Nobody hid this defection. A startup CEO announced it on X, a Fortune 500 exec called it “fast and cheap” on the record, and Congress is already asking two of the highest-profile adopters to explain themselves in writing.
- 01Chinese open-weight models now carry a growing share of U.S. enterprise AI traffic. OpenRouter data reported by CNBC put it above 45% as of July 7 — up from 11.5% at the start of the year.
- 02Congress is already investigating two of the biggest adopters. The House Select Committee on China and the House Homeland Security Committee wrote to Airbnb and Cursor’s parent Anysphere in April over their use of Alibaba’s Qwen and Moonshot AI’s Kimi.
- 03The economics are the whole story.Open-weight Chinese models reportedly run 60–90% cheaper than frontier U.S. models while landing within roughly a point of them on agentic benchmarks.
- 04At least one company has gone all the way.Lindy’s CEO says he moved 100% of the startup’s Anthropic traffic to DeepSeek V4 in a single cutover and saved “millions” while performance went up, not down.
A flag of convenience is a shipowner’s oldest trick: register the hull under a foreign flag with looser rules and a smaller bill, and the cargo doesn’t care whose ensign is snapping over the stern. Nobody in this story is hiding a registry the way that trick implies — the companies switching to Chinese models are saying so in public, sometimes proudly. What’s new is how fast the fleet is doing it, and how openly Washington has started asking why.
01 The reflag
The number doing the most work here is the change, not the level. OpenRouter — the API gateway that brokers traffic across dozens of model providers for thousands of companies — reported that Chinese labs’ share of the U.S. enterprise tokens crossing its platform sat at 11.5% on January 1. By July 7 it had crossed 45%. That’s not a drift; a share can’t roughly quadruple in six months without a structural reason, and the reason on offer is unusually blunt: price.
That last figure is the tell. A model that goes from under 1% to 17% of token volume in a single month, while its share of actual dollars spent barely moves, is being adopted for exactly one reason: it’s cheap enough that using ten times the tokens still costs less than the alternative.
02 The manifest
Four names keep coming up in this story, and each one switched a different way — a full cutover, a fleet-wide quiet migration, and two companies built the dependency directly into a product they shipped.
| Vessel | New flag | Cargo | Reported toll |
|---|---|---|---|
| Lindy | DeepSeek V4 — off Anthropic entirely | AI agent workflows | “Saves us millions of $… increase in performance,” per CEO Flo Crivello |
| Coinbase | Chinese open-weight models, ~1,200 agents | Internal automation | Cut AI spend in half, per reporting |
| Airbnb | Alibaba’s Qwen | Customer-service agent | “Fast and cheap,” per CEO Brian Chesky |
| Anysphere (Cursor) | Moonshot AI’s Kimi, inside its own Composer 2 | Code-completion model | Disclosed only after Composer 2 had already shipped |
03Two ships’ logs
Nobody disputes that the workloads moved, or that price is why. What the two sides disagree about is whether the savings are worth what rides along with them.
Cheaper, and not obviously worse
- Lindy, a 25-person startup, moved off Anthropic entirely in a single cutover after its AI bill grew larger than its payroll.
- Coinbase runs roughly 1,200 AI agents on Chinese open-weight models and says it has cut its AI spend in half.
- Vercel’s AI Gateway saw DeepSeek’s share of token volume jump from under 1% to 17% in a single month, while its share of actual spend stayed near 1%.
The savings are the risk
- The House Select Committee on China and House Homeland Security Committee jointly wrote to Airbnb and Anysphere in April, demanding an accounting of their model choices and any communications with Chinese providers.
- Airbnb built a customer-service agent on Qwen and called it fast and cheap; Anysphere’s Composer 2 turned out to be built on Kimi, disclosed only after it had already shipped.
- Rep. Andy Ogles, who chairs the cybersecurity subcommittee involved, has called for a “serious strategy” to keep U.S. models a “real alternative.”
Nobody disputes that the workloads moved, or that they moved because of price. The entire fight is over whether a legal regime that can compel a lab to cooperate with state authorities is a cost line or a different kind of dependency altogether.
