Even the most blissfully childlike among us know that the AI labs are losing money. In fact, if you just examined Anthropic’s or OpenAI’s revenue against their expenses, you wouldn’t just think that they’re unprofitable; you’d come to the conclusion that they’re magnificent money-incinerating machines. It’s hard for a nation-state like Afghanistan to spend a couple billion dollars, yet each of the AI labs spends that much year after year without blinking.
On the other hand, Anthropic recently raised money at a $380 billion valuation. Some of the most sophisticated investors in the world believe that Anthropic has the potential to earn more money than old-school companies like IBM, Coca-Cola, or even Goldman Sachs. The same story is playing out for OpenAI, except it recently raised at an $852 billion valuation, more than twice Anthropic’s valuation. The theory is that, eventually, these AI labs will turn the corner from horrifically unprofitable to legendarily profitable.
Each of the AI labs has their own theory on how they get there.
Anthropic’s CEO, Dario, has thrown out the math that Anthropic is profitable on a model-by-model basis. OpenAI is trying to monetize its 900 million weekly active users with ads. Google’s own AI efforts seem to be mostly around defending their core search engine business. And all three companies have software engineering agent projects at the core of their business.
With any new technology, commercialization is always a guessing game. Famously, Yahoo tried to monetize their audience by becoming a media portal that served on-page ads. In hindsight, the correct play was to build Google’s CPC ad engine and ignore CPM-based advertising. Even earlier, IBM assumed that the BIOS was the only layer they needed to lock down and signed a non-exclusive license with Microsoft for their DOS system. A few years later, dozens of computer manufacturers reverse engineered the IBM BIOS, and Microsoft became the dominant player in the PC market.
The same is true for AI. The AI labs are now generating significant amounts of revenue (though still only a fraction of their expenses). As their revenue grows, their commercialization strategies have begun gradually converging to the same thesis. Namely, enterprise deals, agents, API servicing, and coding harnesses. And the exact way these labs make money says a lot about their prediction of the future and the role LLMs play in it.
Intel and the CapEx Problem
Before we talk more about AI labs, let’s go back to an earlier period of tech history. If you’ve read my other articles, you’ll know that I like to reference what happened in the past as a starting point for thinking about our current situation. In particular, I think we can learn quite a few lessons from the previous titan of technology, Intel.
A few decades ago, Intel was the pinnacle of the tech industry. Every machine worth its salt, from Apple Macs to Dell PowerEdges, used Intel chips. Andy Grove’s Only the Paranoid Survive was a bestseller. It was Intel’s world, and we were all finding our place in it.
Over the years, competitors and management mistakes carved away at Intel’s invincibility. They turned down the opportunity to supply iPhone chips, failed to enter the GPU market, and faced delays in adopting advanced node production. Intel’s fatal flaw, however, was not recognizing what rising CapEx meant for their business.
Semiconductor fabrication facilities (fabs) have always been costly. In 1980, a single fab took around $100 million to construct. [0] Once done, however, these fabs became money-printing machines. Raw material costs for each wafer were minuscule (less than 2%) compared to the prices they demand. Over time, the equipment side of fab construction became more and more costly. By the late 90s, the price for leading-edge fab was more than $1 billion. Today, a modern fab can cost $10-$20 billion or more.
Intel, after pivoting away from memory, is a vertically integrated company. They designed various chips, built manufacturing fabs, and sold directly to OEMs like Apple and Dell. In a world where Intel is the only player in town, this meant that they took a high percentage of the money spent on semiconductors.
The exponentially rising CapEx costs left Intel with a decision. They could either sell more chips than ever to justify their exponential investments. Or they could increase their margin by raising prices, squeezing more juice out of every sale.
They chose the latter. And, given that they weren’t confident in matching the exponential explosion in cost, they simply… didn’t. Intel began investing less aggressively than their foundry competitors like TSMC and Samsung. In 2016, TSMC’s CapEx surpassed Intel’s and permanently took the lead in 2020.
In fact, TSMC guided towards $52 billion to $56 billion in CapEx for 2026, which is potentially higher than Intel’s revenue for the whole year. On the other hand, Intel’s projection for 2026 CapEx is $9.1 billion, less than 20% of TSMC’s.
