The current belief is that the AI industry has finally cracked the revenue code, and the unprecedented growth companies like Anthropic or OpenAI are experiencing will continue for years. The growth rate is so impressive that if it sustains for just one more year, Anthropic will have a $1 trillion annual revenue run rate by the end of 2027.

And while that won’t happen, it’s not crazy to believe that Anthropic will reach a $100 billion/year run rate by the end of 2026, around double today’s value, and on track to represent a 10-fold increase for the year, an unheard-of growth rate at such values.

However, I fear all this might be a temporary illusion, which will lead to a much greater need for liquidity or… else.

And this liquidity will come from where? From a place most people don’t expect, and the answer is going to make you feel very nervous about the future of AI and the global economy.

Let’s dive in to understand why.

The growth of AI labs is historically unusual because the leading companies are scaling from research organizations into cloud-scale revenue machines in only a few years.

As you can see below, OpenAI and Anthropic reached $1 billion in revenue much faster than all other tech companies in history.

Moreover, the growth rate is even more impressive the higher they go, as not only were they the first to reach $1 billion, but the gap widens with each subsequent milestone. OpenAI/Anthropic reached a $20 billion run rate almost two decades faster than Netflix.

The catalyst, the key product, has been coding agents: AIs that are used to write code. One could argue it is the product right now, and most other Generative AI use cases, with the exception of search, are mere satellites compared to this giant planet.

But the product itself isn’t the only star of the show explaining the huge revenue growth. And it’s precisely these unacknowledged ‘stars’, the ones that investors conveniently ignore, that are the problem.

On the one hand, as any start-up trying to grow at all costs would, these companies have happily traded unprofitability for rapid growth, sacrificing the bottom line (profits) to boost the top line (revenue).

They are doing this by offering these products at massive discounts relative to the actual prices required to make money.

This is an acceptable business practice, but with obvious implications. If you set up a candy store, you will get some guests. But if you go out into the street and give candy for free, the amount of “customers” skyrockets, but you aren’t making more money; you are losing money.

So while you can boost your revenues a lot by massively decreasing your price, that doesn’t say very good things about your business; if you’re having to sell it for less than what it costs you, you don’t have a business at all, you’re just pretending to have one.

A perfect example of this is subscriptions, the most popular form of subsidy (or, more accurately, revenue-opportunity loss relative to the API business).

For example, if you ran usage to drain the ChatGPT $200/month subscription using APIs only, you would spend $14,000. Careful, this isn’t saying OpenAI burns $14k on a Pro subscription to earn $200; it’s describing the “revenue opportunity loss” for them; it’s the money they could be making from you for that usage.

This raises an important question: would these same people be willing to pay $14,000? I think not.

So, what will happen when AI is priced accordingly? And even then, will the API revenues shown above be enough?

It’s important we insist on this because most people misunderstand the difference between margins and cash flows. Put simply, you can be profitable and still lose money.

But how?

Frontier Labs are profitable at a gross level, meaning they charge more than they spend to operate and serve you tokens. If the gross margin is 50% (which seems like a good figure for where these companies stand), that means they make a dollar for every 50 cents they spend to serve you with AI models.

Down the line, they do have a clear path to operational profitability, too, meaning that at current growth rates, they’ll soon make money even after accounting for expenses like salaries and marketing. I could even see that happening in the next two years, meaning that for every $10, they might be making something like $1 or $2.

Not great, but profitable.

Most people see these numbers and immediately conclude that these labs are two years away from making money from AI. However, that is not necessarily true, because the key to understanding return is considerably smaller than what they actually spent on you.

But what do I mean by ‘actually’?

The problem most people miss is that they don’t understand cash flows. If I make 10 dollars on something I spent $5 on, I look super profitable. However, if I spent $100 that year to purchase the assets that allow me to exist as a business in the first place, I’m still losing $95.

Of course, one could interject and say, “Sure, but those $100 come from the company reinvesting to grow, and they could just stop spending eventually and be insanely profitable AND generate cash.”

But that’s not how the AI business works, guys. If you stop spending, you lose. Spending unlocks more compute, which unlocks better AIs, which unlocks more revenue (or dare I say, allows you to continue to compete). You fall behind, and your business goes to zero.

In other words, this industry is very capital-intensive and will remain so for the foreseeable future. And if that is true, cash flows are all that matter. Margins can be good or bad, not as a nominal value, but relative to how much cash they allow the company to generate. For all I care, these companies could have 100% margins and still lose money.

