Core theses · Updated 2026-09-25

US and China in AI: one cycle, two positions

01The claim

The US and China sit at two positions in one AI cycle: the US late in its mania phase, China early to mid. The gap is about 14x in capital, about 2.7% in model capability and about one third in valuation. Founders who bill in dollars are closing that discount themselves.

02Mechanism: why it happens

Concentration holds only in the US, because the pricing anchors sit in different places. Late in a mania phase capital pays only for revenue that is already proven, so the river narrows: US venture invested $412.7 billion in the first half of 2026, 86% of it in AI, and OpenAI plus Anthropic took 43% of global venture. China is early to mid, with few revenue anchors that can be priced, so money spreads across more bets: 64.6% sat in early rounds. Two positions give capital opposite shapes, and the two sides cannot run on one clock.

Cognition: on September 8, 2026 it announced a $2 billion Series E at a $48 billion post-money valuation, double the $1 billion Bloomberg had reported on September 2; annualized revenue rose from $492 million in May to nearly $900 million (confirmed, Bloomberg and Reuters).

The exit window opened on the Chinese side first because Hong Kong produced a public pricing anchor first. Zhipu and MiniMax listed in Hong Kong in January 2026, and Zhipu's intraday peak reached 25.6 times its issue price; 155 Chinese companies went public in the first half, raising RMB 233.985 billion, 82 of them in Hong Kong. No AI company listed in the US in Q1 2026, and the anchor waits on Anthropic, whose roadshow reportedly starts in mid-October at the earliest. Until then the US exits through acquisitions: incumbents cannot buy time, so they buy positions.

NVIDIA's acquisition of Hugging Face (FutureX portfolio): announced September 3, 2026 at about $12.93 billion, expected to close in the first half of 2027; per company disclosure the platform has 18 million developers and 3 million models (confirmed, SEC 8-K).

Control of the cost curve changed hands, and open source is why. A closed model's moat is capital expenditure, which must be refinanced to survive and is weakest at the top of a cycle; an open model's moat is a living community, which needs no refinancing. DeepSeek trained a GPT-4-class model for $5.6 million, and Chinese models held 58% of OpenRouter token share in July 2026. An application's gross margin is set by the cheapest model that is good enough, so each step down the cost curve moves value from the model layer to the application layer.

Zhipu's GLM-5.3-Flash: open-sourced on August 26, 2026, 320B total and 18B active parameters, priced at $0.15 per million input tokens and $0.50 per million output, about 1/40 of Claude Opus 4.8, scoring 57 on the Artificial Analysis index (confirmed, Zhipu release).

The discount closes from inside the company. Chinese teams built 44% of the global top 50 AI apps yet trade at about a third of US comparables. The discount is a buyer's price for exit and regulatory risk, and it attaches to the currency of revenue and the address of the customer. Global AI companies put Chinese engineering density and global pricing inside one company: dollar revenue, more than half from overseas, teams on both sides of the Pacific, leaving buyers only overseas comparables. They do not wait for anyone to correct the discount; their revenue line is the correction.

Genspark and Dify (FutureX portfolio): per company disclosure, Genspark reached $200 million ARR eleven months after launch; Dify went from 30,000 GitHub stars and zero revenue to $10 million ARR in 18 months, with 57% of revenue from overseas (per company disclosure, as of August 2026).

Europe is becoming a third pool of capital, and what it buys is sovereignty. Governments, industrial investors and an EU-backed fund entered one open-weight model company in the same round, with the money going to owned data centers in France and Sweden; the sovereign model moved from policy narrative to capital fact. The two-column US-China table therefore gains a third column: open weights plus sovereign capital. Yet €3 billion is about $3.6 billion, a fraction of the tens of billions US frontier labs raise in one round, and it does not rewrite the cycle.

Mistral (FutureX portfolio): on September 8, 2026 it closed a €3 billion Series D led by Samsung Electronics at a post-money valuation above €21 billion, the largest equity round by a European technology company; per company disclosure it expects ARR of €1 billion by the end of 2026 (confirmed, TechCrunch, CNBC and Euronews).

