The Information Machine

Chinese AI Models and Products Gain Structural Ground on US Rivals · history

Version 15

2026-07-01 18:25 UTC · 278 items

What

Chinese AI models now account for over 45% of OpenRouter token traffic (up from under 2% in late 2024), with open-source model processing on the platform reaching 65% of total traffic in June 2026 driven primarily by Chinese adoption [2][1]. Chinese models are priced as low as $0.18 per million tokens against a $4 frontier average, and Coinbase CEO Brian Armstrong disclosed his company is routing Chinese open-weight models as defaults for execution workloads [1][4]. Meituan's LongCat-2.0 (1.6T parameters) was trained from scratch on 50,000 domestic Huawei Atlas-950 chips—the first large Chinese model to use domestic hardware for both pre-training and inference, not just inference [6]. A research study and a public acknowledgment by Eric Schmidt—who helped design US chip export controls—each suggest the controls have accelerated China's open AI ecosystem rather than constraining it [12][13].

Why it matters

LongCat-2.0's domestic-chip training removes one of the remaining credible arguments that export controls keep large-scale Chinese AI training dependent on US hardware. Combined with named enterprise defections, Citibank and J.P. Morgan pricing data, and Schmidt's acknowledgment, the evidence base for the argument that export controls are producing a net strategic benefit has narrowed further in a single week.

Open questions

  • Can LongCat-2.0's domestic chip training success—50,000 Huawei Atlas-950 chips for a 1.6T-parameter model—be replicated across the broader Chinese AI ecosystem at frontier scale [6]?

  • Will enterprise adoption of Chinese open-weight models (Coinbase's LLM gateway default, UBS survey's 60% migration rate) produce durable lock-in, or remain price-conditional and reversible [4][5]?

  • If Eric Schmidt—who helped design US chip export controls—has acknowledged they have not achieved their intended effect [13], what policy revision follows, and on what timeline?

  • Can CXMT scale its buried-wordline DRAM architecture toward HBM production for AI accelerators at competitive cost relative to Samsung, SK Hynix, and Micron, as it approaches its STAR Market IPO [20][21]?

Narrative

Chinese AI providers have gained ground on US rivals across pricing, performance, and enterprise adoption over roughly 18 months. Citibank Research finds Chinese models charge as little as $0.18 per million tokens versus a $4 average for top frontier models [1], consistent with J.P. Morgan's finding of a 50x price differential [2]. Open-source processing on OpenRouter grew from 34% in January 2026 to 65% in June, driven primarily by Chinese model adoption [1], while Bloomberg documented a parallel fall in US model token share from approximately 70% to 30% in roughly one year [3]. Coinbase CEO Brian Armstrong disclosed that his company is experimenting with defaulting to Chinese open-weight models including GLM 5.2 and Kimi 2.7 through its internal LLM gateway, routing frontier models only to planning tasks while treating them as unnecessary for execution [4]. A UBS survey found 60% of enterprises monitoring AI budgets are migrating to cheaper or open-source Chinese models [5].

At the model layer, Meituan's LongCat-2.0 (1.6T parameters, 1M token context) became the first large Chinese model trained from scratch on domestic hardware—50,000 chips using Huawei's HCCL communication library and Atlas-950 architecture—for both pre-training and inference [6]. This distinguishes it from DeepSeek-V4-pro, which used domestic chips only for inference. LongCat-2.0 ranks in the top 3 on OpenRouter by call volume [6]. Z.ai's GLM-5.2 (753B parameters, MIT license) topped the Artificial Analysis Intelligence Index v4.1 at approximately $4.40 per million output tokens [7]; Nathan Lambert called it the first open-weight model to work credibly as a general coding agent, while Zvi Mowshowitz argued it is very likely heavily distilled from Claude Opus and occupies an awkward commercial niche [8][9]. DeepSeek raised $7.4B at a $50B valuation; The Information reported that a preview of Anthropic's Mythos model prompted CEO Liang Wenfeng to pursue the round to remain competitive [10][11].

The debate over US chip export controls has accumulated evidence that the policy has produced effects contrary to its stated aim. A research study measuring GitHub activity, research papers, and company publications found that Chinese developers increased activity on open LLM projects significantly more than US developers after major export controls took effect [12]. Eric Schmidt—whose SCSP think tank helped design the export control architecture—publicly acknowledged the controls have not achieved their intended effect [13]. Nvidia CEO Jensen Huang argues controls have functioned as industrial stimulus for Huawei and cede operating-layer standards to China [14]. Perplexity CEO Aravind Srinivas argues controls are the primary reason a roughly 12-month gap exists between Chinese open-source and US frontier models, and that by forcing China to build domestic infrastructure—where power, permits, and labor present no constraints—controls may be producing a more capable long-run competitor at the physical layer [15][16]. Anthropic CEO Dario Amodei holds the opposing position, calling restriction of China's AI access clearly in the US national security interest and dismissing counterarguments as 'fishy' [17].

