Who really holds the power in the AI ​​race: the modeler or the person in charge of the computing infrastructure

The same week when the news that Anthropic surpassed OpenAI in revenue was spread everywhere, I read a detail that made the story of winning and losing much more complicated: Anthropic, the company that had just been hailed as the winner, was paying $1.25 billion a month to xAI, Elon Musk's AI company, just to rent computing power from the Colossus 1 data center near Memphis. Google, one of the companies that owns the largest cloud computing infrastructure on the planet, is also paying $920 million per month to SpaceX to borrow more than 100,000 Nvidia graphics processing chips (GPUs) located in xAI data centers. Reading here, I started to ask myself the question that this article wants to answer: who is really in charge of the AI ​​race, the one who makes the smartest model, or the one who has the infrastructure to run that model?

The revenue flip doesn't tell the whole story

First of all, the revenue numbers are real and worth talking about. Anthropic has reached an annual revenue of more than $30 billion, surpassing OpenAI's $24 to $25 billion, an impressive comeback when at the end of 2025 this number was only about $9 billion. The difference lies in the strategy: OpenAI focuses its efforts on individual users through ChatGPT, while Anthropic chooses to plug deeply into the business workflow from the beginning. As a result, for the first time, Claude was paid more by American businesses than ChatGPT, and Anthropic's average revenue per user was about fifteen times higher than its competitor.

But revenue is only one layer of the story. To generate that revenue, Anthropic had to run its model on someone else's hardware, and the spending on that hardware was so large that it could eat up a significant portion of the financial advantage it had just gained.

Even the winner has to hire an accountant

The detail I find most thought-provoking is the map of infrastructure lease agreements that are overlapping each other in a rather surprising way. Anthropic pays xAI $1.25 billion per month until 2029 for exclusive use of the full capacity of Colossus 1, a data center xAI initially built for itself, in exchange for more than 300 megawatts of additional capacity and over 220,000 Nvidia GPUs. At the same time, Anthropic also has a separate agreement with Amazon with a scale of up to 5 gigawatts, and another agreement with Google and Broadcom also at 5 gigawatts, expected to operate gradually from 2027. Meanwhile, Google, despite being one of the largest owners of AI computing infrastructure in the world, still has to pay SpaceX $920 million per month from October 2026 to mid-2029 to borrow more work. capacity from the same data centers that xAI once built for itself. Google explains this as "bridge capacity" to meet the demand for Gemini Enterprise, which is growing faster than their own expectations.

Ultimately, the power is in the hands of companies with infrastructure like SpaceX - XAi Looking at this picture, I realize a paradox: even the company that is leading in model revenue, and even the company that owns the most massive cloud infrastructure, does not have enough computing power of its own to meet demand, and must queue up to rent it from a third source. That third source, in many recent cases, has been the newly merged infrastructure empire between SpaceX and Elon Musk's xAI, which is selling the very computing power that was supposed to be built to compete directly with those models.

As the race moves from algorithms to concrete and wires

Another detail that further strengthens this perspective is that Amazon has just submitted documents to issue corporate bonds with a scale of at least $25 billion, divided into eight terms, the longest up to 40 years, maturing in 2066. The amount of money mobilized, as the market understands, will almost certainly be poured into building data centers for AWS. Amazon's investment budget this year has reached $200 billion, up sharply from $131 billion last year. Taking on long-term debt rather than just cash on hand, with a maturity of up to four decades, was a symbolic financial decision: it showed that the AI ​​race is now measured not only by who is better at training models, but also by who can borrow the most money, build the most data centers, and lock in the most stable power for years to come.

Amazon has also just raised money to invest in AI infrastructure With that same logic, I recall at the end of June, when it was reported that Google had to limit the Gemini capacity granted to Meta due to computational shortages, causing Meta to ask employees to save the number of tokens used and transfer the workload to their own internal Muse Spark model. Even Sundar Pichai has publicly admitted that Google's cloud computing revenue could have been higher if it had enough capacity to meet demand. When the head of one of the world's largest infrastructure companies has to admit a shortage, I think it says a lot about the fact that the real limit of the entire AI industry today lies not in ideas or algorithms, but in the number of chips, the amount of electricity, and the speed of physical construction.

So who is really in power?

If we look across the events of the same week, we see that power in the AI ​​industry is distributed in a different order than what the model rankings or revenue numbers show. AI labs like OpenAI, Anthropic, or Google DeepMind still create products that users see and pay for, but they increasingly rely on a small group of physical infrastructure holders, companies that own land, electricity, Nvidia chips, and the ability to borrow long-term capital to build it all before the need actually arises. In that group, the new SpaceX-xAI empire emerged as a surprising factor, when the data center originally built to serve Elon Musk's private AI ambitions has now become a huge source of rental income from his competitors. For me, the question worth watching in the coming years is no longer which model is smarter, but whether the concept of "independent AI labs" still makes sense, when all of them, including the biggest names, have to line up to rent computing power from the same small group of infrastructure owners. If that's true, the biggest prize in the long-term AI race may not go to whoever writes the best algorithm, but to whoever builds the most data centers, locks up the most electricity, and borrows the most capital before the race ends.