03Evidence

01AI venture investment in 2025: $194 billion in the US against $14 billion in China, about 14x. (2025 · 已证实 · PitchBook / NVCA / 中国信通院;图版《开源破局》第 18、26 页)
02In Q1 2026 three US companies took 67.3% of global AI funding, while 64.6% of China's funding sat in early rounds. In the first half, US venture reached $412.7 billion, 86% of it in AI; OpenAI ($122 billion) and Anthropic ($95.6 billion) together took 43% of the $510 billion in global venture. (2026-06-30 · 已证实(公开统计) · 图版第 12、18 页;报告 us-china-ai-primary-capital-2026-h1 摘要与关键发现(PitchBook / Crunchbase))
03Exits run in opposite directions: no AI company listed in the US in Q1 2026; 155 Chinese companies went public in the first half, raising RMB 233.985 billion, 82 of them in Hong Kong. Zhipu and MiniMax listed in Hong Kong in January 2026; intraday peaks versus issue price: Zhipu 25.6x, MetaX 9.9x, Moore Threads 8.2x, MiniMax 8.1x. (2026-06-30 · 公开统计 / 公开市场行情 · 图版第 9、16 页(Wind,截至 2026 年上半年))
04The capability gap between the top US and Chinese models is about 2.7% (LM Arena, March 2026); Chinese models reached a record 58% of OpenRouter token share in July 2026, up from an earlier reading of more than 45%. (2026-03 / 2026-07 · 已证实 · 图版第 11、18 页;报告 ai-compute-token-economics-2026 关键发现)
05Moonshot AI confidentially filed an A1 application with the Hong Kong Stock Exchange in early September 2026 while advancing a pre-IPO round at roughly $50 billion pre-money, above its July Series F post-money of $35 billion; it is the third Chinese model company in the Hong Kong queue after Zhipu and MiniMax. The company declined to comment. (2026-09-02 · 据报道 · 晚点 LatePost 与 IT之家 2026-09-02、凤凰网 2026-09-03;positions-and-updates.md [china-ai-going-global-2026]、[frontier-models-ai-sovereignty-2026] 9 月上旬更新)
06Mistral closed a €3 billion Series D at a post-money valuation above €21 billion (about $24 billion), led by Samsung Electronics with EQT's Scaleup Europe Fund and PSG as co-leads and the Grand Duchy of Luxembourg as a new investor; the company calls it the largest equity round in European tech history, with proceeds going to data centers in France and Sweden. (2026-09-08 · 已证实 · TechCrunch、CNBC、Euronews 2026-09-08;positions-and-updates.md [us-china-ai-primary-capital-2026-h1]、[frontier-models-ai-sovereignty-2026] 9 月上旬更新)
07Nscale filed an S-1 with the SEC on September 18, 2026: a GPU services agreement with Anthropic signed August 25 worth up to about $44.6 billion; H1 2026 revenue of $140.6 million and a net loss of $1,020.1 million; $103.4 billion of active and contracted TCV as of August 31; NVIDIA subscribed to $2.1 billion of unsecured convertible notes on September 15 and committed a further $1.0 billion. Debt and circular financing on the US side sit in one prospectus. (2026-09-18 · 已证实(S-1 原文) · https://www.sec.gov/Archives/edgar/data/0002110365/000119312526395475/ck0002110365-20260918.htm(2026-09-18))

04The strongest counterargument

Tony Peng, Recode China AI, May 11, 2026

From 2020 to 2025 Chinese AI startups faced three structural exit barriers: Nasdaq closed behind geopolitics and a regulatory risk premium, domestic anti-monopoly rules kept big tech from acquiring startups, and VCs pulled back for lack of exits. Only after Zhipu and MiniMax listed in Hong Kong in January 2026, raising a combined $1.2 billion and trading at multiples of that, did later rounds unlock. His evidence shows the discount is priced by the exit channel; taken to its end, what the channel gives, the channel can take back, and founders cannot change that layer.

Our reply: We accept his diagnosis of 2020 to 2025; 'the channel can take it back' is our extension of his logic, not his words. The exit component of the discount is indeed set by the channel, and whether Hong Kong keeps absorbing listings is one of our falsification tests. Our claim is narrower than his: for companies with dollar revenue and more than half of it overseas, less of the discount attaches to the passport and more to the revenue line. That part founders can change; whether they can close it to the same revenue multiple as US peers in the same sector, we do not yet have sample-level evidence.

NDRC's block of the Manus acquisition (April 2026), as read by Morgan Lewis on May 6, 2026

In April 2026 China's NDRC used the Measures for the Security Review of Foreign Investment to block a major US technology company's roughly $2 billion acquisition of Manus, the first publicly confirmed use of that mechanism to unwind a cross-border AI deal. Even a company that has moved abroad and bills in dollars can have its exit vetoed. 'Founders fix the discount' ignores that vote.

Our reply: This is the hardest single fact against our view, and we do not soften it. It raised the cost of one exit route: a team with Chinese-origin technology selling to a US strategic buyer. Our reading is that value then lands more on the revenue line and the Hong Kong route; Manus returned to independence in August 2026 with ARR above $100 million (reported). We added 'which regulator can veto your exit' to our diligence list; a second block of this kind before June 2027 would raise our estimate of the structural share of the discount.

Brad Gerstner, founder of Altimeter Capital, August 8, 2026 (BG2 podcast)

Field data runs opposite to the story that open models hollow out the frontier: revenue is concentrating at the leading labs, and the intelligence gap may widen over the next two to three years. He has tried Kimi, Qwen and GLM and finds them strong, yet they grow token volume while the economics flow to frontier labs, which pay three to five times market rates to lock up compute. If the gap widens, China's cost-curve advantage is a low-end position and the discount is deserved.