In decreasing CapEx, Intel could no longer maintain the technical superiority, which meant that their end chip products were less desirable. That meant there was less money to invest in leading-edge nodes, which repeated the cycle. They’ve had to cancel a $30 billion fab in Magdeburg, which led to the German government taking back 9.9 billion euros of planned subsidies, and a $4.6 billion project in Poland, foregoing another 1.7 billion euros in state aid.
In the past few years, Intel has been bleeding money at almost the same rate as top AI research labs. They lost almost $10 billion in 2022, $14 billion in 2023, and $15 billion in 2024.
TSMC, on the other hand, has been riding high. Being a pure-play foundry that doesn’t involve itself in the design or marketing of chips, they only need to execute on building the fabs and then pressuring its many customers to source leading-edge wafers as quickly as possible to amortize the investment. In the end, TSMC takes very little risk despite its massive CapEx figures because the capacity is already keyed to orders from existing customers. Even better, the customers themselves are diversified across the whole spectrum of consumer and commercial computing goods. When demand for one potential use of wafers slips, other types are there to pick up the slack.
At this point in time, we can pretty confidently say that Intel's strategy of vertical integration has failed in the face of exponentially rising CapEx.
AI and the Frontier Game
AI research CapEx figures are also exploding. Where GPT-3 cost less than $4 million to train, the GPT-5 series models are estimated to cost somewhere in the billions to train. Costs are rising, and risk with them; we’re seeing the same fab dynamic play out with frontier AI models, but with one key difference:
The economics for AI models is much, much worse.
Let’s say that you’re a new player in the semiconductor market today and you decide to build a chip fabrication facility. Your options of how you’d go about it are limited. It would be hard to jump straight to the leading edge when even state actors like China’s SMIC are stuck on 14 nm nodes that are three to four generations behind.
That’s okay, though, since there are still plenty of companies that will buy chips produced on older technology. Nvidia likes to stay 1-2 generations behind leading-edge nodes for their AI processors, and advanced driver-assistance systems (ADAS) function fine with 14 nm chips. You’ll still have buyers, even if you have to sell the wafers at bare-metal prices to compete with Intel and TSMC, who have already paid off their fixed-cost investments.
However, the ability to sell the obsolete does not also apply to an AI lab. As an AI lab founder, you’ve spent billions to train the latest frontier model, and you can only count on making money on it so long as it remains among the cutting-edge entries in the field. Worse, the clock has already started ticking on the model’s relevance. OpenAI has introduced a better frontier model roughly each year, which means your frontier model has a maximum shelf life of two years before it’s functionally obsolete.
Sometimes, the timeline is even worse than that. OpenAI is already reduced to earning negligible amounts on their GPT-4 series models, the latest of which came out about a year ago. As crazy as it is to say this, OpenAI is making a rational economic decision. Their GPT-5 series models cost less in terms of inference, which means that given the chance between serving a more expensive and less performant GPT-4 model and a cheaper and better GPT-5 model, OpenAI will always pick the GPT-5 models. It also means that OpenAI has essentially run the course with their GPT-4 models that cost hundreds of millions to train!
So as a frontier lab, you somehow need to generate enough demand to cover all the fixed and variable costs of any new model in an incredibly short timeframe. You need to compete for adoption against Google’s Gemini, OpenAI’s GPT, Anthropic’s Sonnet/Opus, and even XAI’s Grok to earn enough money for the next round of research. Unlike the semiconductor manufacturers, AI labs are sharks that have to keep swimming. If one slows, even for a moment, its competitors will overtake it, steal away its token sales, and sink the company completely.
OpenAI and the Trap of Verticalization
For a while, OpenAI swam the AI waters perfectly. They became the modern-day Intel. Anyone who did business with OpenAI saw their stock jump by double-digit percentages due to the mere potential that association held. For most people, their first, second, and only interactions with AI were through ChatGPT.
OpenAI’s lead was so large that they could broaden their ambitions beyond the scope of a mere AI research lab. They released Sora, a video generation model that once again took the world by storm. They acquired io, the creative studio founded by Jony Ive, and began work on a hardware device. They partnered with others to incorporate Stargate, a venture that will invest up to $500 billion in AI infrastructure.
In sum, they’ve tried to become the vertical AI company, from compute to AI end-usage.
For a while, this strategy worked. Google fumbled their Gemini and Bard releases. Anthropic couldn’t finance the same level of commercial intensity, to the point where people called it the Lyft to OpenAI’s Uber. And Facebook’s own research team failed to stay at the frontier.
