The Take — You can't ban your way to an open-weight ecosystem

Share
The Take — You can't ban your way to an open-weight ecosystem

The House report on Chinese AI models gets the diagnosis exactly right and the remedy backwards. It found that American companies buy Chinese open weights because nothing American competes on price, customization and near-frontier performance — then recommended excluding the Chinese models without building the alternative first. That is not a supply policy. It is a price hike with a national-security preamble.

The facts are narrow and worth stating plainly. A report from the House Committee on Homeland Security and the House Select Committee on China, shared with CNBC ahead of publication on Friday, says Cursor plans to "explore phasing out PRC-developed open-weight models because of geopolitical risk," according to CNBC. Cursor, which SpaceX acquired at an implied $60 billion equity value in August, built its Composer 2 coding model on top of Moonshot AI's Kimi K2.5. Airbnb and DoorDash also engaged with the investigation. None of the companies would discuss it on the record.

What the committee actually wrote

The report's finding is unusually candid: cost, customization and near-frontier performance drove adoption, and the United States lacks a competitive open-weight alternative "at price and performance the market demands." That sentence is an indictment of American labs, not of Chinese ones, and the report knows it, which is why the recommendations split in two.

The good half is a "national American open-weight initiative" combining public computing, shared evaluation infrastructure, and "lawful access" to American frontier models for training. Public compute and shared evals are cheap, uncontroversial and genuinely useful — the evaluation layer is where a government can buy down risk without picking winners.

The other half is a prohibition: extend the October order that banned DeepSeek from federal civilian systems into a bar on any "PRC-developed AI models" in federal information systems and in the performance of federal contracts.

On its own, that second half is defensible. It is the government's own money, spent on its own systems, and governments refuse suppliers for reasons short of proven malice all the time. What I object to is the sequence — exclusion now, supply later, with an initiative whose design contradicts its own name.

Cursor's answer is the tell

Ask what Cursor would phase into and the story stops being about open weights at all. SpaceX develops the Grok family, and Cursor has been involved in training it. So the "American alternative" in this specific phase-out is a model from the parent company, not a downloadable competitor anyone else can use. That is an affiliate transfer dressed as de-risking: it raises Cursor's own cost of goods and removes the price anchor a cheaper model provided.

That is the shape of the policy in practice. Exclusion does not create an American open-weight ecosystem; it creates a captive internal market for whoever already has a frontier model, priced accordingly.

The case I have to take seriously

Dependency really is the point, and the report says so: "Beijing is now seeking to turn those technological gains into global market share and long-term dependence." That is not paranoia, it is a coherent strategy with a visible mechanism.

Open weights cannot be recalled. A licence can be revoked, an API can be cut off, but a downloaded checkpoint keeps working, and a fine-tune can strip whatever refusals a lab shipped with it. China's 2017 National Intelligence Law obliges domestic organizations to assist state intelligence work. And the surrounding moves are real: Commerce published an interim rule on Monday closing remote access to American chips through third-country data centers, and Washington has been circling the distillation route for months. A federal contractor writing code against an endpoint answerable to a foreign government is a legitimate worry, and the narrowest possible lever — the government's own procurement — is the right place to act on it.

I also have to concede that a House committee report is not a ban. It binds nobody. The committees did not propose reaching into private-sector deployments, and the White House's fallback has been disclosure rather than prohibition. If I am arguing against a ban, I should be honest that no ban exists yet.

Why the take still holds

First, prohibition without supply moves usage, not capability. The October DeepSeek order did not send federal-adjacent buyers to American open weights — there were none at the price. Chinese models were the three most-used by tokens on OpenRouter in October, and open-weight models account for more than 71 percent of tokens among its most-used, as Today in AI — November 18, 2026 noted. That same roundup recorded the administration weighing a name-and-shame rule requiring companies to post when they use Chinese models — a disclosure regime is what you adopt when you have concluded you cannot enforce a ban.

Second, an "open-weight initiative" that runs on permission is not an open-weight initiative. "Lawful access to American frontier models for training" means the labs must agree to lend the thing that makes them valuable. The labs that would grant that access are the same ones that gain from excluding a cheaper competitor. You cannot convene a cartel and call the output an ecosystem, and a program that funds compute but cannot buy a release decision will produce benchmarks, papers and pilot projects, not downloadable weights at Kimi's price.

Third, the alternative has to exist at the edge cases, not just the benchmark. Suno Visuals turns songs into video, priced in separate credits covers a US product with more than 100 million users whose video model, H3 Move, is based on MiniMax's H3 — Chinese weights whose licence excludes the US, EU, UK and South Korea from self-hosting. Suno did not pick that because it loves Beijing. It picked it because no American model did the job at that price, and the licence problem was the cost of admission.

Fourth, voluntary phase-outs ahead of rules are the worst regulatory form: no definition of "PRC-developed," no timetable, no standard to test, no appeal. "Considering exploring a phase-out" costs Cursor nothing today and shifts a burden onto every buyer who now has to justify a cheaper model to a procurement officer and a board. And the bill does not vanish. As I argued in The Take — The open-weights gap is real, and it is a price list, the gap at the downloadable frontier is real — roughly 26 index points in October — but the price list is the whole reason anyone buys there. Delete the cheap tier and the spend moves to Grok and to the American labs at American list prices. That is a defensible policy outcome. Nobody should sell it as security.

What would change my mind

Three things, and I would take any of them.

Ship the substitute. If an American lab releases weights at Kimi-class quality and a comparable price, with no access agreement attached, and buyers still choose Chinese models, then the preference is political and exclusion is the honest response. I do not think that is what happens, but it is a testable claim and I will report the result.

Show me the incident. A checkpoint carrying a trigger, an exfiltrating fine-tune, a hosted endpoint with a kill switch someone actually found — any documented case converts an architectural worry into evidence, and evidence changes the arithmetic. The risk so far is legal and structural, not demonstrated.

Publish the delta. If Cursor names the replacement model and the cost difference is small, then geopolitics really was the price and I was wrong about the motive. "Exploring" is not that disclosure, which is exactly why it should not be treated as compliance.

Until one of those lands, the ordering matters more than the intent. Build the open-weight alternative first, and the Chinese models lose on merit. Ban them first, and the only thing that gets cheaper is the argument for buying American.

Is your team buying Chinese open weights for the price, or in spite of it? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

The Take — Chinese data isn't the risk; the last fine-tune is

The Take — Chinese data isn't the risk; the last fine-tune is

Nous Research took Alibaba's Qwen3.8-27B, put about fifty million tokens of supervision through it, and cut the share of answers carrying Chinese state propaganda from 22 percent to 0.9 percent without losing measurable capability. Most coverage read that as a story about Chinese censorship. I think it is a story about where a model's politics actually live — and it is not where the people writing export rules think it is. As our brief on the release laid out, Arete-27B is supervised…

The Take — Stop blaming open weights for the patch gap

The Take — Stop blaming open weights for the patch gap

Six consecutive Firefox releases fell to a downloadable model at roughly $1,000 a version, and the reflex will be to reopen the open-weights liability argument. That reflex is wrong. The vulnerability that made this work was not the model — it was a release calendar that hands attackers a machine-readable map of every flaw before it hands users a fix. The experiment is Epoch AI's, and it is clean enough that the argument can be had on evidence rather than vibes.