July 2026: The Month Intelligence Got Cheap

On the sixteenth of July, a company most people in the West cannot name gave away the largest artificial mind ever built. Moonshot AI, out of Beijing, released Kimi K3—2.8 trillion parameters, free to download—and by one independent ranking it stood fourth in the world, ahead of Anthropic’s Opus, behind only that lab’s flagship and OpenAI’s best. You could have it for the cost of the bandwidth. That same week, a Taiwanese company that makes the chips a mind like that runs on told its investors it would spend another hundred billion dollars—not on research, not on models, but on buildings, in the Arizona desert. Hold those two facts in one hand. The intelligence was free. The ground it runs on now costs a quarter of a trillion.

This was not a fluke of timing. In eleven days that July, at least five frontier-grade models landed—OpenAI’s long-gated GPT-5.6 finally going public, Musk’s Grok 4.5, Moonshot’s Kimi K3, Meta’s Muse Spark, Alibaba’s Qwen3.8—each roughly as capable as the last, several of them cheap, one of them free. For three years the only question that mattered in this industry was whose model was smartest. In July that question quietly stopped being worth asking, because the answer had become: everyone’s, more or less, by next week. And so the money did the only thing money does when the prize turns into a commodity. It went looking for whatever was still scarce. It went down a layer—to the silicon, the power, and the permission to buy them.

The Fortnight of Frontiers

Take the models one at a time and the sameness is the story. Grok 4.5, trained on data from the coding tool Cursor, was pitched by Musk as Opus-class—as good as the best anyone sold—only faster and cheaper, at two dollars a million tokens. Kimi K3 debuted at number one on a widely watched coding leaderboard, past Anthropic’s Fable, and set a date to hand its full weights to anyone who wanted them; on a broad independent tally it landed fourth overall, the first openly published model ever to crowd the very top of the rankings. GPT-5.6 arrived in three tiers named Luna, Terra, and Sol, from merely brilliant to frontier. Line them up and you cannot easily say which is best, and—this is the point—it no longer pays to know. A year ago a new top model was an event. In July it was a Tuesday.

The clearest sign of the shift was how careless the boasting had grown. On the nineteenth, at a conference in Shanghai, Alibaba previewed Qwen3.8-Max—2.4 trillion parameters, which it declared second only to Anthropic’s best model anywhere—and shipped that claim with no benchmarks, no model card, and no weights anyone could yet download, at a tenth of the going price. A world-beating mind, asserted and unproven, tossed out like a flyer. When a product is scarce you guard it; when it is abundant you shout about it and cut the price. By the UK AI Security Institute’s reckoning, as reported that month, the gap between the best freely available model and the best locked-up one had narrowed to something like four to seven months, down from the better part of a year. The frontier was not a fortress anymore. It was a queue, and the line was getting shorter.

The frontier is no longer a place you reach. It is a price—and the price is falling.

Where the Money Actually Went

If the mind is abundant, advantage moves to whatever the mind cannot be run without. That is a chip, and in July every serious lab was trying to build its own. Days before the month began, OpenAI had unveiled its first custom processor—codenamed Jalapeno, designed with Broadcom in nine months, the fastest such effort anyone could remember. On the second, word came that Anthropic was in talks with Samsung to fabricate a chip of its own on a two-nanometre line. Google has quietly made its own tensor chips for years; Meta has its MTIA; a startup called Etched has raised most of a billion dollars on the bet that a chip built for one kind of model can bury the general-purpose graphics cards Nvidia sells. Even China’s labs are at it—DeepSeek designing its own inference silicon, another firm unveiling a homegrown processor the same month. Broadcom, which built OpenAI’s chip and helps make Google’s, is now the silicon partner to two of the three largest American model labs—a quiet kingmaker in a war supposedly about software. The GPU monopoly that funded this whole era is, for the first time, attacked from every side at once.

But a chip design is only a drawing. Someone has to etch it into silicon, and almost no one on Earth can etch it at the size that matters. On the sixteenth, the company that can—Taiwan Semiconductor—reported a quarter in which profit rose seventy-seven percent to a record, and announced it would pour another hundred billion dollars into its fabrication plants in Arizona, lifting its American commitment to two hundred and sixty-five billion and its building budget for the year toward sixty-four billion—and still it could not keep pace with a demand it described as only intensifying. This is the thing you cannot copy over a weekend. You can clone Kimi K3 with enough disk space; you cannot clone a two-nanometre fab, which takes years, tens of billions, and a body of tacit skill held by a single company on an island that two great powers are circling. The model got cheap. The place that makes the machine to run it became the scarcest ground in the world.

