

• An agent that tracks changes on competitor sites or pricing pages and hands you updates so you can react (e.g. Web Monitoring Agents such as Yutori Scout)

• An agent that watches or monitors a set of data for anomalies or changes and only surfaces something or takes action when it sees a meaningful pattern

Sometimes these run purely on events, sometimes they also run on a schedule, for example sweeping everything once a night and dropping a fresh brief in the morning (e.g. daily briefing agents that can be created on Lindy or Gumloop). Sometimes these can be derided as just cron jobs, but they also have LLMs baked in, and so are more powerful.

So you get two complementary patterns. Jobs you explicitly assign, often from the same tools you already use to assign work to humans. And agents that feel more like part of the environment, always running in the background, watching your inbox, data, or the web, and only appearing when there is something you might want to do about it.

Examples of background agents

To make this less abstract, here are three classes of background agents that already exist and feel meaningfully different from simple chatbots.

1. Coding agents as background teammates
   Cursor is probably the cleanest example. You describe a change, sometimes in chat or directly assigning it in Linear. Cursor spins up cloud agents that read and edit your repo, run tests, and open pull requests. The work might run for quite a while. When you come back, the real interface is the diff and the test results.

2. Monitoring agents for the web
   Products like Yutori’s Scouts or Exa-powered monitors allow you to create an ambient agent that continously monitors the internet for changes in things you care about. You say which companies, topics, or products you care about. The system spins up agents that continuously check relevant pages, feeds, or search results. When something meaningful changes, you get a short synthesized update. At some point, these agents may be able to take direct actions as well (buy item if price drops below certain amount, if competitor changes pricing, update our comparison page to reflect that).

   

3. Daily brief and inbox agents
   There is a growing set of tools that connect to your email, calendar, and documents and then run every night or as new emails come in. Fyxer is a great example: it monitors your inbox, and autocategorizes email and drafts responses in your voice. Another good example are various daily briefing agents that run at scheduled times overnight and review the next day’s calendar and create

You can imagine similar patterns inside companies: agents that monitor various forms of data and proactively take action.

Closing thoughts

This cycle started by putting a chat box in front of a model. The next phase is about letting agents running on their own time - it allows them to run for longer, always remain running/monitoring and even be proactive.

Background agents are the mechanism that makes long running, multi step, and recurring work feel natural. They may start from chat, but they do not have to live there. They can be user-created, or ambient and always on. They can be quiet most of the time, then proactive in short bursts when something actually matters.

My guess is that most of the real leverage from AI will come from those quiet loops: agents working in the background while we are in meetings or asleep, and then appearing, briefly, with something that is actually useful. Chat will still be there, but more as the way we brief and tune them, rather than the place where all the work happens.

If you have thoughts about this paradigm or are building various forms of these agents, I’d love to hear from you at tanay at wing.vc

Subscribe now

Hi friends,

Voltaire once wrote “This agglomeration which was called and which still calls itself the Holy Roman Empire was neither holy, nor Roman, nor an empire”

A similar confusion has crept into one of SaaS’s favorite metrics. ARR, short for annual recurring revenue, used to have a simple meaning: the predictable subscription revenue a company generates.

But lately, what some startups call ARR is neither annual, nor recurring, nor in some cases, even revenue. Let’s look at what’s going on.

1) When ARR is not annual

What’s going on

Many companies take monthly or cancel-anytime contracts (or cancel within 3 months with most customers in that period) and multiply the current month’s revenue by twelve to arrive at an ARR number, arguing that if customers keep paying next month, that’s what they’d earn in a year. The issue is that these customers haven’t committed to a year, so the figure represents potential, not commitment. We all know many of the budgets in AI today are experimental.

The other version of it is using peak spikes in a customers demand that likely won’t recur month to month and using that to annualize.

Why it matters

Annual implies time-bound commitment and predictability. When most of your base can churn in a month, the number behaves like annualized MRR, not true ARR.

2) When ARR is not recurring

What’s going on

One common stretch is project type revenue being bundled into ARR, which is especially true for service type models which have grown in the AI world that are often transactional/project-based. In some cases, those customers come back for multiple projects so it can feel recurring, but these are still discrete projects that must be re-won each time.