At the semiconductor and infrastructure layer, China reclaimed the world's top-ranked supercomputer for the first time since 2018, built in Shenzhen without US GPUs [18][19]. CXMT is approaching China's largest semiconductor IPO on the STAR Market, with DRAM technology scaled toward the 10nm class tracing to approximately 2.8TB of Qimonda documentation [20][21]. Apple is reportedly seeking US government approval to buy memory chips from CXMT—which is on the US entity list—as AI-driven DRAM demand pushes prices higher [22]. The US raised concern with ASML that a banned EUV lithography tool reached China; ASML categorically denies it, citing active tracking of all approximately 314-340 EUV units worldwide [23][24].

Timeline

  • 2026-06-14: SemiAnalysis SMIC N+3 teardown finds TSMC N6-class logic density via DUV multi-patterning but at higher cost; Kirin 9030 Pro trails current flagship SoCs. [30]
  • 2026-06-16: DeepSeek raises $7.4B at a $50B valuation; founder Liang Wenfeng personally contributed approximately $3B. [10][33]
  • 2026-06-16: Z.ai releases GLM-5.2 (753B parameters, MIT license), topping the Artificial Analysis Intelligence Index v4.1 at approximately $4.40 per million output tokens. [7]
  • 2026-06-19: US Commerce Secretary Lutnick raises concern with ASML that a banned EUV tool reached China; ASML publicly denies it, saying it tracks all approximately 314-340 EUV units worldwide. [23][24][34]
  • 2026-06-23: SemiAnalysis reports CXMT is approaching China's largest semiconductor IPO; its DRAM technology traces to approximately 2.8TB of Qimonda documentation, scaled toward 10nm class. [20][21]
  • 2026-06-26: The Information reports Anthropic's Mythos model preview prompted DeepSeek CEO Liang Wenfeng to pursue the $7.4B fundraise to remain competitive. [11]
  • 2026-06-26: UBS survey finds 60% of enterprises monitoring AI budgets are migrating to cheaper or open-source Chinese models. [5]
  • 2026-06-27: J.P. Morgan report: Chinese models up to 50x cheaper per token than US equivalents; Chinese firms held over 45% of OpenRouter traffic by April 2026, up from under 2% in late 2024. [2]
  • 2026-06-27: Bloomberg reports US model token share on OpenRouter fell from approximately 70% to 30% in roughly one year. [3]
  • 2026-06-27: DeepSeek publishes DSpark, achieving 60-85% faster per-user token generation via Markov head and confidence scheduler. [28]
  • 2026-06-28: China reclaims the world's top-ranked supercomputer for the first time since 2018, built in Shenzhen without US GPUs. [18][19]
  • 2026-06-28: 'Owl Alpha' on OpenRouter reported to be Meituan's LongCat-2.0-Preview: a 1.6T-parameter MoE ranking #1 on Hermes Agent and #2 on Claude Code by usage. [29]
  • 2026-06-29: Apple reportedly seeks US government approval to buy memory chips from blacklisted CXMT as AI-driven DRAM demand pushes prices up. [22]
  • 2026-06-29: Jensen Huang argues chip export controls stimulate China's domestic semiconductor industry; Dario Amodei calls restricting China's AI access clearly in US national security interest. [14][17]
  • 2026-06-29: Perplexity CEO Srinivas argues export controls compressed the frontier model gap to approximately 12 months while forcing China to build superior data center infrastructure. [25][16][15]
  • 2026-06-30: Meituan's LongCat-2.0 (1.6T parameters) confirmed trained from scratch on 50,000 domestic Huawei Atlas-950 chips—the first large Chinese model to use domestic hardware for both pre-training and inference. [6]
  • 2026-06-30: Coinbase CEO Brian Armstrong discloses Coinbase is experimenting with Chinese open-weight models (GLM 5.2, Kimi 2.7) as the default in its internal LLM gateway for execution tasks. [4]
  • 2026-06-30: Citibank Research: Chinese models as cheap as $0.18 per million tokens vs $4 average for frontier models; OpenRouter open-source processing share grew from 34% in January 2026 to 65% in June. [1]
  • 2026-06-30: Eric Schmidt—whose SCSP think tank helped design US chip export controls—publicly acknowledges the controls have not achieved their intended effect. [13]
  • 2026-07-01: Research study finds US chip restrictions accelerated China's open AI ecosystem: Chinese developer activity on open LLM projects grew significantly more than US developer activity after major export controls. [12]

Perspectives

Jensen Huang (Nvidia)

Chip export controls don't prevent China from developing AI, function as industrial stimulus for Huawei, and cede operating-layer standards competition to China; the long-term risk is a world where US technology is absent from systems America most wants to influence.

