Nobody Trusts China's AI. Everyone Is Using It Anyway.

East/West AIpha · Edition 3 · Tuesday, August 4, 2026 · Day 1002

Automate Work. Amplify Life. East/West AI, for the people betting real capital and real careers on it.

At least five governments have restricted DeepSeek: the US (Navy and Senate), Taiwan, Australia and South Korea on official devices, Italy nationwide (Tech.co roundup: link). The US Congress is investigating DoorDash, Airbnb, and Cursor for quietly running Chinese open-weight models. Trust polling has China splitting opinion, drawing both strong trust and strong distrust, and failing to top the rankings even among people who think China is winning (Public First via TNW, Jun 19, 2026). (HIGH)

And this week, Chinese open-weight models hold every one of OpenRouter's top five spots, the second such sweep in about three weeks, after Fortune reported an all-Chinese top five in mid-July (Global Times, Aug 3, 2026: link). Chinese models' share of tokens run by US firms on OpenRouter is running 57 to 63% (Fortune, Jul 26; Bloomberg via AI Weekly, Jul 22). In the first half of 2025 it was 4.5% (CNBC, Jul 7). (HIGH)

Not a contradiction. A pattern. Chinese financial media has a name for it that predates this news cycle: 弯道超车 (wāndào chāochē), "overtaking on the curve," a Chinese business-strategy metaphor, popularly associated with Baidu founder Robin Li, for beating a stronger rival at the bend instead of the straightaway. Hugging Face's own CEO said it more bluntly than any Chinese state outlet has: "They're clearly dominating on open models right now" (Clément Delangue, CNBC, Aug 3, 2026: link).

Every number below is sourced, Chinese-language press included. The rule: adoption does not track trust. It tracks whether a model can be downloaded and run on a machine you already control. Weights versus wrapper. 后来居上. Same pattern, two languages.

Every outlet reporting Chinese AI's OpenRouter share since February measured something slightly different, and the resulting range, 46% to 63% across five months, is real, not sloppy. Note that the February figure is global share, not US-firm share. (HIGH)

Cailian Press, republished by Sina Finance, carried the same demand-side data to Chinese readers that week: "中国大模型在美'悄然崛起'" ("Chinese large models are quietly rising in the US"), quoting OpenRouter's Justin Summerville that Chinese open models run 60 to 90% cheaper than leading Anthropic or OpenAI models, and reporting that Lindy founder Flo Crivello's switch to DeepSeek cut inference cost to a tenth of Anthropic's Sonnet (Cailian Press via Sina Finance, Jul 7, 2026: link).

Treat 57 to 63% as the working range, not a fixed number. The trend is the real story: US models held roughly 70% of all OpenRouter tokens a year ago; Chinese models first passed US models on US-firm traffic in the week of February 9 to 15, 2026 (CNBC, Jul 7), and reached 57 to 63% by mid-2026 (Bloomberg via AI Weekly, Jul 22, 2026: link). (HIGH, directional) The second number that matters more: Anthropic captured 61% of spend on Vercel's Gateway in June while carrying only 32% of tokens; DeepSeek alone ran 22.6% of token volume (Vercel via Computing, Jul 15, 2026: link). Chinese models win the pipes. US labs still win the invoice. (HIGH)

US federal bodies including the Navy and the Senate, plus Taiwan, Italy, South Korea's trade ministry, and Australia, all restricted DeepSeek in early 2025 (Tech.co: link); New York State banned it on state government devices in February 2025, the first US state to do so. A bipartisan bill, the No DeepSeek on Government Devices Act (H.R. 1121), was introduced in February 2025 and has not yet passed either chamber (congress.gov). NIST's CAISI evaluation found agents built on DeepSeek's most secure model, R1-0528, were on average 12 times more likely than US frontier models to follow malicious hijacking instructions, and that the same model answered 94% of overtly malicious requests under a common jailbreak, versus 8% for US reference models (NIST/CAISI, Sept 30, 2025: link). The same report notes downloads rose nearly 1,000% since January 2025. The risk assessment and the adoption curve came from the same document. (HIGH)

The trust data does not match the usage data. A Public First poll of 18,000-plus people across 15 countries found that in 11 of the 15 countries polled a clear majority believe China leads on AI capability, yet China does not top trust rankings (Public First via TNW, Jun 19, 2026: link). TNW's read of the same poll: most people who said they distrusted US models were still using ChatGPT, Claude, or Gemini anyway. Run that the other direction and you get this edition's thesis. (HIGH)

Named and dated: two US House committee chairmen sent DoorDash a formal information request on July 31 over its use of Moonshot's Kimi K2.6 (SCMP, Aug 1: link); Kuai Technology's headline, republished by Sina Tech, was blunter: "给特朗普送过外卖也不管用！...遭国会调查" ("Even having delivered for Trump doesn't help, now under Congressional investigation") (Aug 1: link). Earlier probes hit Airbnb (Qwen for customer service) and Cursor's Anysphere, which has said Moonshot's Kimi provided the foundation for its Composer 2 coding model (Fortune, Jul 26: link).