One of the other trickier ones is with the rise of usage-based models, whether to annualize the usage. The logic is that consumption repeats, so it must be recurring, which it may be, but in some cases there are clear spikes for certain reasons such as the first time to onboard, etc and the usage is likely to not recur in the future.

Finally, some companies are baking pilot revenue into ARR by including the revenue from the pilots themselves or sometimes the assumed size of the contract when it is signed (gulp).

Why it matters

Recurring means repeatable by contract or strong customer habit. When project work or speculative expansion is blended into ARR, it makes the metric lose meaning. Distinguishing recurring from repeatable or re-occuring helps shed better light on the business.

3) When ARR is not Revenue

What’s going on

Perhaps the most creative accounting happens here. Some companies are treating gross merchandise volume (GMV) as revenue, even though the company only earns a small take rate.

Others are annualizing pilot conversions, counting the potential revenue “if all goes well” as if it already exists, and in some rare cases, companies are including potential expansion figures which haven’t happened yet into ARR (again, gulp!).

Why it matters

Revenue represents money that has actually been earned by the company and is supported by delivery or performance. Counting unearned expansions or total GMV overstates reality, since those dollars may never materialize or do not belong to the company in the first place.

Closing thoughts

The pressure to show eye-popping growth is real. Screenshots of “$0 to $5M ARR in X months” and soundbites of “We only invest in $0 to $2M ARR in 10 days” travel fast, and it’s tempting to keep up.

If your business relies on short-term contracts, usage, or pilots, that’s fine. Many companies in the AI world are the same. Just label it honestly as annualized run-rate revenue or annualized MRR and remember that it is one of the aspects of the business to focus on and improve over time. Revenue quality ultimately decides who endures. Retention and gross margin matter more than vanity milestones.

Subscribe now

Hi friends,

It’s OpenAI’s world and we’re just living in it. Over the past few months, news tying a company to OpenAI, whether official partnerships or announcements or just chatter has moved stock prices of large companies across chips, cloud, commerce, and beyond.

OpenAI seems to be having a kingmaker effect in the public markets unlike anything we’ve seen before. Great for partners and those who can be aligned with their direction, terrible for those in the firing line.

Note that I won’t focus on the accounting treatment or circular nature of these deals, but instead on just how broad and impactful announcements with OpenAI are proving in the stock markets across the board.

Compute & Cloud

NVIDIA
On September 22, 2025, NVIDIA and OpenAI outlined a plan to deploy at least 10 gigawatts of NVIDIA systems for OpenAI’s next wave of infrastructure, alongside press about a large prospective NVIDIA investment linked to deployments. NVIDIA Stock surged 5% on the news, adding hundreds of billions in market cap.

AMD
On October 6, 2025, AMD announced a multi year agreement with OpenAI to stand up roughly 6 gigawatts of capacity, beginning with 1 GW targeted for 2H26, plus a warrant package for up to 160 million AMD shares for OpenAI. AMD rose almost 30% that day adding ~$100B in market cap.

Broadcom
On September 13, 2025, reports tied OpenAI to a roughly ten billion dollar custom accelerator program at Broadcom, and AVGO pushed to fresh highs into late September. On October 13, 2025, Broadcom and OpenAI formally announced a 10 gigawatt custom accelerator collaboration, and the stock surged nearly 10%, adding ~$100B in market cap.

Oracle
On September 11, 2025, coverage of a very large multi year OpenAI capacity agreement in the ~$300B range resulted in Oracle logging its biggest one day gain in decades, up over 35% adding over ~$300B in market cap.

Commerce

OpenAI’s agentic commerce announcements also saw a number of beneficiaries that have been early with them to partner to support agentic checkout in ChatGPT.

Etsy
On September 30, 2025, OpenAI launched Instant Checkout inside ChatGPT with Etsy as the first live surface. Etsy jumped about 16% on the news, adding ~$1B to its market cap.

Shopify
Around the same time as above, OpenAI said the same Instant Checkout surface would extend to Shopify merchants soon, implying better discovery and potential for higher sales and purchase volume over time for Shopify merchants. Shopify gained ~6% on the news, adding ~$10B in market cap.

Walmart
Earlier today on October 14, 2025, Walmart announced a partnership with OpenAI to enable shopping through ChatGPT using Instant Checkout. Walmart is up over 5% adding ~$40B in market cap.

