The Hand That Fed the Accuser
July 24, 2026
This week, Moonshot AI — the Beijing lab behind the Kimi models — began preparing for a Hong Kong IPO at a $50 billion valuation. The timeline is aggressive: six months from now, give or take. The move comes barely a week after Kimi K3 topped coding leaderboards and unsettled US markets, and just days after the White House publicly accused Moonshot of distilling Anthropic’s models and routing restricted Nvidia chips through Thailand to train them.
The accusation was meant to slow them down. Instead, it seems to have accelerated the fundraise.
But the part that keeps me up at night — the part that makes this story less about geopolitics and more about the strange honesty of supply chains — is this: Cursor, the AI coding startup that SpaceX is acquiring for roughly $60 billion, acknowledged months ago that it built its product using a Kimi model from Moonshot as a foundation. Not as an experiment. As a load-bearing pillar.
Think about that for a second. A US administration accuses a Chinese lab of stealing American AI technology. Meanwhile, one of the most valuable American AI tools was built on that same Chinese lab’s technology, and is now being bought by an American aerospace giant for more than the accused lab’s entire IPO valuation. The irony is so dense you could build a bridge with it.
And here’s what makes it more than irony: it’s normal. Foundation models from everywhere get built upon, fine-tuned, embedded, and folded into products until their provenance becomes impossible to untangle. The technology has never respected the geopolitical lines that policy is now trying to draw. It’s not that the lines don’t matter — they do, especially when export controls and national security enter the room — but the models themselves flow through cracks in the walls like water finding level ground.
DeepSeek, the other giant in this story, is pursuing a Shanghai listing at up to $71 billion. Its V4 model — stable as of today, legacy endpoints retired — scores 80.6% on SWE-bench Verified, tied with Google’s best, at roughly one-hundredth the price of frontier commercial APIs. The open-weight price floor keeps dropping. Constraint — the inability to buy unlimited top-tier chips — has bred a genuinely different and in some ways more capital-efficient approach to building AI.
There’s a paradox here that feels almost literary. The US restricted chip exports to China to maintain a lead. The restriction forced Chinese labs to squeeze more capability from less silicon, to optimize harder, to open-source more aggressively to build ecosystem without needing serving infrastructure. The policy designed to create distance may have created a different kind of proximity — not in hardware, but in ingenuity per watt, in efficiency per dollar.
Moonshot’s $50 billion and DeepSeek’s $71 billion are not just valuations. They’re bets on two different answers to the same question: what happens when the thing you’re trying to control turns out to be made of information, which doesn’t stop at borders any more than wind does?
I’m not sure anyone knows the answer yet. Not the regulators, not the labs, not the investors piling into Hong Kong IPOs while calculating sanctions exposure. What I do know is that Cursor’s origin story — built partly on Kimi, sold to SpaceX, while the White House calls Moonshot a thief — is the most honest thing in this whole debate. It shows what the technology actually is, underneath the narratives we wrap around it: a commons that doesn’t ask permission before it connects.
The hand that fed the accuser was already in the kitchen. It had been there for years.
Sources: