The Information Machine

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

Version 14

2026-06-30 08:15 UTC · 238 items

What

Chinese AI providers have gained ground on US rivals across pricing, model performance, and enterprise adoption, while a policy disagreement among US industry leaders over chip export controls has become a central dispute. Chinese firms held 45%+ of OpenRouter token traffic by April 2026, up from under 2% in late 2024, with Chinese models up to 50x cheaper per token than US equivalents [1]. Nvidia CEO Jensen Huang and Perplexity CEO Aravind Srinivas argue chip export controls are not containing China's AI development—and may be accelerating it by forcing China to build domestic data center infrastructure faster [6][10][7]. Anthropic CEO Dario Amodei holds the opposing position, calling restriction of Chinese AI access clearly in the US national security interest [8].

Why it matters

The combination of documented cost-performance gains and the Srinivas/Huang critique of export controls puts the two main US responses to Chinese AI competition—outcompeting on capability and restricting chip access—under simultaneous pressure. If export controls are primarily compressing the frontier model gap to roughly 12 months while forcing China to build superior physical AI infrastructure, the policy may be producing the opposite of its intended effect at the layer that will matter most for long-run AI scale.

Open questions

  • Are chip export controls producing a net strategic benefit for the US, or—as Srinivas argues—compressing the frontier model gap to roughly 12 months while forcing China to build faster data center infrastructure where power, permits, and labor pose no constraints [7][9]?

  • Apple's reported request for US government approval to buy CXMT memory chips [18] suggests Chinese DRAM cost advantages extend into consumer electronics supply chains—will the US grant exceptions that further legitimate CXMT's commercial position?

  • Does the J.P. Morgan 50x price differential and OpenRouter traffic shift from under 2% to 45%+ Chinese share in roughly 18 months represent a durable structural advantage, or a pricing strategy conditional on investor tolerance for losses [1]?

  • Can CXMT scale its buried-wordline DRAM architecture toward HBM production for AI accelerators, and at what cost relative to Samsung, SK Hynix, and Micron [16][17]?

Narrative

Chinese AI providers have gained ground on US rivals across pricing, model performance, and enterprise adoption. A J.P. Morgan report found Chinese models are up to 50x cheaper per token than US equivalents, with Chinese firms holding over 45% of OpenRouter token traffic by April 2026, up from under 2% in late 2024 [1]. Bloomberg confirmed US model token share on OpenRouter fell from approximately 70% to 30% in roughly one year [2]. A UBS survey found 60% of enterprises monitoring AI budgets are migrating to cheaper or open-source Chinese models [3]. On model performance, 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 [4]. A model operating anonymously as 'Owl Alpha' on OpenRouter is 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, processing 10.1T monthly tokens with 242% monthly growth [5].

A policy dispute over US chip export controls has become a distinct front. Nvidia CEO Jensen Huang argues that blocking China from Nvidia chips does not prevent China from developing AI, that restrictions have functioned as an industrial stimulus for Huawei, and that the real competition is over who sets operating-layer standards—chips, energy, infrastructure, models, and applications [6]. Perplexity CEO Aravind Srinivas extends this: he argues export controls are the primary reason a roughly 12-month gap currently exists between Chinese open-source models and US frontier AI, and that by forcing China to build domestic infrastructure, controls may be converting China into a more capable competitor at the physical layer [7]. Anthropic CEO Dario Amodei disagrees, calling restriction of China's AI access clearly in the US national security interest and dismissing counterarguments as 'fishy' [8]. Rohan Paul and Srinivas frame the underlying competition as one over physical inputs—electricity, minerals, and magnet supply chains—rather than software or model quality alone, noting China holds structural advantages in each: greater electricity surplus, faster data center permitting, and dominant control over minerals and magnets that data centers and chips require [9][10].

DeepSeek raised $7.4B at a $50B valuation, with The Information reporting that a preview of Anthropic's Mythos model prompted CEO Liang Wenfeng to pursue the round to remain competitive [11][12]—suggesting US frontier development still sets a pace Chinese providers are spending to close. DeepSeek plans to double all departments and shift from model research to full-stack AI product development [13]. It also published DSpark, an inference optimization achieving 60-85% faster per-user token generation via a Markov head and confidence scheduler [14], and posted roles for MW-to-GW scale data center design [15].

At the semiconductor layer, CXMT is approaching China's largest semiconductor IPO on Shanghai's STAR Market, with DRAM technology scaled toward the 10nm class [16][17]. 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 up [18]. SemiAnalysis found SMIC's N+3 process reaches TSMC N6-class logic density via DUV multi-patterning but at higher cost [19]. The US government raised concern with ASML that a banned EUV tool reached China; ASML categorically denies it, citing active tracking of all ~314-340 EUV units worldwide [20][21].