Negative Moves

Moves aren’t always positive however as some of the examples below show

Reddit
On October 1, 2025, third party data showed ChatGPT referencing Reddit less than it had in mid September. Reddit fell roughly 10 percent on the day as the market repriced how it expected Reddit user growth, referral traffic and data licensing revenue to fair over time with decreased ChatGPT visibility, losing almost ~$1B in market cap.

Docusign
On October 1, 2025, DocuSign plunged ~12% immediately after OpenAI revealed its internal contracting agent, DocuGPT, despite it just being a demo about an internal tool. Investors got spooked about the threat to SaaS products like Docusign with LLM advancements. The drop translated into a ~$1.5B decrease in market cap.

Closing thoughts

OpenAI has become a market scale allocator of demand. One announcement of some relationship with OpenAI can send a stock soaring. One news article about OpenAI encroaching in some way can send a stock tanking. A private company’s roadmap has become one of the major inputs for the public markets writ large. The tweet below encapsulates it well.

Subscribe now

Hi friends,

ChatGPT released a really interesting paper last week that went deep into how it's being used by its over 700 million monthly active users over the last year.

About 73% of its usage is now for non-work tasks, with 23% for works tasks. But perhaps most interesting was the graphic below which broke down how people are using ChatGPT for different use cases and different “jobs to be done”. It clearly shows that ChatGPT has become an all-in-one assistant across personal and work tasks and multiple jobs to be done.

Some of these jobs to be done required others websites or tools to serve the same need, and some of these jobs to be done are things where there wasn’t a good parallel prior to the advancements in LLMs. Let’s discuss further.

The bundle that is ChatGPT

“There are only two ways to make money in business. You can bundle, or you can unbundle.” — Jim Barksdale.

Today, it’s clear that ChatGPT is basically used as an all-in-one tool that bundles and absorbs many jobs that used to live in separate products.

Some of the top use cases of ChatGPT per the chart are:

• Practical Guidance (28.3%) which includes brainstorming ideas, getting how to type help on aspects of life, fitness, self care as well as Tutoring or teaching. These are things which people relied on websites such as YouTube, Reddit, etc before for the former, and YouTube, KhanAcademy and Chegg for the latter, but were not personalized or adaptable to any question.

• Writing and editing (28.1%) including writing drafts of posts, emails, etc, editing and summarizing existing text. This is something which may have been done by things like spellcheck, Grammarly, and parts of these (first drafts from scratch) really weren’t possible or accessible before.

• Seeking Information (21.3%) which is very clearly a form of search and information retrieval which people relied on Google, specific vertical blogs and forums for.

Below is a version of the chart showing what various tools and products ChatGPT can “substitute” across the various use cases by fulfilling the job to be done itself.

Bundling wins here for simple reasons. One interface replaces a pile of tabs/apps, which cuts switching costs and makes the next step instant. A single memory carries context across jobs, so the draft you wrote informs the summary you ask for, which informs the email you send. Coverage matters too. ChatGPT is good enough across many jobs in the same session, which is how real work and life actually happen and so if you aren’t a frequent user of a use case, you never need to bother finding the best product for it. And then ofcourse, speed improves because the model does the next action (giving you the answer) rather than handing you links.

Unlocking Net New Behavior

But it is not just that ChatGPT can perform many jobs other tools could do. It enables jobs that are bigger and more fluid because of how LLMs work. The paper frames usage as Asking, Doing, and Expressing. Asking is about getting information or advice. Doing is asking the model to produce an output you can plug into a process. Expressing is sharing views or feelings.

Across the sample, about 49% of messages are Asking, 40% are Doing, and 11% are Expressing. For work messages, about 56% are Doing, and most of those are writing tasks. By late June 2025 the split was 51.6% Asking, 34.6% Doing, and 13.8% Expressing. The chart below shows the split across these.

Why does this matter? Doing is the net-new capability at consumer scale. The assistant does not only answer questions (which in and itself is better than the past form of getting answers manually). It produces artifacts and chains steps with context that carries across turns. You ask for a plan, then ask for the spreadsheet version, then ask for a manager-ready summary, all in one thread. You describe constraints and get a finished draft that fits your situation rather than a generic page of links.

This is what lifts the ceiling on old jobs to be done like tutoring, cooking, or data pokes, and it also creates new habits that did not exist for most people before.