Evolution: Consistent.

[14]

Aravind Srinivas (Perplexity)

Export controls are the primary reason a roughly 12-month frontier gap exists, but by forcing China to build domestic infrastructure—where power, permits, and labor present no constraints—controls are converting China into a more capable competitor at the physical layer; the right US strategy is investment in open-source models and nuclear energy.

Evolution: Consistent.

[25][16][15]

Dario Amodei (Anthropic)

Restricting China's AI access is clearly in the US national security interest; counterarguments are 'fishy'; Anthropic has actively lobbied for AI chip export controls.

Evolution: Consistent.

[17][15]

Brian Armstrong (Coinbase)

Coinbase is experimenting with Chinese open-weight models as defaults for execution tasks, routing frontier models only to planning; frontier models are overkill for routine AI workloads at current pricing.

Evolution: New voice this pass.

[4]

Rohan Paul (@rohanpaul_ai)

Tracks and amplifies Chinese AI gains across OpenRouter traffic, pricing data, enterprise migration, and LongCat-2.0's domestic chip milestone; frames the competition as one over physical inputs—electricity, minerals, and magnet supply chains—not just model quality.

Evolution: Scope expanded to include LongCat-2.0 domestic training milestone, Coinbase adoption, and Citibank pricing data.

[26][10][27][5][11][3][2][28][29][14][17][25][16][15][6][4][1][12]

Nathan Lambert (Interconnects)

GLM-5.2 is the first open-weight model to perform credibly as a general coding agent, comparable in significance to DeepSeek R1; it puts economic pressure on Anthropic's Claude Code revenue.

Evolution: Consistent.

[8]

Zvi Mowshowitz

GLM-5.2 is the strongest open-weight model available but is very likely heavily distilled from Claude Opus, trails frontier closed models substantially, and occupies an awkward commercial niche—not cheap enough for bulk tasks, not strong enough for the hardest ones.

Evolution: Consistent.

[9]

SemiAnalysis

Covers Chinese AI and semiconductor infrastructure with technical grounding: SMIC N+3 reaches TSMC N6 density but at higher cost; CXMT's DRAM traces to Qimonda documentation and is approaching a major IPO; coverage spans memory, AI accelerators, and logic chips.

Evolution: Consistent.

[30][20][21]

Tensions

  • Jensen Huang, Aravind Srinivas, and Eric Schmidt argue US chip export controls have not contained China's AI development—and a research study found controls accelerated China's open AI ecosystem [14][15][13][12]; Dario Amodei argues restriction is clearly in the US national security interest and counterarguments are 'fishy' [17]. [14][15][13][12][17]
  • Nathan Lambert argues GLM-5.2 is the first open-weight model to match closed frontier performance in coding agent harnesses [8]; Zvi Mowshowitz argues it is very likely heavily distilled from Claude Opus and trails frontier closed models substantially [9]. [8][9]
  • Enterprise migration data—UBS 60% migrating, Coinbase default routing, Citibank pricing—treats Chinese model adoption as a structural shift [5][4][1]; developer-community commentators argue the gains reflect spot-market price routing conditional on investor tolerance for losses [31]. [5][4][1][31]
  • US hyperscalers are projected to spend approximately 8.3x more than Chinese hyperscalers on AI infrastructure by 2027 [32]; UBS enterprise migration data, Citibank's pricing differential, and GLM-5.2's benchmark performance suggest the dollar gap is not translating proportionally into capability or market separation [5][1][7]. [32][5][1][7]
  • The Information's report that Anthropic's Mythos preview prompted DeepSeek's $7.4B raise implies the US frontier still holds a lead Chinese players are spending to close [11]; the same period's OpenRouter share data and Citibank pricing show Chinese models are winning the cost-performance layer where most enterprise work runs [1][3][5]. [11][1][3][5]
  • The US government believes a banned EUV lithography tool reached China; ASML categorically denies having ever shipped one, saying it tracks all approximately 314-340 EUV units worldwide [23][24]. [23][24]

