David Sacks, a Trump adviser and venture capitalist, has been circulating cases of American companies quietly switching to Chinese LLMs—because the domestic models refuse to do the work.
Sit with that for a second. The same guardrails the US government mandated to keep American models "safe" are pushing American firms toward Chinese ones to close security gaps. The protectionist logic eats itself: the safety regime built to advantage US labs is now a reason to leave them.
That contradiction is the whole story this week.
Last week, in "Agents Become Infrastructure," we tracked whether the frontier labs would reabsorb the stack technically—swallowing the harnesses, orchestration layers, and tools that sit on top of raw models. This week the reabsorption bid went regulatory: the labs are now trying to capture the stack through Washington rather than through code.
The trigger was Kimi K3. On July 16, Moonshot AI released a 2.8-trillion-parameter open-weight model that landed at #3 on Artificial Analysis's Intelligence Index—beaten only by Claude Fable and GPT-5.6 Sol, and cheaper than both.
Nathan Lambert, who visited the Kimi team, put it bluntly: the open-to-closed gap, long debated at six to nine months, has collapsed to something like three to five. The AEI's Ryan Fedasiuk went further—"a matter of weeks."
And Kimi is not distillation cosplay. Lambert's read after seeing the model: if adversarial distillation from US frontier models helped, it was marginal. Chinese labs can now build, not just fast-follow.
That is the backdrop for the moment the subtext became text.
Dean Ball, OpenAI's head of strategic futures, publicly argued that the US government should find a pretext to manufacture regulatory "fear, uncertainty, and distrust" around the new open models—on the explicit logic that open weights "must necessarily deter capital spending by the frontier labs."
Read that reasoning again. The stated harm is not to national security. It is to capex.
Ball retracted within days, after Yann LeCun and Martin Casado savaged the argument—walking back both the claim that a regulatory crackdown was the White House's "best strategy" and the claim that open weights necessarily slow the technology.
But you cannot un-say the quiet part. The justification he reached for first was margin defense; only the pushback dressed it back up as something else.
The justification he reached for first was margin defense; only the pushback dressed it back up as something else.
So examine the something else. The "security" case against Kimi K3 comes in three flavors, and each collapses on contact.
One: data leakage to Beijing. The US banned modern Chinese EVs over data-gathering fears, and the analogy gets stretched to models. But experts note that open-weight models running on US servers are unlikely to phone home—the weights are static, the inference is yours. Not impossible, but not the EV telematics problem either.
Two: implicit PRC bias baked into the model. Maybe. But it is genuinely unclear what pro-Beijing bias means for, say, refactoring a React component. The bias worry evaporates the moment you look at what these models are actually used for.
Three: missing guardrails. Chinese models supposedly lack the safety constraints Washington mandated—through an opaque process—to stop LLMs from exploiting systems or building weapons. Except this is exactly where Sacks's examples bite: those same guardrails make US models refuse legitimate security work, and companies route around them to Chinese models. The guardrail gap is a sales advantage for Kimi, not a threat vector.
Three security arguments, three collapses. Meanwhile the driver nobody in the lab wants to name is sitting in the pricing tables.
Blend input and output at a 90/10 ratio and the median open-weight frontier model runs about 15% cheaper than GPT-5.2. The cheapest, DeepSeek V4 Flash, runs roughly 90% cheaper.
That is the fear. Not backdoors—commoditization.
Braden Hancock of Snorkel AI says it without euphemism: "Strong, frontier-caliber open source models will place a squeeze on the margins and will bring down the prices of the frontier companies." Crucially, he adds, this does not shrink AI usage—the opposite. It just relocates the spend outside the closed labs.
That relocation is precisely the threat to a company that has raised on the premise that intelligence is a scarce, high-margin good. If frontier-adjacent capability is available at a 90% discount on independent infrastructure, the return on a multibillion-dollar training run compresses. Ball's "capital spending" line was honest about which capital.
Which is why Sam Bresnick at Georgetown's CSET frames the whole exercise as a category error. "Why should the weight of the U.S. government be aimed at protecting these companies from competitors that are being locked out from the U.S. market based on their origins?"
And here is the strategic own-goal, laid bare by Bresnick's own prescription.
If Washington genuinely wanted to slow China, the lever is chips, not apps. Stop selling Nvidia H200s to China. "That," Bresnick says, "could potentially keep us out of this thorny debate about banning open source technologies that huge numbers of US companies want to use."
That is the tell. The one measure that actually constrains Chinese frontier training—export controls on the compute—is the one the labs' allies are not pushing. Kimi K3 was built under "persistent hardware/compute capacity constraints," per Bank of America; more H200s ease exactly that bottleneck.
So the proposed policy gets the incentives backwards on both ends: ban the cheap models Americans want to use, keep selling the chips that let China build the next one. You would struggle to design a posture more precisely tuned to protect lab margins while doing nothing for national security.
Now the part that should actually worry a strategist, and it has nothing to do with weights leaking secrets.
Hancock's frame is ecosystem capture, and his analogy is PyTorch. It became the industry standard because it was open—the whole community contributed, and the rival libraries "kind of died in comparison." Open didn't lose. Open ate everything.
Apply that to models. The danger from open Chinese models "is less that they're sneaking in back doors, and more that they are owning the innovation." Open weights create an expanded, unpaid workforce building on your foundation.
And the diffusion is already underway. US graduate programs, Hancock says, now mainly build on open-weight Chinese models, and roughly half the papers students study come from Chinese institutions—as American frontier labs grow "increasingly reticent about sharing their work widely."
That is the compounding loss. Every grad student who fine-tunes on Kimi, every paper that benchmarks against it, every tool built to its interface deepens the dependency. The commons tilts, and commons don't tilt back quickly.
Xi Jinping made the strategy explicit this week, committing China's AI future to open source and global diffusion at the World AI Conference—the same week as the strongest open-weight release to date. Clem Delangue of Hugging Face argues the American answer of restriction "wouldn't make AI safer. It would simply hide the risks, concentrate power in the hands of a few."
Note what banning Kimi does to that fight. It does not slow Moonshot, which trains and ships regardless. It removes American researchers, startups, and academics from the ecosystem forming around these models—ceding the contributor base to everyone else.
The Sequoia framing captures the trap: the weights are open, but the dependency is not. You can download Kimi; you cannot un-download the research culture that grows up around it.
None of which means the closed labs are doomed. Tomasz Tunguz's read is more interesting: closed models pull ahead, open models catch up, the whole market accelerates. OpenAI already cut inference costs 50%; Kimi shipped a new attention architecture; Anthropic is about to post its first profitable quarter. Competition, not protection, is keeping the race fast.
The honest position is that neither business model is solved—open or closed, in the US or in China, everyone is still hunting for durable revenue against rising training costs. That uncertainty is exactly why a lab might prefer to litigate the question in Washington rather than in the market.
But a policy built to defend margins will not defend the frontier. It will just make the frontier someone else's commons.