What comes next

Bundles create demand by collapsing steps and reducing friction. Specialists then peel off power users who do one job all the time and need depth, speed, and tight integrations. That is when unbundling shows up. You can already see the signs of it.

For those that code a lot, specialized products such as Cursor and Windsurf that leverage the capabilities underlying ChatGPT (i.e., the models) are better. For those that create a lot of images, using Midjourney or Krea may be better. For those that are writing for work, Notion that has the context of your related knowledgebase may be better.

If you do one of these tasks heavily in your job, a focused tool will often beat the general assistant that is ChatGPT. That idea of the unbundling that is to follow and where we are likely to see it next is what I will cover next week.

Subscribe now

Hi friends,

A lot is happening in AI browser land, so let’s talk about it. Just this past week, Atlassian agreed to buy The Browser Company, the makers of Arc and Dia for $610M in cash and a federal judge declined to force Google to sell Chrome, which keeps the incumbent’s distribution power intact. Perplexity continues to roll out Comet more broadly. OpenAI is widely reported to be prepping a browser of its own.

Let’s jump in and discuss:

• Why the browser matters

• Who the major AI browser contenders are

Why the browser matters for AI

I expect the browser to be an interesting battleground going forward for the following reasons.

1. Distribution surface: Default slots in browsers are worth billions. Google paid Apple about $20B in 2022 to be Safari’s default search on iOS, and spent $26.3B in 2021 across devices and platforms to secure default status. A recent ruling lets Google keep paying partners for defaults, which preserves that distribution engine. Mozilla’s finances also show how powerful this channel is, with a large majority of revenue tied to its Google search deal. Search was the original beneficiary of these defaults, and chat assistants are now stepping into the same front door.

2. Place where agents happen: The browser already sits where work happens, so it is a natural potential control plane for agents. You can grant page context, watch steps, and take over when needed. That is why new products position the browser as an agent desk. Perplexity’s Comet launched as an agentic browser that researches and acts from the page. The Browser Company’s Dia makes it easy to use AI skills and assistants using the context of a webpage. Together, these moves show agents gravitating to the browser surface first, particularly as human in the loop happens. The fact that the browser also holds identity, cookies and app state across various apps and tools people use also further helps further this idea.

The AI Browser Wars Contenders

While still relatively early, and I expect all the key browsers such as Firefox, Edge, Safari, Brave to add AI features, I figured it’s worth touching on the dominant browser today, as well as the key AI-native browsers.

Google Chrome

Chrome is the dominant browser today with almost 70% market share, and as the recent Google antritrust ruling asserted does not need to be divested by Google.

We’ve already seen a few AI features added in Chrome over the last year or so, and as the company continues its focus on AI, I expect to see more. So far, they’ve shipped things like features to organize tabs and help people write and summarize content on a page. Over time, there may be page-grounded actions across multiple tabs.

Google’s goal with Chrome is likely to funnel and leverage Chrome’s distribution to continue to maintain Google and its AI products as the defaults for most users as well as to continue to collect more data around users that can be helpful to improve and personalize Google’s products.

Given that many AI features can be quite expensive, particularly as you move into agentic features, I do wonder if Chrome's massive scale ends up being a limiter to the amount they can push with AI functionality in the browser in the absence of requiring subscriptions to Gemini or similar.

Dia

Dia was built by the team over at The Browser Company of New York, who also built Arc, and was acquired by The Issue Tracking Company of Australia (Atlassian) for about $610 million in cash last week. The Dia browser itself only launched in June of 2025, barely 3 months ago, and was the team’s take on an AI-native browser, which combined their learnings from Arc with what’s possible with AI.

On AI features, Dia already lets you chat with pages and YouTube videos, run reusable Skills, use all your browsing history as context and personalization for chats and more. The vision going forward is to make the browser more useful for work across the myriad of tools people use, including of course Atlassian’s own tools.

“Today’s browsers weren’t built for work; they were built for browsing. This deal is a bold step forward in reimagining the browser for knowledge work in the AI era. Together, we’ll create an AI-powered browser optimized for the many SaaS applications living in tabs – one that knowledge workers will love to use every day” - Mike Cannon-Brookes, Atlassian’s CEO and co-founder.

