The Best AI Will Be Free, Chinese, and Banned in America

Kimi K3 just matched the frontier. If you watched who built DeepSeek, you saw this coming 18 months ago.

So this morning Moonshot AI released Kimi K3. 2.8 trillion parameters, which makes it the largest open-weight model anyone has ever shipped, and the weights go public in ten days. On the long-horizon benchmarks (the ones that actually resemble real work) it keeps up with Claude Fable 5 and GPT-5.6 Sol. On a few of them it wins.

And honestly, I wasn’t surprised. Not because I’m a genius, but because this was basically announced in January 2025. You just had to pay attention to who built DeepSeek instead of what it scored. That’s what I want to walk you through, because it changed how I think about speed, efficiency, and what “survival of the fittest” actually means.

Not a clean sweep, but an open-weight model beating closed frontier models on several benchmarks.

The tell was who built DeepSeek

Quick rewind to January 2025. DeepSeek releases R1, publishes that the base model cost about $5.6 million to train, and Nvidia loses something like $600 billion in market value in a single day. Everyone spent that week arguing about whether the $5.6M number was real. I got stuck on a completely different detail: DeepSeek wasn’t built by a tech giant. It came out of High-Flyer, a quant hedge fund. The name literally means “magic square” in Mandarin.

Which is genuinely weird, right? China has Alibaba, Tencent, ByteDance. Armies of engineers, basically unlimited money. And the lab that cracked frontier-level AI on rationed hardware was run by a fund manager. Liang Wenfeng had quietly bought around 10,000 Nvidia A100s before the export ban hit in October 2022. His own business partners thought it was an eccentric side project.

timeline
    2016 : High-Flyer founded, AI-driven quant trading
    2019 : Firefly One cluster, 1,100 GPUs for trading models
    2021 : Liang buys ~10,000 A100s, "the side project"
    Oct 2022 : US export controls hit, top chips banned for China
    2023 : DeepSeek spun out of High-Flyer
    Jan 2025 : R1 released, Nvidia drops ~$600B in one day
    Jul 2026 : Kimi K3, open weights match the frontier

Once you know that, R1 stops looking like a shock and starts looking like a prediction. And Kimi K3 today is the same prediction paying out again.

Why the quant background matters

Okay, so why do I care so much about the quant part? Because a quant’s whole job is making things unreasonably fast. If you want to feel what that actually means, read Flash Boys by Michael Lewis (I recommend it to anyone who touches software). It opens with a company secretly drilling a fiber line in a dead-straight path from Chicago to New Jersey. Through mountains. Under rivers. Because a straighter cable saves about three milliseconds, and traders paid fortunes for those three milliseconds.

You don’t win that game by writing clean Python. You win it by knowing exactly what the hardware is doing: cache lines, network hops, kernel bypasses. And here’s the part people miss. Quants aren’t just low-level engineers. Every microsecond they shave has to show up on a real scoreboard, the market. So you get deep systems knowledge that’s permanently connected to consequences. If someone hands you rationed GPUs and says “compete,” that’s exactly the person you want.

The speed-over-efficiency doctrine

Now compare that to the American side. Silicon Valley literally has its operating system written down in a book: Blitzscaling, by Reid Hoffman. The core definition is, almost word for word, prioritizing speed over efficiency in conditions of uncertainty. Grow first, understand later, let the mess pile up, capital will cover it.

I don’t buy it, and here’s my honest reason: look at what Hoffman actually built. LinkedIn. A network that grew through novelty and growth hacks (remember the “import your entire address book” era?) and has been a mediocre product for twenty years. Blitzscaling gets you distribution. It doesn’t get you craft. And in AI the same instinct turns into: the run underperforms, you don’t ask why, you buy more compute. If your product is a black box to you, improvement is just more compute, more data, and praying.

I see the mini version of this in my own agent work all the time. Tokens feel free, so nobody audits the context window. Tools pile up, prompts bloat, and the fix is always “use the bigger model.” The first time I put a hard budget on one of my pipelines, a third of the context turned out to be dead weight. Free resources hide what you don’t understand.

Chinese labs don’t compete like American labs

Everyone knows the export restrictions capped what Chinese labs could buy. The part that gets missed is the second ingredient: Chinese labs don’t compete against each other the way American labs do. OpenAI, Anthropic, and Google guard their weights like state secrets, so every efficiency trick gets discovered three separate times behind three separate walls. The Chinese labs open-source. DeepSeek at 1.6T, Xiaomi at 1.02T, Alibaba, GLM, now Kimi K3 at 2.8T with weights public in ten days. So when one lab figures out how to squeeze more out of a crippled interconnect, all of them have it next month. The constraint set the problem, and open weights let them solve it as a group.

