China AI Going Global: 23 Weeks Ahead on Usage, Not Yet on Pricing Power
FutureX Research · AI Lab · 2026.07.06 · 24 pp · preview 6 pp · Revised 2026.11.02
Listen · Audio Summary
5-8 min · AI narration in English · abstract + all key findings
Bottom Line
Chinese models have won global distribution, but pricing power must still be earned in dollar revenue; model-layer monetization is only halfway there.
- Why now
- On October 7, 2026, Anthropic released Claude Haiku 5.5 at one-tenth of Haiku 4.5's list price, narrowing the price gap behind China's cost-focused Flash models.
- What it means
- For founders and investors, the yardstick for Chinese AI abroad is shifting from downloads and token usage to dollar revenue and overseas cloud revenue share.
Abstract
Chinese models have won global distribution; pricing power still has to be earned in dollar revenue. By National Business Daily's count, Chinese models led U.S. models in weekly OpenRouter usage for a 23rd straight week in the week to October 4. Zhipu's GLM 5.3 went live on Amazon Bedrock on October 5, yet Zhipu's first-half revenue came in about 30% below analyst estimates. On October 7, Anthropic released Claude Haiku 5.5 at one-tenth of Haiku 4.5's list price. DeepSeek's valuation talks reached about $71B, and Bloomberg says Anthropic could list as soon as mid-November. (Data as of 2026-10-11)
Key Points
- 01In the week of September 28 to October 4, 2026, Chinese models processed 57.46 trillion tokens on OpenRouter versus 16.2 trillion for U.S. models, a 23rd straight week in the lead (National Business Daily estimate; Reported).
- 02Per reports on August 31, 2026, Zhipu's first-half revenue was RMB 954 million, up about 400% yet about 30% below analyst estimates; an August 15 report showed Alibaba's Qwen had topped 3 billion cumulative downloads (Verified).
- 03On July 27, 2026, Moonshot AI released the full weights of Kimi K3 (2.8T parameters, ~1.56TB), which topped Hugging Face's trending chart within 30 minutes (Reported).
- 04DeepSeek's valuation rose from about $52B post-money in its June 2026 first round to about $71B pre-money in July talks; Bloomberg reported on July 14 that it had begun onshore IPO preparations (Reported).
- 05On October 7, 2026, Anthropic released Claude Haiku 5.5 at one-tenth of Haiku 4.5's list price (Verified); the next day MiniMax fell 13.74% in Hong Kong while the Hang Seng Tech Index fell 2.89% (Public market data).
I. Kimi K3 Opens Its 2.8T-Parameter Weights; Open Source Is Still the Main Channel Abroad
Kimi K3 (a 2.8T-parameter MoE with a 1M-token context) was unveiled on July 16 and delivered on its open-weights promise late on July 27. The full weights (modified MIT license, ~1.56TB) went up on Hugging Face, alongside three open-sourced training-infrastructure components: MoonEP, FlashKDA and AgentEnv. Within 30 minutes K3 topped the trending chart with more than 4,000 likes, the fastest climb in the platform's history (Reported).
Support arrived on day one. Huawei Ascend, Alibaba Cloud supernodes, Nebius, Baseten and Fireworks went live the same day, and Cursor and Devin announced integrations.
Performance and pricing are converging on closed models. K3 topped the frontend-coding blind test at 1679 Elo, above Claude Fable 5's 1631, though it still trails the two closed-source flagships overall. API pricing of $3 per million input tokens and $15 per million output tokens ends the 'China discount' (Reported).
The race keeps tightening. On July 19 at WAIC, Alibaba previewed the 2.4T-parameter Qwen3.8-Max, claiming it is second only to Fable 5 without publishing benchmarks; its open-weights pledge was still undelivered as of Aug 1. On July 24, Jensen Huang amplified an a16z-initiated open letter arguing that frontier open and closed models should coexist (Reported). Open source is now a U.S. policy question, not only a Chinese strategy.
Small models still account for most downloads. The two-tier structure holds: flagships set the benchmark, small models drive adoption.