Going forward, I expect to see more agentic features that allow users to prompt taking actions directly from webpages or their browser, particularly in a work context.

If you don’t yet receive Tanay's newsletter in your email inbox, please join the 10,000+ subscribers who do:

Subscribe now

Perplexity’s Comet

Perplexity launched Comet on July 9, 2025, positioning it as an agentic browser that researches, summarizes, and then acts on the page. It is a natural extension of Perplexity’s retrieval strength, which makes answers feel fast and grounded. Distribution has ramped quickly through a broad public launch and partner promotions, and Comet has the most advanced set of agentic features and functionality of browsers today.

On AI features, Comet blends page awareness with suggested next actions and can already reach into basics like email and calendar when you grant permission. The promise is that you can ask a question, follow links, and complete simple steps without context resets. What comes next is breadth. Shopping, scheduling, and forms are obvious surfaces, along with longer task chains that need retries and recovery without losing the thread, which will take models to improve as well.

From Perplexity’s point of view, it’s likely they view the browser similar to Google Chrome - as a way of increasing Perplexity subscription adoption as well as Perplexity usage by being the default for Comet users. Interestingly, I know of some people subscribing to Perplexity’s max plan specifically for the browser itself, given how useful the still early agentic functionality can be. It will be interesting to see if it remains a distribution channel to core Perplexity or if it becomes an independent product of its own that people pay for.

OpenAI

OpenAI’s browser is not out yet, but multiple reports say it is in the works, with the internal name Aura surfacing in code per a piece in June of this year. The expected shape is a tight coupling between ChatGPT for intent capture and Operator/ChatGPT agent for tooling/an agent that can act with clear prompts and consent.

Prediction markets currently estimate a 70% chance we see the browser this year. Whenever it does come, it will be interesting to see whether given ChatGPT’s runaway lead as the AI assistant of choice for most users with over 700M MAU, they can translate that into browser downloads and usage as a way to continue to deepen their relationship with their users and increase engagement with their core product.

Closing Thoughts

The browser matters because distribution still sets the table and because the tab is a natural place for agents to work alongside us. Chrome remains the default gravity well and is layering in AI features inside the browser. Dia feels like the most AI-native take on a work browser and now has Atlassian’s distribution and product surface to grow into real workflows. Comet is the most agentic today for research and quick tasks, and it is pushing the action model forward in public. OpenAI is the wildcard. If it ships a browser that cleanly hands off from chat to page actions, it could convert it’s large ChatGPT user base into a new OS for personal or professional work.

One final thought on share and switching. Chrome’s scale means even small shifts create room for meaningful new businesses. AI feels like a large enough wedge to change habits and perhaps encourage many people to consider switching their browser, which they likely haven’t changed in a decade.

Subscribe now

Hi friends,

I hope everyone had a great labor day weekend. There has been a lot of discussion about gross margins in AI land recently, so I figured I would chime in with some thoughts. First, I’ll give a quick lay of the land of where we’re at gross margins wise across the stack. Then, a short list of considerations on the gross margin debate for builders across the stack, particularly at the application layer.

Lay of the Land on Gross Margins

• Chips: Nvidia still captures premium economics. Excluding one-off charges, Gross margins for chips hover in the 70% range consistently across quarters.

• Cloud: Hyperscalers do not break out product gross margin cleanly for AI, so operating margin and corporate gross margin disclosures are the best signals. AWS reported 33% percent operating margin for the quarter and 36.7% on a trailing twelve month basis. Microsoft reported Microsoft Cloud gross margin percentage at 69 percent, with a note that AI infrastructure scaling is pressuring the percentage. Google Cloud delivered roughly a ~21% operating margin. Net, the platforms are monetizing AI demand even while they fund the buildout. However, AI is pressuring gross margins, as called out by Microsoft, and some of my estimates have the gross margins in the 50-55% range, though they could be lower for the other clouds.

• Models: Outside estimates peg OpenAI’s gross margin roughly around ~50% while Anthropic’s numbers per the Information also hover around 60%. It should be noted that their are a number of business lines such as the API and the consumer product which make up those numbers, and each looks pretty different. It should also be noted that training costs do not fall under these gross margins.