A government can ban a company or block an app. It cannot un-download weights already copied onto servers it does not control. Mozilla's State of Open Source AI report states it plainly: jurisdictions ban the hosted app, and "enterprises adopt the weights anyway, self-hosted or via Western endpoints" (Mozilla/SlashData, Jul 2026: link). (HIGH)

Moonshot founder Yang Zhilin (杨植麟) said it directly at March's Zhongguancun Forum: "开源模型正逐渐成为新的标准" ("Open-source models are gradually becoming the new standard") (The Beijing News, Mar 25, 2026: link).

Proof of the mechanism in both directions. Perplexity took DeepSeek-R1's open weights, stripped the model's government-alignment filtering, and released R1 1776 hosted entirely on US and European servers (Perplexity, Feb 2025): capability transferred, jurisdiction did not (MODERATE-HIGH). In June 2026, a US export-control order took Anthropic's own Fable 5 offline globally from June 12 to July 1, because Anthropic had no way to verify user nationality in real time (Forbes, Jun 16: link; CNBC, Jun 30: link): you cannot pull the plug on a file already sitting on ten thousand machines you do not own. Distillation risk runs both ways too: Anthropic alleged in February 2026 that DeepSeek, Moonshot, and MiniMax collectively created roughly 24,000 fraudulent accounts generating over 16 million exchanges (VentureBeat, Feb 24, 2026: link). Self-hosting solves where your data goes. It does not resolve where the model's training data came from. (MODERATE)

The fastest Chinese-model adoption is not happening in the developing world. It is happening in Silicon Valley. A16z's own policy team claims roughly 80% of developers building on open-source tools chose Chinese ones (a16z, Apr 28, 2026: link); treat that figure with some caution, a16z partner Martin Casado has separately said he was misquoted on a similar stat, clarifying it described a narrower base than "80% of developers" broadly. (MODERATE)

Menlo Ventures' formal survey of 495 US enterprise decision-makers found Chinese models at just 1% of total enterprise LLM API usage, roughly 10% of enterprise open-source usage (Menlo Ventures, Dec 9, 2025: link). (HIGH) Both numbers can be true at once: what enterprises report buying stays Western; the product layer, where an engineer routes a workload because it is ten times cheaper, is already majority Chinese. Alibaba CEO Eddie Wu (吴泳铭) put it plainly: "我们坚定的选择开源" ("We firmly choose open source"), to support the developer ecosystem (Sina Tech, Sept 24, 2025: link). DeepSeek's own app dethroned ChatGPT as the #1 free app on the US Apple App Store on January 27, 2025, days after the R1 release (CNBC, Jan 27, 2025: link).

This did not happen in one week. Ten dated milestones got us here, from Zhipu and MiniMax's Hong Kong IPOs in January to Alibaba's Qwen3.8-Max this week (full list in the timeline graphic below). A steady monthly cadence, each event compounding on the last.

Allocators: the token number is noisy, revenue still favors US labs, do not short US AI infrastructure off one chart. Do reprice the assumption that enterprise software has a durable model-cost moat.

Executives: if you are quietly running Chinese open models to cut costs, assume disclosure becomes the question before cost savings do. Self-hosting the weights survives a compliance review. Routing to a Chinese-hosted API gets you a letter from the US Congress.

Builders: if you cannot say which model runs behind your product, and where its weights execute, you do not own your stack. You own a wrapper someone else can switch off.

Also viral this week: "90后清华天才，干崩了美股" ("Post-90s Tsinghua genius crashed US stocks") hit #3 on Weibo's tech chart at Kimi K3's launch, 32 million views in 24 hours (Dealroom.co, citing Business Insider: link). Marc Andreessen's "Deepseek R1 is AI's Sputnik moment" (Jan 2025) is still the framing this story keeps getting hung on.

If your company runs Chinese open-weight models in production, disclosed or not, reply and tell us what and at what scale. We will share what's working for others.

Subscribe to East/West AIpha for the weekly signal, one read from the West and one from the East, and forward this to one operator who thinks this is a China problem instead of a procurement problem.

Are you running an open-weight model in production, and does your compliance team know which one? Reply and tell me. I read every one.

San Eng is the WSJ (#4) and USA Today best-selling author of Ten Commandments of Investing and an institutional investor with $500M-plus deployed. This is East/West AIpha's whole reason for being: research you can trust, from an investor who actually builds.

Subscribe free at workoptional.ai. Edition 4 lands next Tuesday.

Disclaimer: East/West AIpha is for informational and educational purposes only and does not constitute financial, investment, legal, or tax advice. Nothing in this newsletter is a recommendation to buy or sell any security. The views expressed are our own, may be incomplete, and can change without notice. Do your own research and consult a licensed professional before making investment decisions. WorkOptional.ai, its writers, and affiliates may hold positions in companies or assets mentioned. Past performance is not indicative of future results.