Imagine you have a machine you have to purchase every year for $1,000, because it only lasts 1 year, and you manage to get $1 back for every cent you spend to use it. That’s a 99% margin. But if you only manage to sell $100 for the year, you’re still down $900.

Nonetheless, both OpenAI and Anthropic have entered into several-GW deals with many suppliers that amount to more than a trillion dollars in committed spending.

I continue to insist that it isn’t clear to me at all how these companies will ever make money unless they either raise prices like crazy or spend less capital.

In a nutshell, this is a very long way of me saying this business sucks, and you rely on SoftBank and other investors to cover your losses. Without them, you die, probably in a few months from now if cash dries up today. There’s a reason OpenAI raised $122 billion in one round, and Anthropic has raised $95 billion year-to-date alone.

Nobody talks about this because we all have to pretend otherwise so that the hype train doesn’t run out of fuel.

But I digress. Besides financial engineering, the other key factor is narrative, specifically the remarkably stupid ‘tokenmaxxing’ strategy that VCs in Silicon Valley somehow convinced the world for a few months that it was a good idea.

A bloodbath later, companies like Uber and ServiceNow had spent their annual budgets in a few months, with little proof of return. Simply put, what I’m trying to say is that, for a brief period, the stars aligned for these companies.

Combined, yeah, you’re going to see revenues explode, which is what happened. Sadly, however, people are still treating this growth as sustainable and these behaviors as indicative of future pricing behavior, which, in fact, I believe could be the opposite.

Besides the commoditization pressures that Chinese models put on US prices, something I won’t get into today because I’ve done so several times recently, which is a huge problem in itself because now Anthropic and OpenAI finally have a credible threat that gives 90% of the performance for 10% (or sometimes, below 5%) of costs, with models like GLM-5.2 actually giving frontier US models a run for their money in raw performance, the more interesting question here is whether Anthropic and OpenAI’s current metrics are sustainable by themselves.

And I’m not talking about growth rates, which will obviously decline over time; I’m talking about the revenue itself, which could stagnate or even decline, too.

For starters, as mentioned earlier, I believe a non-negligible amount of current revenue comes from enterprises simply testing AI without necessarily committing. “Testing out AI” has three implications here:

Don’t get me wrong, you will have to spend more tokens to generate better and bigger outcomes from AI, but not as a “strategy". Don’t forget that just a few months ago, Hyperscalers had ‘Token rankings’ to measure employee “productivity” based solely on token spend.

Where are those rankings today? They went as fast as they came.

Long story short, not all revenue is equal, and you should have every incentive to measure the “quality” of these revenues. However, that revenue is still recognized for what it is, revenue. And in the very stereotypical Silicon Valley tradition, conveniently expressed as MRR (Monthly Recurring Revenue) immediately.

But is it actually recurring? Are these companies going back for more the next month?

We don’t yet have net dollar retention metrics from Anthropic or OpenAI, but I do have my personal thoughts on this matter.

For the most part, unless you train on them, models are fungible, meaning you can easily switch if you find a better option. This was proven by OpenRouter data, a popular LLM aggregator platform (a platform that lets you switch between models with ease). Across several models, OpenRouter sees “high churn and rapid cohort decay.”

Therefore, Anthropic and others seek to build product-level stickiness, with examples such as Claude Code, Claude Design, and OpenAI's Codex.

Furthermore, creating the best products on top of AI models is surprisingly hard. In fact, time and time again, we see third parties beating these Labs at their own game, despite owning the models. Good examples include Cursor and Factory, which many believe have the best coding harnesses despite using Anthropic and OpenAI models underneath.

In other words, I don’t think products are sticky. Therefore, while I do believe OpenAI and Anthropic do have much better retention rates than all these application-layer companies (e.g., Lovable or Replit), I’m growingly concerned about the actual sustainability of this business as is.

Then there’s budgeting. Uber is the prime example once again, having curtailed spending to $1.5k per user per month.

This isn’t a churn per se, as Uber remains a customer, but it’s a considerable reduction in projected revenues from one of your most important clients. But we probably both agree that right now, I’m saying a lot with little to back up my claims.

For what it’s worth, I could be wrong, and these businesses may have extremely low churn and sustainable revenues for years. My gut tells me this isn’t true, but you would be wise not to take my hunches as gospel.

Instead, let me explain in a more principled, reasoned way why I don’t believe model serving might not be a great business.