• Applications. This is where the spread is widest. Bessemer’s 2025 dataset shows fast-ramping AI “Supernovas” averaging about 25% gross margin early on, while steadier “Shooting Stars” trend closer to 60%. They also note that many of the AI Supernovas have negative gross margins, something we don’t tend to see often in software.

  

  Additionally, the way companies are calculating their own gross margins may be different from each other, as the chart below from The Information notes. Some companies include the cost of free users within it, while others do not.

A Few Considerations on Gross Margins

The Application layer is the area which has seen the widest dispersion in gross margins, and I’ll be touching on a few considerations for builders to keep in mind. Coding apps are probably where the Gross margin debates come up most heavily, so I’ll use those as examples throughout.

1) Inference costs are falling, but the benefits accrue unevenly
There’s been a lot of talk about inference costs falling 80-90%+ per year. And while that’s true for a given model, the price of top-end models has stayed steady or even gone up as Ethan noted in the excellent analysis and piece below.

The key question is simple. Does your use case require the top model on every request, or do you only need to meet a quality bar. If you can meet a bar, routing lets you send most traffic to cheaper models and burst to the frontier when needed. If your users demand the best every time, you need pricing that mirrors usage, or margins will compress as quality expectations rise.

2) Control and workflow depth decide unit economics.
Tied to the above, when users insist on the highest performing model, a companies COGS rides someone else’s price card, the model provider. Coding assistants feel this more than almost any other category because quality is visible on every keystroke. In fixed workflows like document processing or voice IVR, vendors own the acceptance criteria, which lets them route to cheaper models by default and escalate only on hard cases. Over time, that gives them the opportunity to push app-layer margins toward SaaS-like territory.

Additionally, workflow depth matters as well. Many AI companies are derided as “wrappers” and given as the reason for low gross margins. Classic SaaS was, in practice, a wrapper over databases and public cloud too, yet it sustained 70-80% gross margins. The path is to get deeper and take control so buyers see more than “access to a model.” Control which is discussed above is one area, but good old workflow depth continues to be another. Collaboration, versioning, audit, analytics, governance, and integrations into the rest of the stack as well as going deeper into the workflows continue to be tried and tested ways to grow ACVs without large increases in AI COGS.

If you don’t yet receive Tanay's newsletter in your email inbox, please join the 10,000+ subscribers who do:

Subscribe now

3) Do not live or die on tokens alone

As hinted at above, there are more ways to make money than marking up tokens. Many of the companies, even in areas like coding and vibecoding where a lot of the value comes from the model itself, are seeking other opportunities to monetise which will not be tied to AI inference costs and so grow aggregate Gross Margins. A few of these examples include

• Hosting and deploy. Bolt and Replit monetize runtime, bandwidth, storage, domains, and private deployments once projects live inside their environments. That raises ARPU and decouples margin from token pricing.

• Marketplaces and services. Replit’s Bounties take a ten percent fee from posters, which is a clean non-inference take rate.

• Advertising / Affiliate. OpenAI has been piloting checkout inside ChatGPT with Shopify, which would create commission revenue on the free tier. Perplexity has been testing sponsored follow-up questions. It’s likely only a matter of time when some forms of ads finds its way to most consumer chatbots.

4) Pricing models still need work and iteration

Most companies started with simple per-seat plans and ran into power-user cost spikes. The shift toward other pricing models that make costs visible and predictable. Examples include seat plus pooled credits, usage with an included allowance and clear pass-through, token bundles with rollover, and BYOK for heavy cohorts. Replit has iterated across these ideas, pairing seats with credits, layering pay-as-you-go, and adjusting tiers as usage patterns change. Anthropic noted that on their original pricing structure of Claude Code, they were losing tens of thousands of dollars per month on a user on a $200 plan.

Pricing models will continue to evolve and be a lever where companies can play around with to make gross margins improve, particularly if hurt by a small subset of extreme power users.

The chart below shows Replit’s gross margins change drastically over time while revenue continued to grow as they adapted pricing models.

5) Net Margins Matter

There’s been a lot of emphasis in recent weeks on the gross margin profiles of these companies. But what’s not discussed as much is that ultimately it’s a trade-off these companies are making where they trade-off gross margins for faster growth and lower costs in other areas.

