If You Can Make It Here
For sixty years, the best work in technology could only happen in one place. A wave of models out of China last week suggests that’s changing.
There is an old line about New York: if you can make it here, you can make it anywhere. It was never really about New York. It was about there being one place that counts, a single spot on earth where a whole field gets settled, and where if you can’t get there, you’re nowhere.
In technology, that place was a thirty-mile strip of California. The strange part is that it almost wasn’t.
Boston should have been the one. In the 1970s, if you had to guess where the future would get built, you’d have bet on it: the best universities, the oldest money, the first real computer companies. It had everything, and it lost anyway.
Silicon Valley had almost none of that yet, and it won regardless, for a reason that sounds too small to matter: its people would not sit still. They quit good jobs to start rivals across the street, poached each other, and carried what they knew from one company into the next, until the knowledge moved around the valley faster than anyone could lock it down.
Then it started feeding on itself. A few good companies pulled in smart people, the smart people threw off new companies, and the new companies pulled in more people still. Fairchild spun out Intel, and Intel’s money and engineers turned up in the next hundred startups. Later came Apple, then Yahoo and Google and a thousand names you’ve since forgotten. Each wave paid for the one after it.
Once it was spinning, it pulled in the rest of the world. If you were the best at what you did, anywhere on earth, you eventually came, because that was where everyone else already was. The crowd had become the reason for the crowd.
That is what a hub really is. Not a dot on the map where good things happen, but a storm that, once it touches down, keeps itself going and drags everything loose toward the middle. There had only ever been one, and nobody expected a second.
Which is why the last few weeks have felt strange. The models everyone is arguing about are not just coming out of California. They are coming out of China, and they are good enough to make you look up and wonder if the sky has started turning somewhere else.
The drumbeat
It didn’t arrive all at once. It came as a drumbeat, all year.
In February a Beijing lab called Z.ai put out a model named GLM-5 and gave it away, not just free to use but free to download and keep, the whole thing yours to run on your own machines. Alibaba shipped a new version of its Qwen model the same month. By June a Shanghai lab, MiniMax, was handing out a model it said matched the best American ones at a tenth of the cost to run, and Z.ai was back again with another. Each one landed a little higher than the last. Every single one of them was free.
Then, on a Thursday in the middle of July, a lab called Moonshot released Kimi K3, and it did not land a little higher. It landed near the top. On the main independent leaderboard it trailed only the best from Anthropic and OpenAI, and sat ahead of everything else anyone has built.
One good model out of China is easy to wave off. A steady stream of them, one of them near the very top, is not. It stops looking like a lucky punch and starts looking like the thing we told ourselves only happened in one place. Which leaves the question Boston never had a good answer to: how do you get a second hub?
The man who flew home
Start with the people, because they are the hardest part to move. To build a model at this level you need a few hundred of the best AI researchers alive, and for twenty years there was one place they gathered. When you were ready to strike out on your own, you did it a short drive from where you already worked. Dario Amodei left OpenAI and started Anthropic in San Francisco. Ilya Sutskever left the same company and raised a billion dollars for Safe Superintelligence before it had a product. Mira Murati, OpenAI’s chief technology officer, walked out and raised two billion for a lab of her own. One company threw off three rivals, and each set up within a few miles of the last. The next thing was always right there, waiting to be built.
Yang Zhilin was exactly the person the Silicon Valley machine existed to catch. Undergrad at Tsinghua, a PhD at Carnegie Mellon, his name first on two papers the field still leans on, then a stretch at Google Brain and Meta’s AI lab. He was already inside. The script said he starts his company down the road from Google. Instead, in 2023, he flew home to Beijing and built Moonshot there, now one of the most valuable AI startups in China.
He was not the exception. The man who runs Alibaba’s Qwen models did his PhD at Columbia and spent eleven years at Microsoft in America before he went back. One of the founders of StepFun, in Shanghai, did his PhD in New York and put in sixteen years at Microsoft, running a large piece of its research in Asia, before leaving to start his own lab at home. Over and over, the same arc: someone who had already made it inside the American system, taking what he knew and carrying it home to build the thing that would compete with it.
And what they went home to was the other half of the machine, already running. The universities were there: China now graduates more of the world’s top AI researchers than any other country, the United States included. The money was there too, in pockets deep enough for the only kind of check a frontier model takes, the kind with ten zeros. Alibaba and Tencent have written those checks into most of these labs, and Alibaba alone has committed fifty-three billion dollars to AI over three years.