I’m of the opinion that the Fable ban has done way more harm than we yet realize, because for the first time, there’s a discussion to be had about sovereignty.

And although this has geopolitical ramifications (e.g., the Pax Silica, or incentivizing the creation of a European champion, which I don’t believe is possible, but we’ll leave that for another day), I’m talking about enterprise sovereignty.

Hate to be that guy, but I’ve been pounding this idea in this newsletter for years; you should strive to own as much of your AI use as possible, and outsourcing your entire technological stack to a third-party, be that OpenAI, Anthropic, Google, or who knows, is a terrible idea.

The reasons are many, from geopolitical to regulatory, but the most important reason is purely technological, of performance.

Frontier AI Labs has convinced the world you don’t need to train AIs on your data to reach top performance, and that is a lie.

In fact, the hard reality that nobody in Silicon Valley wants to acknowledge is that AI remains very much indeed a deep technology, and by deep, I mean that depth beats breadth.

There are already countless examples showing how you can take a “worse” model and make it fit your use case simply by training the model on your data. Just to name a few:

Why this happens is quite simple to understand with a human analogy. No matter how much your junior employee knows, she can have 4 STEM degrees for that matter, you’re still going to run her through on-the-job training for a few weeks or more.

Without it, all that broad knowledge is useless, and contextual and in-company knowledge are essential to the tasks at hand.

Basically, with the sole exception of coding agents, which clearly have product-market fit for companies, the incentives are misaligned, and what enterprises really need isn't what OpenAI or Anthropic are building to raise their valuations.

In a recent interview, Nikesh Arora, Palo Alto Networks’ CEO, echoed a similar idea, stating that consumer users have a much higher tolerance for false positives (e.g., hallucinations) because questions are broad and only need to be broadly helpful. But companies need things to be executed perfectly, something current AIs are NOT built for.

Put another way, two crucial factors are at play here:

Instead, you’re stuck with a model that, to close the knowledge gap it has about your data, has to be very large and way more knowledgeable about the world than you actually need.

In layman’s terms, to deal with the fact that you can’t train the model, you’re hiring an overqualified individual who might not need any training but costs way more than hiring a less qualified individual and just training them on the job.

What companies are, unknowingly, “asking for” is a relatively small, affordable base model they can train on each individual task and optimize for that task, and a model that can be safely stored within your organization.

Thinking Machines Labs is a perfect example of a Lab that ‘gets it’. From the very beginning, they centered their business on unlocking enterprise adoption by reducing the complexity of training models on their data.

This is terrible news for Frontier AI Labs, which wished you would simply just outsource your entire AI stack to them, having zero control over your models and overpaying for tokens like groupies at a Bad Bunny concert.

They really have little option; their entire business relies on this being true. But I just don’t know how it could be true.

Why pay 50 times more for perhaps even worse performance? Enterprises may not understand AI, but they do understand budgets, financial KPIs, and business cases, and none of this makes sense if you’re relying on Frontier Lab tokens for everything.

As an example, hundreds of models have been released since then, and I still use Gemini 3 Flash and Gemini 3.1 Flash Lite to extract invoices, yet neither would even come close to the top 50 models today.

But they work just fine for that task and at a great price, so why would I change them unless it was for an even cheaper model?

Soon enough, I believe most companies won’t need to chase the latest model for most tasks and will prefer cost-effective older models.

And can’t the Frontier model Labs simply drop costs and compete? They can, if they’re willing to burn even more of the money they are already burning, but not as a sustainable business strategy.

All things considered, how should we expect the AI market to evolve over the next few years?

I’m sorry, but I only see it one way: a very strong pivot toward enterprise-sovereign AI, AI that belongs to the company, not to outsiders. An AI that can be controlled, governed, optimized, and updated at the enterprise’s pleasure, not because OpenAI decides to sunset the model to make room for newer ones.

And what are the implications? Put simply, I don’t view AI revenues for frontier labs as sustainable under the current direction. Tokens aren’t getting cheaper; quite the contrary, and instead of closing the intelligence-per-cost gap with China, it’s only getting worse.

So if revenues are eventually confirmed to be less promising than we had hoped, and we aren’t at hundreds of billions in yearly revenue by the end of the year, then what?

In that case, we need liquidity from elsewhere. And that elsewhere is precisely one of the biggest sources of concern for this entire industry. And let me tell you, you aren’t going to like what I’m about to show you.