II. U.S. Rivals Now Compete in Courtrooms, on Store Shelves and on Balance Sheets at Once
Legal (Verified): On July 10, Apple sued OpenAI, its hardware unit io Products and two former Apple employees in the Northern District of California, seeking damages and an injunction. The two are OpenAI hardware chief Tang Yew Tan and former engineer Chang Liu; OpenAI denies wrongdoing. Reportedly (FT), Apple has sent legal letters to dozens of former employees who left. The complaint's allegations are one-sided claims, and no substantive ruling had been issued by end-July, so treat the case as unresolved.
Hardware: On July 15, OpenAI released the $230 Codex Micro, a limited-run agent-control keypad built with Work Louder. It sold out in about 12 hours, shipped from July 24, and briefly resold on eBay at multiples of list price.
Capital: On March 31, OpenAI announced a $122B financing anchored by Amazon, Nvidia and SoftBank. It disclosed that enterprise revenue is above 40% of the mix and that Codex has more than 2M weekly users, up 5x in three months. Its IPO has reportedly slipped to 2027. Anthropic filed a confidential S-1 on June 1, and Bloomberg reported that bankers had begun investor meetings for a listing as soon as October. On October 1, Bloomberg moved the earliest listing date to mid-November; see Latest developments.
Models: On July 18, Musk said xAI is training the 2T-parameter Grok 4.6, aimed squarely at surpassing K3.
III. DeepSeek's Valuation Moves From ~$52B to ~$71B: Private Markets Are Pricing Narratives
DeepSeek became the pricing anchor of mid-to-late July. Per Bloomberg (July 14) and the WSJ (July 15), it has begun onshore IPO preparations: a Shanghai listing, a filing possibly this year and a 2027 debut, with accountants and banks already engaged. Its valuation moved from about $52B post-money after June's first round (about $7B raised) to about $71B pre-money (about RMB 480B) in July's talks. On July 26 it reportedly paused new fundraising to prioritize the IPO.
Moonshot AI is switching from a performance story to a capital-markets story. Per National Business Daily, it plans to open pre-IPO funding talks in August at a $50B pre-money target (the current round is about $31.5B pre-money). It has reportedly sent investors a listing proposal for a HKEX debut within six months. ARR reportedly crossed $300M in June, which still implies a P/ARR around 100x.
The two premiums have different sources. DeepSeek's carries technology influence and strategic scarcity; Moonshot's rests more on financial and performance logic. Markets are pricing narratives more aggressively, and the SSE's June guideline applying the STAR Market's fifth listing standard gives such narratives a domestic exit. Correction risk and institutional tailwinds are building at the same time.
Qwen Passes 3 Billion Downloads While Zhipu Misses Estimates by About 30%: Adoption Runs Ahead of Revenue
Open weights buy global reach while the application layer collects the cash, and that call largely holds, but adoption is running well ahead of revenue.
Verified (Fortune, Aug 15; China Daily, Aug 16): Hugging Face's state-of-open-models report shows Alibaba's Qwen family has topped 3 billion cumulative downloads, with over 300,000 derivative models. Downloads on the platform in 2026:
| Model family | 2026 downloads on Hugging Face |
|---|---|
| Alibaba Qwen | ~2.045 billion |
| 418 million | |
| Meta | 227 million |
Verified (Caijing, Sina Tech, Aug 19): Kuaishou's Q2 results put Kling AI's quarterly revenue above RMB 850 million, up more than 200% year on year, with over 100 million users across 224 countries and territories as of June.
Verified (Bloomberg, SCMP, Aug 31): Zhipu's first-half revenue was RMB 954 million, up about 400%, with API sales at 86.5% of the total. It still came in about 30% below analyst estimates, on a half-year loss of RMB 2.07 billion.
Chinese video apps are earning real money abroad, and Chinese open models own the download charts. The model layer has not yet turned that adoption into profit.
FutureX Position · Founders Billing in Dollars Are Closing China's Valuation Discount Themselves (Xiamen keynote, 2026-09-03)
This keynote supplies the mechanism behind the report's 'monetization is still halfway' call and narrows that halfway point to the model layer. Application-layer monetization did not stall. It took a different route, with Chinese founders billing in dollars abroad.