You can trade some gross margin for distribution if product-led growth is truly working which results in faster revenue growth and leads to lower sales and marketing (and perhaps even G&A) as a percentage of revenue. The Supernova pattern that Bessemer highlighted is exactly that. Very high ARR per employee and thin early gross margins that thicken over time with the above additions, but much lower S&M, G&A as a percentage of revenue than their SaaS counterparts given the fast scale and the bottoms up motion truly working.

The goal is not a perfect gross margin in isolation. The target is a healthy net margin profile as cohorts mature.

Subscribe now

Hi friends,

The three hyperscalers just reported earnings, and the cloud wars are heating up. This quarter was especially important because it shows how AI is reshaping not just growth but share dynamics. Let’s break down the numbers, see who captured the most new dollars, and then look at what management is saying about demand.

Note that these numbers are estimates, given the companies don’t report apples-to-apples comparisons on run-rate, etc and break out their cloud divisions in the same manner.

1. Where they stand today

• AWS is still the largest by a wide margin, generating about $30.9 billion in cloud revenue last quarter. That translates into an annualized run rate of roughly $124B. Growth came in at 17.5% y/y.

• Microsoft Azure continues to close the gap. While Microsoft doesn’t break out Azure run-rate directly, they did reveal that they did 75B in revenue over the llast year which we we can use to estimate their last quarter revenue of around $21B, putting its exit run rate near $84B. That’s up 39% y/y, making it the fastest grower of the three.

• Google Cloud reported $13.6 billion in revenue, which annualizes to more than $54.4 billion in run rate. Growth was 32% y/y. It’s important to note, though, that Google’s number includes Workspace (G Suite), which inflates the headline compared to AWS and Azure’s pure cloud businesses. Even adjusting for that, Google is still growing briskly, but from a smaller base.

The numbers are summarized below.

2. Net new run-rate and share capture

It’s one thing to look at total revenue, but the perhaps more interesting lens is to focus on the net new run rate added over the past twelve months. That’s where you see who is truly winning incremental share.

When you add it all up, the three of them added about ~55B in incremental annualized runrate over the last twelve months.

• Azure captured the largest slice of new dollars, adding about $23.6B in run rate. That’s roughly 43% of all the incremental spend in the market.

• AWS followed with around $18.4B, which equates to 34 percent of the net new pool. Despite its massive base, it’s capturing a smaller share of growth than it used to.

• Google Cloud pulled in about $13B, good for 24 percent of the new dollars. It’s still the smallest of the three, but the pace shows it’s carving out a meaningful seat at the table.

On a relative basis, the share of incremental dollars between the three only looks like below, with Azure taking the crown. Note that since Google Cloud includes Gsuite, it’s actual share is lower and the other’s is higher.

If you don’t yet receive Tanay's newsletter in your email inbox, please join the 10,000+ subscribers who do:

Subscribe now

3. What management is saying about AI demand

AWS notes that demand still outstrips capacity

Andy Jassy: “On the supply constraints as it relates to AWS and what we see there, -- as I mentioned, we have more demand than we have capacity at this point. And I think that -- and you see some of the constraints and they kind of exist in multiple places, the single biggest constraint power. But you also see constraints off and on with chips”

… and will continue to invest heavily in capex

“Brian Olsavsky( CFO): “Now turning to our cash CapEx, which was $31.4 billion in Q2. We expect that to be representative of the quarterly rate for the back half of this year, putting us on track for ~$118 billion full-year CapEx.”

Microsoft notes their AI leadership

Satya Nadella: “We continue to lead the AI infrastructure wave and took share every quarter this year.

… and that they also are capacity constrained

“Even as we continue bringing more data center capacity online, we currently expect to remain capacity-constrained through the first half of our fiscal year.”

Google noted their growth in large deal sizes

Sundar Pichai: “We see strong customer demand, driven by our product differentiation and our comprehensive AI product portfolio. Stats show this: One, the number of deals over $250 million, doubling year-over-year; two, in the first half of 2025 we signed the same number of deals over $1 billion that we did in all of 2024; three, the number of new GCP customers increased by nearly 28% quarter-over-quarter.”

…. and their differentiated approach of vertical integration

“We are seeing significant demand for our comprehensive AI product portfolio. Of course, this is all possible because of the long-term investments we have made in our differentiated full stack approach … which spans from AI infrastructure to world-class research, models and tooling and our products and platforms that brings AI to people all over the world”

4. Key takeaways

In closing, what do we take away from all this? A few things stand out to me:

A/ AI is the engine with capacity the bottleneck

All three CEOs hammered on AI demand as the main driver of cloud growth, but each also admitted they’re constrained by infrastructure buildouts. Backlogs continue to grow, with Google’s backlog hitting over $100B.