And then the cost of exponentially rising CapEx caught up with OpenAI.
There’s one more difference between frontier AI labs and semiconductor manufacturers. For any given chip, the raw material costs are less than 5% of the total value. Most of the money from chips is funneled towards paying back the CapEx.
For inference in frontier models, the energy and GPU costs can sometimes be 20x higher than the revenue that they’re bringing in! The cost of each token doesn’t stay constant, which means that one user can cost magnitudes more money to service than the next without an obvious way to distinguish between the two.
That became one of the first flaws of OpenAI. Where OpenAI is incredibly gifted at building consumer products and engineering viral distribution, its users are quite price sensitive. The percentage of users converting from free to paid has stayed constant despite massive improvements to the intelligence of OpenAI’s models.
That meant OpenAI was often taking on huge losses to keep free users or not fully breaking even on paying customers. Earlier this year, they shut down Sora with reports that they were losing $1 million every day on compute spending for video generation. And, in an echo to Intel’s woes, shutting down Sora lost them a $1 billion investment from Disney.
The second flaw of OpenAI repeats Intel’s footsteps. One of the key moments in Intel’s flaw was turning down the Apple iPhone chip business. Later, they tried to rectify that with a $1.4 billion acquisition of Infineon's wireless business in 2011 only to exit the industry and sell the subsequent division to Apple for $1 billion eight years later.
Likewise, OpenAI fumbled a key partnership with Apple. In 2024, they negotiated access to serve the GPT models alongside Siri. Two years later, Apple shifted its primary AI partnership for Siri to Google’s Gemini. If we accept the initial assumption that OpenAI cannot become a vertical player (such as producing their own AI device) due to the exploding CapEx involved in running an AI research lab, then this development limits their future in the entire smartphone market. After all, Apple’s iOS and Google’s Android run a duopoly for smartphone OS’s.
Switching gears, let’s talk about OpenAI’s primary startup competitor, Anthropic. Early on, Anthropic took a different path in becoming a horizontal company within a specific vertical. They focused on developers and optimized Sonnet and Opus models to be the best at coding-related tasks. As a result, they became the default model for coding harnesses like Cursor and Cognition, and could ride the growth of their application-layer customers in the same way that TSMC grew with its customers.
Anthropic then doubled down on that, publishing research and insights meant for technical audiences. Where OpenAI had to market to the lowest common denominator of its users, Anthropic intentionally focused on higher-value audiences. This strategy allowed them to bypass the stranglehold OpenAI had on consumer AI and jump straight to enterprise adoption.
Anthropic is now putting up such incredible revenue numbers that the narrative around the two AI labs has completely shifted. People believe that Anthropic will win the AI wars. They’ve found a better revenue strategy to support the exploding CapEx. As more and more industries begin to adopt AI, whether through application layer companies or direct contracts with AI labs, Anthropic is positioned to take the lion’s share of that growth. After all, they’re the horizontal player to support that.
This, in turn, makes it possible to imagine a future where Anthropic takes the lead not only through revenue but also through the ability to take more risks. Where TSMC has been able to offset the risk of building new fabrication facilities with its wide base of customers, Anthropic can now do the same with the suite of application layer and enterprise company customers.
Anthropic has become the TSMC to OpenAI’s Intel.
Now, Anthropic isn’t the only player looking to spoil OpenAI’s party. Google and Facebook began to leverage their resources to draw away ChatGPT users and hire away OpenAI researchers. ChatGPT growth has thus stalled in recent months, while Gemini has been making large gains in monthly active users. And Facebook recently released Muse Spark, which is roughly the equal of OpenAI’s latest GPT model.
Closing Thoughts
In more ways than one, this essay is merely a proxy for what’s actually happening. There are dozens of messy pieces that could throw the theses stated here into disarray.
Maybe one of the OpenAI researchers has a breakthrough and they develop AGI first. Then, it’s game over for everyone else. OpenAI is also not sitting idle while its competitors make large gains. They pivoted their efforts toward coding with Codex and doubled down on their enterprise efforts. Anthropic has also stepped towards verticalization with their release of Claude Code, essentially competing with their initial customers like Cursor.
But that’s the best part about being at the frontier. Everything’s messy, and just like the best sports games, even in the last seconds, anyone can win.