FutureX's read: the new wave of global AI companies is largely founded by Chinese entrepreneurs. What they share is Chinese engineering density and global pricing inside one company.
| Company | Valuation | ARR / revenue | Basis |
|---|---|---|---|
| Cognition | $26B | above $490M | Reported, as of August 2026 |
| Fireworks AI | $17.5B | above $1B | Reported, as of August 2026 |
| Manus | — | above $100M | Reported, as of August 2026 |
| Genspark | — | $200M ARR in 11 months | Per company disclosure |
| Dify | — | 57% of revenue overseas | Per company disclosure |
The backdrop is a mismatch. In 2025, AI venture investment was $194B in the US against $14B in China, a gap of roughly 14x (PitchBook/NVCA/CAICT). Chinese teams built 44% of the global Top 50 AI products, yet are valued at roughly one third of US comparables (a16z, August 2025). Our judgment: nobody needs to wait for that discount to be corrected. Companies that bill in dollars, with more than half their revenue overseas, are closing it themselves.
For this report: download charts measure distribution, and pricing power shows up on the revenue line. Chinese teams already hold the charts. Pricing power has to be earned in dollar revenue.
Full deck: /reports/open-source-breakthrough (first 5 pages public).
Latest developments, newest first (data current as of 2026-10-11)
- 2026-10-08 · In Hong Kong, MiniMax fell 13.74% to HK$207.2 and Zhipu fell 7.61% to HK$643, while the Hang Seng Tech Index dropped 2.89% (Tencent Securities daily data; International Finance News) · Public market data
- 2026-10-07 · Anthropic released Claude Haiku 5.5: for prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, one-tenth of Haiku 4.5's list price; the company says it costs about 75% less to run on average (Anthropic website) · Verified
- 2026-10-06 · Zhipu's Hong Kong shares closed at HK$715.5, up 7.59% · Public market data
- 2026-10-05 · AWS announced that Zhipu's GLM 5.3 is generally available on Amazon Bedrock for eligible enterprise customers through the US and Global cross-Region inference profiles; the model has 753 billion total parameters, about 40 billion active per token, and a one-million-token context window (AWS website). Red Star Capital Bureau reported that AWS shares revenue with Zhipu based on usage · Verified (revenue-share arrangement reported)
- 2026-10-05 · National Business Daily, calculating from OpenRouter data: in the week of September 28 to October 4, total model usage reached about 166 trillion tokens, up 13.69% week on week; Chinese models on the leaderboard processed 57.46 trillion versus 16.2 trillion for US models, a 23rd consecutive week of Chinese leadership. The anonymous Space Bunny Alpha took first place with 38.7 trillion tokens, up 179%; a tokenizer test suggests it may come from MiniMax, which has not been confirmed. DeepSeek V4.1 Flash ranked second with 25.6 trillion. Four of the top five were Chinese models, three of them speed- and cost-focused Flash models · Reported
- 2026-10-01 · Bloomberg (as relayed by Yahoo Finance): Anthropic aims to list as soon as mid-November, could begin marketing the offering the week of November 9, and is set to meet prospective investors on October 14 · Reported
- 2026-09-30 · DeepSeek open-sourced base components for Huawei's Ascend platform, including Ascend support for its TileLang compiler tooling plus compute and distributed-communication libraries; DeepSeek said they map one-to-one to its open-source components for Nvidia hardware, that compute and communication performance is close to the hardware ceiling in several key test cases, and that Huawei's team provided support (Wallstreetcn, Yahoo Finance and others) · Verified
- 2026-09-29 · A non-public draft of Anthropic's prospectus, seen by Reuters, shows at least about $518 billion in compute and infrastructure commitments over roughly a decade, around 80% of them non-cancellable or payable regardless of usage · Reported
- 2026-09-28 · Yilan Business and National Business Daily, compiling OpenRouter data: in the week to September 27, platform-wide token consumption was about 146 trillion, up roughly 13.2% week on week and a record. The top three were DeepSeek V4.1 Flash (19.6 trillion), Zhipu GLM 5.3 Flash (16.3 trillion) and the anonymous Space Bunny Alpha (13.9 trillion), listed on September 23; tokenizer fingerprint tests suggest the last may come from MiniMax, not officially confirmed. By National Business Daily's count, Chinese models have out-consumed US models for 22 consecutive weeks; by Yilan Business's count, Space Bunny alone took 12.2% of top-20 volume · Verified
- 2026-09-22 · The Information (relayed by Cryptopolitan): the Cyberspace Administration of China is investigating DeepSeek and Moonshot AI over allegations that user queries were sent to Claude without customers' knowledge, and has called in representatives of seven companies: DeepSeek, Moonshot, Alibaba, Zhipu, SenseTime, MiniMax and Xiaomi; the review focuses on cross-border data rules, and no penalty had been announced as of September 22 · Reported
- 2026-09-17 · Sina Tech (September 17) and Yicai (September 20), citing OpenRouter data: in the week of September 7-13, OpenRouter handled 127 trillion tokens, up 10.4% week on week; Chinese models accounted for 61.2 trillion, up 7.9%, ahead of US models at 21.8 trillion for the 20th consecutive week, while US models grew faster at 31.6%. Chinese models held at least seven of the top ten slots with a combined share above 70%, and DeepSeek V4.1 Flash, released September 10, ranked sixth in its first week. On another Yicai measure, Chinese vendors' overall share rose from under 20% a year earlier to more than 45% · Reported
- 2026-09-16 · MiniMax Agent, Hailuo AI and MiniMax Audio were added to the SWDA national AI skills program run by the Singapore government and Singtel (Sina Tech, September 20) · Reported
- 2026-09-11 · MiniMax and Genspark released Gen-1 Slides, built on MiniMax M3; the average model cost per deck is about $0.44 (Sina Tech, September 20) · Reported (cost per company disclosure)
- 2026-09-10 · Anthropic published a threat intelligence report describing distillation of Claude by Chinese labs, covering nearly 200 million exchanges across five campaigns; by its account, accounts tied to Alibaba generated 151 million exchanges from May to July, peaking at nearly 3 million a day across about 3,500 accounts; the targeted capabilities were agentic tool use, coding and data analysis, and logical reasoning; the coverage names Alibaba, Moonshot AI and DeepSeek (TechCrunch) · Verified
- 2026-09-08 · Three U.S. security agencies, the NSA, CISA and FBI, issued a joint statement accusing Chinese AI companies of aggressive and systematic industrial-scale distillation of U.S. frontier models; it named six companies (DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Zhipu), gave no damage estimate, announced no new sanctions or indictments, and set no timeline for further disclosures (Inside AI News, September 9) · Reported
- 2026-09-07 · Alibaba's Qwen team and Huazhong University of Science and Technology open-sourced Qwen-Drive-1.0-4B under Apache 2.0, built on the Qwen3.5-4B vision-language model with added 3D perception and trajectory-planning modules; code, weights and demo data are on GitHub, with imitation-learning and reinforcement-learning versions of the planner. Qwen's open-source coverage now reaches autonomous driving (TechNode) · Verified
- 2026-09-03 · Reuters (via Data Studios): Moonshot AI has confidentially filed for a Hong Kong listing seeking about $3 billion, with a concurrent private round negotiated at about $50 billion. It has raised more than $5.5 billion to date, including a May round above $2 billion joined by Meituan, China Mobile and CPE; Goldman Sachs, CICC and Deutsche Bank are the sponsors; before filing it unwound its red-chip structure for an onshore domicile. The same report says Moonshot is in talks with Microsoft, Amazon and Google on revenue-sharing arrangements to host Kimi K3, with no final agreement · Reported
- 2026-09-02 · Alibaba released Qwen3.8-Max-0902, a post-training upgrade with the architecture unchanged (2.4 trillion parameters, 1M-token context; base weights opened in August), focused on coding and long-horizon agent tasks. By Alibaba's own numbers, it ranks first on Code Arena WebDev at 1,691, ahead of Claude Opus 5 Max at 1,687; TerminalBench 3.0 rises from 11.3 to 29.0, QwenSWEBench V2 reaches 70.0, and WorkArena Elo rises from 1,348 to 1,468; API pricing stays at $2 per million input tokens and $6 per million output tokens (DataCamp) · Reported
- 2026-08-31 · Zhipu's first-half revenue was RMB 954 million, up about 400%, with API sales at 86.5% of the total; it still came in about 30% below analyst estimates, on a half-year loss of RMB 2.07 billion · Verified (Bloomberg, SCMP)
- 2026-08-19 · Kuaishou's Q2 results put Kling AI's quarterly revenue above RMB 850 million, up more than 200% year on year, with over 100 million users across 224 countries and territories as of June · Verified (Caijing, Sina Tech)
- 2026-08-15 · Hugging Face's state-of-open-models report: Alibaba's Qwen family topped 3 billion cumulative downloads, with over 300,000 derivative models; 2026 downloads on the platform were about 2.045 billion for Qwen, 418 million for Google and 227 million for Meta · Verified (Fortune, Aug 15; China Daily, Aug 16)
OpenRouter weekly usage: Chinese models stay ahead (trillion tokens)
| Week | Platform total | Chinese models | US models | Consecutive Chinese lead | Evidence |
|---|---|---|---|---|---|
| Sep 7-13 | 127 | 61.2 | 21.8 | 20th week | Reported |
| Week to Sep 27 | ~146 | — | — | 22nd week | Verified |
| Sep 28-Oct 4 | ~166 | 57.46 | 16.2 | 23rd week | Reported |
What This Means for Founders
Distribution is no longer the bottleneck: Qwen has passed 3 billion downloads, Chinese models have led weekly OpenRouter usage for 23 straight weeks, and GLM 5.3 now sits in Amazon Bedrock's enterprise catalog. The hard part is getting paid: Zhipu's first-half revenue came in about 30% below estimates, while companies that bill in dollars and earn more than half their revenue overseas are closing the China valuation discount themselves. Exit routes and rule risk are both moving: Zhipu and MiniMax are listed in Hong Kong and Moonshot AI has reportedly filed, while the distillation dispute has reached enforcement on both the U.S. and Chinese sides, with no penalties yet. Two signals to watch: whether China's Flash models hold their weekly usage lead after Haiku 5.5's price cut, and whether Zhipu's overseas cloud revenue share shows up as visible revenue in its results.
What This Means for Investors
Pricing power still sits with private markets: DeepSeek's valuation rose from about $52B post-money in June to about $71B pre-money in July talks, and Moonshot AI, with ARR reportedly above $300M in June, is still valued at a P/ARR around 100x; the roughly $50B valuation in talks at its Hong Kong filing matches the target reported in July. Hong Kong is the first test: MiniMax rose more than 17% on July 27 and fell 13.74% on October 8 (the Hang Seng Tech Index fell 2.89% that day, so the drop cannot be pinned on Haiku 5.5 alone); re-rating through a listing comes with high volatility. The overseas valuation anchor is moving later: Bloomberg says Anthropic could list as soon as mid-November, and either its public prospectus or a break in China's weekly usage lead would send us back to re-check this report's valuation view. This is not investment advice.
Where We Could Be Wrong
- 01If China's weekly usage lead on OpenRouter breaks after Haiku 5.5 launches, the lead rested mainly on the price gap of Flash models. Our view that open source is the main channel abroad would then narrow, and the valuation section would need a re-check. The baseline is 57.46 trillion vs 16.2 trillion tokens in the week to October 4; Space Bunny Alpha's origin is unconfirmed, and National Business Daily did not say which side it counts it on.
- 02If Zhipu's coming results show no visible scale of overseas cloud revenue share, and revenue still trails analyst estimates by about 30%, the move from developer platforms into a U.S. cloud provider's enterprise catalog stays a distribution story and the model layer's halfway point stretches out.
- 03If, within the second half of 2026, either the U.S. or China imposes penalties, sanctions or indictments over the distillation charges, our policy rating of 'enforcement started on both sides, no penalties yet' would move up, and the valuation of the going-global path would need a re-check.