Our reply: He counts profit, we count load; both are right in different units. In September GPT-6 Astra and Claude Fable 5.1 anchored their flagships at $10 and $50 per million tokens and raised prices together, so the two frontier labs still hold pricing power, and we concede that; if the gap widens as he expects, agent workloads that need the strongest reasoning chain stay at the frontier, and we do not dispute that either. Our judgment sits one layer down: an application's gross margin is set by the cheapest model that is good enough, so once comparable capability costs 1/40 as much, batch, offline and cost-sensitive workloads migrate, and frontier-lab profit and application-layer margin can widen at the same time. FutureX Capital does not bet on which lab wins; it holds the chokepoints every model must pass through.

Joint statement by three US security agencies, September 8, 2026 (reported by Inside AI News, September 9)

The statement accuses Chinese AI companies of 'aggressive and systematic' industrial-scale distillation of US frontier models. US market access and dollar customers are exactly what 'closing the discount' depends on; if the accusation turns into sanctions, what gets cut is the revenue line of Chinese-founded companies.

Our reply: The statement named no company, estimated no loss, announced no sanction or prosecution and set no timetable, so our rating of the policy risk stays at 'more statement than enforcement'. It is a live variable and sits in our falsification tests: if the accusation becomes a named sanction or an entity listing, the judgment changes. On July 24 more than twenty firms including a16z, Microsoft, Meta and NVIDIA signed a letter opposing broad limits on open-weight models, so the US itself has not settled the question.

05What would change our mind

▸If by June 2027 Chinese-founded companies with dollar revenue and more than half of it overseas, measured on publicly reported latest-round valuation and annualized revenue, still raise at revenue multiples below half of their US peers in the same sector, 'founders close the discount' fails, and the structural share of the discount is larger than we estimate.
▸If Hong Kong AI IPOs from the second half of 2026 through the first half of 2027 fall to half the first-half-2026 count, or Moonshot fails to list within 12 months of filing, our judgment that 'the exit window opened in China first' must be rewritten.
▸If before June 2027 the US distillation accusation becomes a named sanction or entity listing, or China blocks a second cross-border AI deal under security review, the wall between the two positions is higher than we drew it, and the 'one cycle' premise must be reassessed.
▸If Chinese models' share of OpenRouter tokens stays below 30% for two consecutive quarters (58% in July 2026), or Meta's open-weight flagship holds the top of the Hugging Face download chart for six months after release, 'China sets the cost curve' must be rewritten.

06What we do, and what we do not do

▸Two clocks for two sides. On the US side we subtract: only application companies with real ARR and chokepoints that span layers. On the Chinese side we add: positions in early rounds, weighted to agents and AI hardware. A seven-signal framework has scored daily since March 2026; when the 30-year Treasury yield touched 5.31% on August 17 and crossed the liquidity threshold, the standing protocol accelerated selling (/answers/us-china-ai-divergence-2026).
▸We back only teams that put both cycles inside one company: dollar revenue, more than half from overseas, people on both sides of the Pacific. Two diligence questions were added: which regulator can veto your exit, and how your retention curve behaves when an upstream model is cut off. Ten questions for founders are at /founders.
▸Own the hubs, do not pick the horse. We hold the chokepoints every winning model must pass through, Hugging Face, Mistral, Dify, OSChina / Gitee; we do not bet on a single Chinese or US lab and stay out of the model-layer funding elimination round.
▸We do not touch debt-financed compute leasing, and we skip the pre-IPO segment priced at 50 to 100 times sales. In humanoid robotics we back only dexterous hands and data, and stay out of the full-body arms race.
▸We track Europe as a third pool and look only at companies that pair open weights with owned compute; a sovereign round is still a fraction of a US one, so we do not price European companies off US valuations.
▸Eight offices, headquartered in Hong Kong, and a 60+ industry-founder network connect the two sides: overseas customers and channels for Chinese teams, Chinese supply chains for US teams. FutureX Capital, founded in Asia in 2018, sits inside a network of 140+ companies, 80+ of them AI-native, and has reviewed 1,500+ projects; 24 research reports and 22 AI agents stand behind the judgments above, each update dated and graded (/research, /debates).

07Open questions

?How much of Hong Kong's scarcity premium for Chinese model companies survives the fourth and fifth listing? Zhipu's 25.6x was scarcity pricing, and the queue behind Moonshot will lengthen; we do not yet have an answer.
?If Anthropic lists in November, the primary market spends two months without a public pricing anchor, with valuations confirmed only by each new round; will they retreat before the public market does?
?Will Europe's sovereign pool grow into an independent third cycle, or rise and fall with the US one? Mistral's €3 billion is one sample, and a sample does not make a curve.
?How much does 'founders close the discount' depend on US market access? If access tightens, can dollar revenue from Europe, the Middle East and Southeast Asia replace US customers? We have not yet broken portfolio overseas revenue down by country finely enough to answer.

08Changelog

2026-09-25First published.