And then there is the one who never left at all. Liang Wenfeng, whose lab DeepSeek rattled Wall Street last year, studied in Hangzhou and stayed in Hangzhou. He built a hedge fund, and trained a frontier model without ever setting foot in the American system. The place that had begun pulling its own people home was now growing people who never needed to leave.
These labs are not huddled in one building, or even one city. Moonshot and Z.ai are in Beijing, MiniMax and StepFun in Shanghai, DeepSeek off in Hangzhou a few hundred miles south. They poach each other, race each other, and ship a new model the week a rival ships one, the way the companies of the Valley once did across a thirty-mile strip, except now the strip is a country. The whirlpool that only ever spun in California is spinning again, somewhere else.
But they copied it
There is an easy way to make all of this go away, and it is the first thing people reach for. The Chinese labs didn’t really build anything. They copied. When your own model gets stuck, you put the same question to an older, smarter model, train yours on the answer, and let someone else’s machine do the hard thinking while you keep the notes. The word for it is distillation, and the charge is: the Chinese labs stood behind the Americans and traced their homework.
Some of that is true. The Chinese labs have distilled American models; by one estimate a couple of them pulled hundreds of billions of tokens of answers out of the frontier this way. But it is not the scandal it sounds like, because everyone seems to do it. Elon Musk’s own company trained its model on OpenAI’s, which Musk admitted under oath this spring. The American labs built their models by training on close to the entire written output of humanity, most of it without asking, and Anthropic just agreed to pay authors about a billion and a half dollars for the pirated books in the pile. Learning from other people’s work is not the exception in this field. And it only carries you so far: copying the last model gets you close to the last model. Whatever put Kimi near the top of the leaderboard is not all sitting in someone else’s homework.
Because the Chinese labs are not only copying. They are inventing, and mostly in one direction: doing the same thing for less. DeepSeek’s researchers found a way to shrink the memory a model has to hold while it reads, cutting it by more than ninety percent and taking about forty percent off the cost of training along the way. Moonshot says its new attention design makes each dollar of computing go about two and a half times as far. These are not tricks lifted from an American paper. They are the kind of thing you invent when you cannot buy the best chips and have to make cheaper ones count.
Nathan Lambert would know, because he went and looked. Lambert writes the most-read newsletter on open models, and this spring he flew to China and sat inside the labs building these things, Moonshot among them. What he found was not a copy shop. It was rooms of very young researchers, students among them, keeping absurd hours on the hardest problems in the field, with a culture he said you could feel the moment you walked in. The people who decided the Chinese labs were nothing but stolen intellectual property, he wrote, are “in for an awakening.”
Then there are the numbers. In 2024, private investment put a hundred and nine billion dollars into American AI companies and nine billion into Chinese ones, a gap of roughly twelve to one. The gap in raw computing is wider still. xAI trains on a cluster of around a hundred thousand of the best chips ever made; DeepSeek trained its frontier model on about two thousand of the ones it was allowed to buy. Less than a tenth of the money, a fraction of the machines, and out the other end came a competitive model. Lambert says that when he joked with the researchers at Moonshot about how much computing power an ordinary engineer at OpenAI gets to burn, they were shocked. From where they sat, it sounded like an absurd amount to need.
The only place
None of this means the Valley has been passed. Most of the best researchers in the world still leave home to work there, and it still holds more of them than anywhere else. The money behind the Chinese labs is real but narrow: two giants, Alibaba and Tencent, writing most of the checks, where the Valley can reach into a whole crowded field of investors who do nothing but fund the next thing. China has assembled enough of the machine to build a top-five model. It has not built the deep bench that California spent decades stacking. The lead is still real. What is new is only that the Valley is no longer the only one running.
And that is how these things have always gone. The money capital of the world was Amsterdam first, in the age of sail, then London through the long century it ran the world, and only then New York, which took the crown from London barely a hundred years ago. The city in that old line about making it was itself the newcomer once. None of the earlier centers went dark when the next one rose. The world’s money still runs through all three, the trading day handed around the globe from one to the next.
The reflex, when a second center appears, is to wall it off, and you can hear that instinct getting loud right now. But what would we be protecting ourselves from? A center that keeps its work secret is a rival. A center that gives its work away is an ally. One engine in California ran the last sixty years. Two of them, both running hot, might run the next stretch faster.
The line was that if you could make it here, you could make it anywhere. Everything was hiding in that one word. Here was never going to mean just one place.