The Price of AI Reasoning

Who Can Afford to Think?

For a while now, a nagging question has lurked behind the AI infrastructure boom. The logic for spending hundreds of billions on Nvidia GPUs for training new models was clear. But what happened after that? The general assumption was that inference—the actual use of a trained model—would become progressively cheaper and more efficient. This created a potential long-term problem: if the ultimate goal is cheap, ubiquitous AI, does the insatiable demand for cutting-edge hardware eventually plateau or even decline?

Then came the news of the AI models from Google and OpenAI achieving gold medal at this year’s International Mathematical Olympiad, which suggest an entirely different future. It turns out that the path to true AI reasoning isn’t just about building a better model; it’s about being willing to pay an exorbitant price to make that model think.

What makes Google’s and OpenAI’s gold-medal performance so striking is that it directly challenges the long-held critique of modern AI. For a while now, the knock against large language models was that they weren't truly “reasoning” but rather engaged in a sophisticated form of pattern-matching that only mimicked intelligence. This achievement challenges that notion. And unlike the AI that conquered Go or Poker, which were hyper-specialized systems trained for a single task, the models that tackled the Olympiad were generalists—the same kind used for language, coding, and science. The secret wasn’t a more efficient algorithm, but rather a profoundly inefficient—and powerful—new method.

The report from Reuters gets to the heart of it:

OpenAI’s breakthrough was achieved with a new experimental model centered on massively scaling up "test-time compute." This was done by both allowing the model to "think" for longer periods and deploying parallel computing power to run numerous lines of reasoning simultaneously, according to Noam Brown, researcher at OpenAI. Brown declined to say how much in computing power it cost OpenAI, but called it "very expensive."

This concept of “test-time compute” is a fundamental shift. Instead of a model producing one answer in a linear fashion, it is given the time and resources to explore a problem from many different angles at once, effectively running multiple parallel chains of thought to check its own work, identify dead ends, and converge on a correct solution. It is the machine equivalent of a human not just blurting out the first answer that comes to mind, but sitting down for hours to reason through a problem from first principles.

This all but guarantees that the demand for high-end compute will not slow down. It creates a new, ongoing, and potentially bottomless market for it. For any application where precision is non-negotiable—scientific research, engineering, medicine, finance—the option will exist to pay a massive premium for a higher degree of certainty.

The catch, of course, is that word: “expensive.” OpenAI has already stated it won’t be releasing this capability to the public for months, because the cost is simply too high. This creates a new and formidable barrier to entry in the AI race. The ability to deploy models capable of advanced reasoning is no longer just a question of having the right algorithm; it’s a question of having the capital to afford the eye-watering compute bill.

Perhaps it no longer matters whether it’s true reasoning or just extremely convincing pattern-matching; if solving the hardest problems is simply a question of how much AI compute you’re willing to buy, the race for artificial general intelligence has truly become a race for capital.

Stargate is Stalled

The AI infrastructure boom can be understood as a project of almost unimaginable scale, requiring presidential announcements and half-trillion-dollar commitments from Japanese billionaires and the world's most famous AI labs. Six months ago, this is exactly what happened. President Trump, SoftBank’s Masayoshi Son, and OpenAI’s Sam Altman gathered at the White House to announce Stargate, a $500 billion joint venture to build the data centers that will power the future.

The problem, it turns out, is that the future is in a hurry and can’t wait for its own official infrastructure project to get started. The Wall Street Journal reported this week that the grand venture has gotten off to a rather sluggish start:

Six months after Japanese billionaire Masayoshi Son stood shoulder to shoulder with Sam Altman and President Trump to announce the Stargate project, the newly formed company charged with making it happen has yet to complete a single deal for a data center.

Son’s SoftBank and Altman’s OpenAI, which jointly lead Stargate, have been at odds over crucial terms of the partnership, including where to build the sites, according to people familiar with the matter.

While the companies pledged at the January announcement to invest $100 billion “immediately,” the project is now setting the more modest goal of building a small data center by the end of this year, likely in Ohio, the people said.

It is worth asking why a project with this much money, momentum, and political backing would immediately get bogged down in disagreements over where to build things. The answer, it seems, is that building the physical world is much harder than announcing plans to build it. The AI race has moved from the realm of software algorithms to the brutal logistics of industrial-scale construction. You don’t just build a 5-gigawatt data center campus. You must first find a place with enough land, then secure a power contract equivalent to what a nuclear reactor puts out, navigate local zoning and environmental permits, and then get in a very long line for Nvidia's chips.

This process is slow, complicated, and apparently too slow for OpenAI. The truly weird part of this story is not that Stargate has stalled, but that OpenAI, its supposed partner, has already moved on and sourced its compute elsewhere. While the joint venture was arguing about site selection, Altman went out and signed one of the largest cloud deals in history with Oracle—a deal for 4.5 gigawatts of capacity that will cost OpenAI some $30 billion a year.

In a sense, the AI boom is so real that the official, headline-grabbing $500 billion plan is being outpaced by the actual, frantic, multi-hundred-billion-dollar scramble happening on the side. The demand for AI compute is not a future problem to be solved by a grand venture; it is an immediate, existential crisis to be solved by whoever can deliver the gigawatts first. Building the future of AI, it turns out, is less like a well-organized moonshot and more like a mad dash for the last helicopter out of Saigon.

The Scoreboard

• AI: Google Clinches Milestone Gold at Global Math Competition, While OpenAI Also Claims Win (Reuters)
• AI: Leaked Memo: Anthropic CEO Says the Company Will Pursue Gulf State Investments After All (Wired)
• Browser: Duckduckgo Now Lets You Hide Ai-Generated Images in Search Results (TechCrunch)