Key Questions
Is Kimi K3 open-sourced, and what are its parameters and license?
Yes. On July 27, 2026, Moonshot AI released Kimi K3's full weights under a modified MIT license: 2.8T total / ~104B active parameters, ~1.56TB of files, plus three training-infra components (MoonEP, FlashKDA, AgentEnv). It topped Hugging Face's trending chart within 30 minutes with 4,000+ likes, with Day-0 support from Huawei Ascend and Alibaba Cloud and integrations from Cursor and Cognition.
When will DeepSeek IPO, and what is its latest valuation?
Per Bloomberg (July 14) and WSJ (July 15) reports, DeepSeek has begun onshore IPO preparations for a Shanghai listing, possibly filing this year with a 2027 debut. Valuation rose from ~$52B post-money in June's first round to ~$71B pre-money (~RMB 480B) in July talks; on July 26 it reportedly paused new fundraising to prioritize the IPO. The SSE issued a guideline on June 17 applying the STAR Market's fifth listing standard to large-model companies.
How did Hong Kong-listed Zhipu and MiniMax perform after lock-up expiry?
Highly volatile. Zhipu rose 13.35% on its first lock-up expiry day (July 8), reportedly adding over HK$100B in market value; MiniMax saw ~50% of shares unlock on July 9, with founders voluntarily extending lock-ups by 12 months. MiniMax then surged 17%+ on July 27, before Zhipu fell as much as 20% intraday and MiniMax over 12% on July 28. Per EY, H1 Hong Kong IPOs averaged a 61% first-day return—re-rating via listing comes with high volatility.
Watch & Listen
▶ Three Paths for Chinese AI Going Global — All Now ProvenYouTube · needs VPN in China
▶ A Window for Chinese AI to Go Global?YouTube · needs VPN in China
▶ Real AI Opportunities Hide Where Big Tech Isn't LookingYouTube · needs VPN in ChinaIn China: search WeChat Channels for 「倩姐投AI」; full library → Qian on AI
Sourcing and standards
Compiled from public sources; data current as of 2026.10.11. The text separates four evidence grades: verified, reported, per company disclosure and public market data, and states the derivation behind every estimate. When we get something wrong, the correction is written into the report body with the original call left visible, and logged publicly.
Research standards & corrections →📄 Full Report
Full report: 24 pages · provided to professional investors & partners only
This is the public preview. The full report includes the sections below. For compliance reasons it isn't posted publicly or offered as a free download. To request a copy, contact the FutureX team.
- 🔒IV. Agent Commercialization: SoftBank × Sierra and Japan's Willingness to Pay
- 🔒V. Zhipu and MiniMax: Post-IPO Re-rating and the Lock-up Stress Test
- 🔒VI. Policy Variables: The 'Tiered Export' Debate and the STAR Market's New Listing Track
- 🔒VII. Value Chain and Beneficiaries (A Neutral Framework)
- 🔒VIII. Risks and Watch List (H2 2026)
- 🔒Data & Sources Appendix
Where we stand on this
Open source breaks the deadlock: the cost curve has a new author
The dividing line of this AI cycle is the cost curve, and open source pushes it down. A closed model's moat is capital expenditure that must be refinanced; an open model's moat is a living community that needs none. The curve's author has changed from US closed labs to Chinese open models; value moves to applications and open-source chokepoints.
Full argument and falsification tests →US and China in AI: one cycle, two positions
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.
Full argument and falsification tests →Questions people ask next
Related Research
AI Agent
Agent Usage Grew Roughly 14x, but Pricing Power Goes to Data and Workflow Owners
Artificial Intelligence / Macro & Capital
AGI Is Here, Capital Is Losing Its Bearings: Valuations Doubled Before Any Public Test
AI Compute Economics
Tokens No Longer Have One Price: Cheap Tier Cut 90%, Fast Tier Costs 6x
Industry research from FutureX Capital's AI Lab, compiled from public information; not investment advice; contains no fund performance, AUM, or offer to raise capital.