Sources

  1. [1] Reuters: Chinese models charge as little as 18 cents per million tokens versus $4 average for top models, says CitiBank … — Rohan Paul Twitter (2026-06-30)
  2. [2] 🇨🇳🇺🇸Chinese AI models are up to 50 times cheaper than their American counterparts on a per-token basis. — Rohan Paul Twitter (2026-06-27)
  3. [3] "the share of tokens used for US models on OpenRouter has collapsed" Bloomberg — Rohan Paul Twitter (2026-06-27)
  4. [4] Coinbase CEO Brian Armstrong said Coinbase is experimenting with defaulting to Chinese open-weight models such as GLM 5.… — Rohan Paul Twitter (2026-06-30)
  5. [5] UBS says 60% of companies now watching AI budgets are moving to cheaper models and open-source Chinese models — Rohan Paul Twitter (2026-06-26)
  6. [6] 🇨🇳China claims a new milestone in locally trained AI, as Meituan rolls out LongCat-2.0. — Rohan Paul Twitter (2026-06-30)
  7. [7] GLM-5.2 is probably the most powerful text-only open weights LLM — Simon Willison (2026-06-17)
  8. [8] GLM-5.2 is the step change for open agents — Interconnects (2026-06-22)
  9. [9] GLM-5.2 Is The New Best Open Model — Zvi's AI Roundups (2026-06-22)
  10. [10] DeepSeek takes the crown as China’s most valuable AI startup after a massive $7.4B raise at a $50B valuation. — Rohan Paul Twitter (2026-06-16)
  11. [11] The Information reports that Anthropic’s Mythos preview spooked DeepSeek into fundraising. — Rohan Paul Twitter (2026-06-26)
  12. [12] U.S. chip restrictions helped push China to build and spread open AI models. — Rohan Paul Twitter (2026-07-01)
  13. [13] Eric Schmidt — whose SCSP think tank helped design the US chip export control architecture — just publicly admitted at h... — reactive:chinese-ai-competitive-rise (2026-06-27)
  14. [14] Jensen Huang explains how blocking China from Nvidia does not mean blocking China from AI. — Rohan Paul Twitter (2026-06-29)
  15. [15] Aravind Srinivas just explained why China’s open-source AI may become more powerful than ever. — Rohan Paul Twitter (2026-06-30)
  16. [16] AI at scale is constrained by physical inputs, and China has more slack in electricity plus dominant control over severa… — Rohan Paul Twitter (2026-06-30)
  17. [17] Dario Amodei has a really hardline view that China shouldn’t have strong AI. — Rohan Paul Twitter (2026-06-29)
  18. [18] China just reclaimed the world's #1 supercomputer for the first time since 2018 — and it did it without a single NVIDIA ... — reactive:chinese-ai-competitive-rise (2026-06-28)
  19. [19] The world's fastest supercomputer just went live in Shenzhen — and it contains zero US GPUs. — reactive:chinese-ai-competitive-rise (2026-06-28)
  20. [20] China’s CXMT Is Set to Challenge DRAM Incumbents — SemiAnalysis Twitter (2026-06-23)
  21. [21] CXMT (ChangXin Memory Technologies), China’s top domestic DRAM maker, is preparing for a major IPO on Shanghai’s STAR Ma... — reactive:chinese-ai-competitive-rise (2026-06-23)
  22. [22] Apple is reportedly seeking U.S. approval to buy memory chips from China's blacklisted CXMT as the AI boom sends DRAM pr... — reactive:chinese-ai-competitive-rise (2026-06-29)
  23. [23] ASML just became the center of a US-China chip fight after Washington said it fears a banned EUV lithography tool may ha… — Rohan Paul Twitter (2026-06-19)
  24. [24] ASML denies US government report that its EUV chipmaking tool ... — reactive:chinese-ai-competitive-rise
  25. [25] Opinion from a former Meta PM. — Rohan Paul Twitter (2026-06-30)
  26. [26] American AI startups are routing far more app traffic to Chinese LLMs. — Rohan Paul Twitter (2026-06-08)
  27. [27] Reuters: DeepSeek is going on a hiring sprint, aiming to double every department. — Rohan Paul Twitter (2026-06-26)
  28. [28] Fantastic, @deepseek_ai just published their new inference optimization method. — Rohan Paul Twitter (2026-06-27)
  29. [29] I’m hearing that "Owl Alpha", one of OpenRouter’s fastest-growing agent models, is actually Meituan LongCat-2.0-Preview — Rohan Paul Twitter (2026-06-28)
  30. [30] Is SMIC N+3’s Metal Pitch Smaller than Intel 18A’s? — SemiAnalysis Twitter (2026-06-14)
  31. [31] The AI model market is turning into an inference spot market. Developers are routing to whatever model gives the best co... — reactive:chinese-ai-competitive-rise (2026-06-08)
  32. [32] China is growing very quickly in AI, but the scale difference is brutal, spending gap is enormous. — Rohan Paul Twitter (2026-06-22)
  33. [33] DeepSeek has completed over 50 billion RMB in financing at a valuation exceeding $50 billion, per The Information. Found... — reactive:chinese-ai-competitive-rise (2026-06-16)
  34. [34] The US says ASML's top chip tool may be in China, but how? — reactive:chinese-ai-competitive-rise