And here’s the thing you can’t fake: you cannot be super efficient without understanding the system really, really well. Every trick in that shared toolbox forces you to learn how the machine actually works:

Each of these is a lesson US labs can skip. Most do.

So one person got forced through every line of that list and now holds a working model of the entire stack in their head. The other person never had to bother and holds, essentially, a procurement relationship. My money is on the first person every single time the game changes even slightly.

Distillation is not copying

Speaking of distillation: most people don’t actually know what it is. The word sounds like theft, so the debate stops at “they copied GPT.” But the real goal of distillation is pushing the Pareto frontier of cost versus performance. Same capability for less money, or more capability for the same money.

The mechanics are almost embarrassingly simple. You generate outputs with your own model, generate outputs from a frontier model, compare, keep the best, train on that. Best-of-n distillation. In hiring terms: instead of hiring the American engineer, you ask the AI model that engineer built to verify your work. The frontier model becomes your unpaid senior reviewer, and that’s how you catch up way faster than figuring everything out on your own. Also, this isn’t some China-only trick. Grok and OpenAI have been doing the same thing.

But here’s what actually changed with Kimi K3, and why I think it matters more than the benchmark table. A distilled model is supposed to be smaller and cheaper than its teacher, and a little less capable. That’s the deal. K3 breaks the deal: it’s bigger than Fable 5 and beats it on some benchmarks. And the timing doesn’t even work for pure student-teacher copying. A frontier model takes three to six months to develop, and Fable 5 simply hasn’t been out long enough to supervise K3’s full training run. They took distillation and improved on it.

The student is supposed to stay smaller and weaker than the teacher. K3 didn't.

We’ve seen this exact movie before, with cars. Chinese cars used to be cheaper and worse, and everyone filed them under “imitation.” Now they’re cheaper and better, and the incumbents are the ones lobbying for tariffs. Cheaper-but-worse is a phase, not a destiny. Kimi K3 is what the end of that phase looks like in AI.

Expect the backlash

Which brings me to the prediction in the title. I mean it seriously, and the reasoning is just following the money. The valuations of the closed American labs, plus a big slice of Nvidia and the hyperscalers, rest on one assumption: frontier capability stays scarce and stays paid. Trillions of dollars depend on that assumption. A free, open-weight model that matches the frontier doesn’t just compete with those companies. It attacks the premise their prices are built on.

And nobody defending trillions says “please stop releasing free models, it hurts our margins.” They say “safety.” So my bet is an anti-open-source movement in the US, dressed in safety language: weight-export rules, liability regimes that make hosting an open Chinese model legally radioactive. Basically the car-tariff playbook, rewritten for weights.

Then add geography. Europe has no frontier lab of its own, and right now its relationship with Beijing is honestly in better shape than its relationship with Washington. If the best free model is Chinese and the American ones cost real money, Europe’s choice makes itself. So the end state is exactly what the title says: the best AI in the world is free, Chinese, and banned in America, while the rest of the world quietly builds on it. I’m half joking when I say teach your kids Chinese. Only half.

Survival of the fittest, read correctly

One more thing, because this reframed the whole story for me. People misread “survival of the fittest.” They picture the fastest, strongest animal. That’s not what it means. Fittest means best fitted to the environment. When the environment shifts, the traits that win shift with it, and yesterday’s speed advantage becomes today’s dead weight.

And the AI environment is shifting right now. Inference cost is starting to matter more than training cost. Energy is becoming the binding constraint even for labs that can buy any chip on earth. Efficiency is quietly becoming existential for everyone. One group of labs was forced to practice it for four years. The other group treated it as optional. So no, Kimi K3 isn’t an anomaly. It’s what being fitted to the environment looks like once the environment actually changes.

If you build with AI, this question transfers directly to you. Here’s the self-test I keep coming back to in my own work:

Is your product a black box to you?

1. When quality drops, is your first move to investigate, or to upgrade the model?
2. Can you name what every tool and token in your pipeline earns?
3. If your compute budget were cut in half tomorrow, would you know where to start?
4. What did you last remove, and what did removing it teach you?
5. Which layer of your stack have you never looked inside?

Question 5 is the uncomfortable one. It was for me.

Speed you can rent. Fitness you have to earn, by actually understanding the system you’re betting on. The quants figured that out before the rest of us. Kimi K3 is the proof.