The Permission Layer

There is one thing scarcer than a fab, and that is the right to buy what it makes. On the fourteenth, a senior official at the U.S. Commerce Department told Congress that Nvidia’s H200 chips—among the most capable parts it is allowed to export—had begun shipping to China for the first time, under a licensing regime, agreed earlier in the year, that taxes each sale twenty-five percent. Around ten Chinese firms had been cleared, Tencent and ByteDance among them, against some ten billion dollars in approved licences. The quantity actually shipped, the official was careful to add, was tiny.

“It’s a very small quantity of chips.”—U.S. Commerce Department official, to Congress, July 2026

The smallness is the tell. A single sentence in a hearing room now decides how many artificial minds a rival nation may run, and the answer for the moment is: a trickle, taxed. In the same days, Nvidia cut its roster of authorised buyers across Asia by more than half and sent auditors into resellers in Singapore, Malaysia, and Japan, hunting chips slipping through the side doors toward Beijing. And the rationing bit at home too: Moonshot, whose Kimi K3 anyone on Earth could download for free, had to pause new sign-ups because it could not get enough compute to serve them—the very export controls aimed at China throttling the Chinese lab that had just handed the world its best open model. You can download the intelligence. You cannot download the permission to run it.

The Cost Is Physical

Follow the money all the way down and it stops being abstract. It turns brutally physical—land, transformers, fuel. In early July, Anthropic signed a twenty-year lease worth some nineteen billion dollars on a data centre being raised on the bones of a shuttered aluminium smelter in Hawesville, Kentucky—the landlord a former Bitcoin miner that had pivoted, overnight, from minting coins to renting out four hundred megawatts of power. It would sink three or four billion into the site to earn back nineteen—the economics of the age in a single line, where the scarce asset is not the code but the concrete and the power hookup. A Texas campus went the same way for nearly ten billion, a full gigawatt spoken for. The old industrial map of America is being quietly repossessed by the AI build-out, smelter by smelter, substation by substation.

And where the grid cannot deliver fast enough, the companies simply make their own power, permits or not. Outside Memphis, across the state line in Southaven, Mississippi, Musk’s xAI ran fifty-nine gas turbines to feed a data centre it calls Colossus 2—without the federal clean-air permits such turbines require, a Reuters analysis found in July. The emissions settle over neighbourhoods that are largely Black and already carry some of the region’s highest rates of lung disease; the NAACP and the Southern Environmental Law Center are suing to shut the turbines down. This is what the cloud is made of, at the bottom of the stack: not weightless computation but burning gas beside somebody’s home. The intelligence may be free. The air it costs is not.

What the Numbers Don’t Say

A month this loud deserves a cold eye. Many of July’s biggest figures are promises, not receipts: the nineteen-billion lease and the ten-billion Texas campus are decades of hoped-for revenue, not cash in a drawer; Etched’s world-beating chip is still pre-production; Anthropic’s Samsung chip is a conversation that yields no silicon before late 2027; Alibaba’s second-best-on-Earth model arrived without a shred of evidence. Nor was all of July about the metal—the FDA cleared new medicine, SpaceX scrubbed its thirteenth Starship flight on the sixteenth to swap two engines, and a widely shared result showed an ordinary laptop cracking a problem long thought to need a quantum computer, a reminder that some frontiers dissolve the instant you look hard. But the month’s gravity was unmistakable, and it all pulled one way: down, toward the chips and the current. The costly skill now is telling a press release from a working thing.

The Ground Beneath the Genius

So here is what July 2026 revealed, once the launch livestreams went quiet. We have spent three years transfixed by the intelligence itself—how it reasons, what it can pass, whether it understands. And in a single month the intelligence became the abundant thing, the commodity, in places literally the free thing, while everything required to make and run it—the fabs, the reactors and turbines, the licences, the water and the wire—turned into the scarce and guarded and fought-over thing. The value did not vanish. It sank. It settled out of the software and into the physical world beneath, where it is harder to see, harder to copy, and owned by very few.

We were watching the mind. The month was about the ground it stands on.

There is an old fear that the machines would begin to think for us. July suggests a stranger inheritance. Thinking, it turns out, is becoming the easy part—cheap, plentiful, handed out for free by a lab in Beijing. What stays hard, and therefore what stays powerful, is everything underneath: the sand milled into chips in one Taiwanese company’s fabs, the current pulled off a straining grid or burned out of unpermitted turbines, the single sentence in a hearing room that decides who may buy what. We keep our eyes on the dazzling, articulate surface. The month’s quiet lesson is to look down—at the desert, the substation, the smelter, the neighbourhood learning to breathe around the machines—because that is where the future is actually being poured, and where the few hands that will hold it are already closing.