What both sides agree on, once you subtract the framingHow the flag changed
The baseline
Chinese open-weight models carry 11.5% of the U.S. enterprise token volume moving through OpenRouter — roughly where the share had sat, on average, the year before.
ORIGINThe first letters go out
Chairmen Moolenaar and Garbarino jointly write to Airbnb and Anysphere, demanding an accounting of their use of Qwen and Kimi and requesting an in-person briefing.
One month, under 1% to 17%
DeepSeek’s share of Vercel AI Gateway token volume spikes from under 1% to 17% while its share of spend barely moves — the cost math working in real time.
Efficiency replaces tokenmaxxing
Reporting describes a broader shift across enterprise AI budgets, from scaling spend first to auditing cost per task.
The number crosses 45%
OpenRouter data show Chinese models’ share of U.S. enterprise tokens above 45% — roughly four times the January baseline.
Lawmakers go public
CNBC’s follow-up widens the story beyond the original two companies, naming the State Department’s concerns and Rep. Ogles’s call for a strategy on the record.
No hearing date yet
Neither Airbnb nor Anysphere has published a response to the committees’ letters — the date this piece’s reporting is current to.
TODAY05 My read
The 45% figure is being reported as a security scandal, but the more interesting number in this story is the first one: 11.5%, six months earlier. That’s not a slow drift — that’s a stampede, and stampedes happen when price gets structurally decoupled from apparent quality. If Coinbase’s internal agents and Cursor’s own model are landing within a point of frontier benchmarks at a tenth of the cost, the real news isn’t that companies moved. It’s that Anthropic and OpenAI priced as if nobody would notice how thin the performance gap had gotten.
But the security case doesn’t evaporate just because the economics are real. A model trained under a legal regime that can compel a lab to cooperate with state authorities is a different category of dependency than a model that’s merely cheap and thinly supported — and “designed to… reflect CCP ideology and values” is a specific, checkable claim, not vague protectionism, even from a committee with an obvious institutional interest in finding one. The uncomfortable truth for both camps: the buyers are right that the frontier labs got complacent on price, and Washington is right that price isn’t the only variable that matters. Neither side’s press release mentions the other’s part.
What to watch
Does either company answer the letters?
Watch whether Airbnb or Anysphere issue a public response addressing model choice and any communications with Chinese providers specifically, rather than a general statement about vendor diversity.
Does the 45% figure keep climbing, or plateau?
OpenRouter’s weekly tracker is the cleanest public signal available. A plateau would suggest the early movers have already moved; a continued climb would suggest this is still accelerating.
Does a frontier lab cut list price early?
The Sonnet 5 pricing window this site has already logged twice closes August 31 on schedule. An unforced price cut before that date, rather than waiting out the scheduled increase, would be the clearest tell that competitive pressure is now setting the roadmap.
Does the investigation produce a proposal, or just a hearing?
Letters and briefings are step one of a process that usually ends in either legislation or nothing. Watch whether a specific open-weight strategy follows Rep. Ogles’s public call for one.
- Sourcing here is aggregated news reporting,not a firsthand read of OpenRouter’s raw dashboards or the committees’ original letters — quotes and figures are relayed through outlets including CNBC and PYMNTS, and through Flo Crivello’s own public posts, not independently re-verified against primary sources.
- The token-share percentage varies slightly by source and by week — different outlets report 45% and 46% for the same rough period. Treat the figure as directional, not a single precise number.
- “Chinese AI models” here means open-weight models built by Chinese labs(DeepSeek, Alibaba’s Qwen, Moonshot AI’s Kimi, Zhipu’s GLM) run through third-party gateways — not a claim about who operates the servers the weights run on.
- Airbnb’s and Anysphere’s own accounts are drawn from prior public statements— Chesky’s comment and Anysphere’s later disclosure about Composer 2 — not a direct response to the committees’ letters, which hadn’t been made public as of this writing.
- This is a fast-moving story; status is current as of July 14, 2026.