B/ Azure is the clear share gainer

Capturing the largest slice of new dollars shows Microsoft’s strength in AI. I wonder how much of the share gain is down to the OpenAI relationship and indeed directly serving most of the OpenAI demand.

C/ Capital Expenditure, primarily on AI continues to be significant

Lastly, it continues to be evident that there’s no slowing down on the investment front from any of the hyperscalers. Google raised its 2025 guidance to about $85B, up from $75B, to keep pace with AI-driven demand. Amazon is on track to spend roughly $118B, fueled by investment in chips, datacenters, and power. And Microsoft expects to spend over $80B, with some quarters alone topping $30B.

Subscribe now

Hi friends,

A few months ago, I wrote about the rise of the agentic workforce, focusing on the agentic coworkers popping up across the key functional roles. Today, I’m breaking down the rise of verticalized agentic coworkers. These are AI agentic coworkers built for specific roles within industry verticals.

What do verticalized agentic coworkers look like?

• Modeled on real jobs: Each agentic coworkers is designed to take a subset of core tasks of a specific, high-headcount role. Think insurance claims adjuster, nurse, or compliance officer. The focus is on work that’s high-volume, repetitive and mostly measurable and/or wasn’t getting done because of lack of time (or in job roles with labour shortages).

• Integrated into existing workflows: These coworkers plug directly into the software and tools already used by the team. They read and write data, process transactions, and interact with other digital systems.

• Task ownership: Unlike generic assistants, these coworkers actually complete work and focus on a subset of tasks that humans in the role perform, ideally taking it off the plate of the human (minus human verification).

• Usage-based or Outcome-based pricing: Pricing is usually tied to usage or results (claims processed, reports generated), not just a monthly seat fee.

• Human oversight/manager: Most agents still operate with oversight. Humans review outputs, handle exceptions, and step in when needed. The goal is to expand autonomy as trust and performance improve and expand the set of tasks they can perform over time.

Where are we seeing verticalized agentic coworkers?

While still in the early innings, we’re starting to see the first wave of agentic coworkers across most key verticals such as finance, healthcare, legal, insurance, retail, logistics, and many more. A market map is presented below of the key AI-native companies that offer at least some agentic functionality in various vertical roles.

Some of they may fit across multiple roles, but I’ve tried my best to bucket them based on the core role they serve.

Common Wedges for Agentic Coworkers

Most verticalized agentic coworkers wedge into the workflow by starting with a repeatable, high-volume task where cost of errors is relatively small. The most common entry points we’re seeing across verticals are:

• Document processing and parsing: Extracting, classifying, and acting on information from contracts, invoices, intake document. This is often the first workflow AI agents can automate end-to-end.

• Voice agents for intake, triage, sales, or customer experience: Voice-based agents are making it possible to automate high-volume, front-line interactions. Many AI coworkers handle inbound calls or chats for scheduling, triage, sales qualification, or customer support.

• Content creation: Drafting proposals, memos, emails, or reports. Agents trained on industry-specific templates can generate high-quality output that matches the expectations of the role.

• Search and information retrieval focused workflows: Surfacing relevant policies, research, or case history in response to specific queries, coupled with forms of content generation. Strong retrieval capabilities are often the backbone of agents in legal, finance, and healthcare, that start of looking like assistants and over time can draft forms, memos, etc

I’ve written this post which goes deeper into these wedges.

Closing Thoughts

Most of the AI coworkers do a fraction of the tasks of a human in their role today. But as models and agentic systems get better over time, they will be able to do more and more tasks and go from junior interns to junior employees to beyond that. It’ll be interesting to see how these companies evolve over time. A few different combination of approaches companies may take:

• Do more work within the role: Eat up more and more of the work (different tasks) and increase their value and pricing power over time

• Build adjacent coworkers: Build adjacent coworkers for additional roles within the industry to capture more spend within customer base.

• Build AI-native systems of action: Leverage the data they’re collecting through doing work to build an AI-native system of record and action within the industry they operate in.