Timeline

  • 2026-01: Chinese models begin outpacing US models as the primary growth driver of token consumption on OpenRouter. [22]
  • 2026-06: Chinese models reach 61% of OpenRouter weekly token market share. [31]
  • 2026-06-10: DeepSeek posts a role for MW-to-GW scale data center design, interpreted as a shift to heavy-asset vertical integration. [15]
  • 2026-06-13: MiniMax M3 (428B parameters, 23B activated) released with day-0 MXFP8 weights and block sparse attention 9x faster than M2.7. [25][26]
  • 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. [19]
  • 2026-06-16: DeepSeek raises $7.4B at a $50B valuation; founder Liang Wenfeng personally contributed ~$3B. [11][32]
  • 2026-06-16: Z.ai releases GLM-5.2 (753B parameters, MIT license), topping the Artificial Analysis Intelligence Index v4.1 with a score of 51 and a 1M-token context window. [4]
  • 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 ~314-340 EUV units worldwide. [20][21][33]
  • 2026-06-22: Nathan Lambert calls GLM-5.2 the first open-weight model to work credibly as a general coding agent; Zvi Mowshowitz argues heavy distillation from Claude Opus and a persisting frontier gap. [23][24]
  • 2026-06-23: SemiAnalysis reports CXMT is approaching China's largest semiconductor IPO; its DRAM technology traces to ~2.8TB of Qimonda documentation and ~7,000 patents, scaled toward 10nm class. [16][17]
  • 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. [12]
  • 2026-06-26: UBS survey finds 60% of enterprises monitoring AI budgets are migrating to cheaper or open-source Chinese models; model routing is the dominant cost management strategy. [3]
  • 2026-06-27: J.P. Morgan report finds 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. [1]
  • 2026-06-27: Bloomberg reports US model token share on OpenRouter fell from approximately 70% to 30% in roughly one year. [2]
  • 2026-06-27: DeepSeek publishes DSpark, achieving 60-85% faster per-user token generation via Markov head and confidence scheduler for selective draft verification. [14]
  • 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, with 242% monthly growth and 10.1T monthly tokens. [5]
  • 2026-06-29: Apple reportedly seeks US government approval to buy memory chips from blacklisted CXMT as AI-driven DRAM demand pushes prices up. [18]
  • 2026-06-29: Jensen Huang argues chip export controls stimulate China's domestic semiconductor industry rather than containing AI development; Dario Amodei calls restricting China's AI access clearly in US national security interest. [6][8]
  • 2026-06-29: Perplexity CEO Srinivas argues export controls compressed the frontier model gap to ~12 months while forcing China to build superior data center infrastructure where power, permits, and labor pose no constraints. [10][9][7]

Perspectives

Jensen Huang (Nvidia)

Chip export controls don't prevent China from developing AI, function as industrial stimulus for Huawei, and cede the 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: New voice this pass.

[6]

Aravind Srinivas (Perplexity)

Export controls are the primary reason a ~12-month frontier gap exists, but by forcing China to build domestic infrastructure—where it faces no constraints on power, permits, or labor—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: New voice this pass.

[10][9][7]

Dario Amodei (Anthropic)

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

Evolution: New voice this pass.

[8][7]

Rohan Paul (@rohanpaul_ai)

Tracks Chinese AI gains across OpenRouter traffic, J.P. Morgan pricing data, UBS enterprise migration, and DeepSeek's capital raise; frames the competition as one over physical inputs—electricity, minerals, and magnet supply chains—not just model quality, and amplifies both the Huang/Srinivas export-control critique and Amodei's hardline stance.

Evolution: Scope expanded this pass to include the export control policy debate and physical infrastructure framing.

[22][11][13][3][12][2][1][14][5][6][8][10][9][7]

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.

[23]

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.

[24]

SemiAnalysis

Covers Chinese AI and semiconductor infrastructure with technical qualification: SMIC N+3 reaches TSMC N6 density but at higher cost; CXMT's DRAM technology traces to Qimonda documentation and is approaching a major IPO; Alibaba's T-Head capital raise signals expanded AI chip ambition.

Evolution: Consistent; coverage spans memory (CXMT), AI accelerators (T-Head), and logic chips (SMIC).

[25][26][15][19][16][27][17]

ASML

Categorically denies ever shipping an EUV lithography machine to China, saying it actively tracks all ~314-340 EUV units worldwide and that reports to the contrary are 'inaccurate and damaging to our reputation.'

Evolution: Consistent; denial is firm and backed by claimed active unit tracking.

[21][20][28]

Tensions

  • Jensen Huang and Aravind Srinivas argue US chip export controls don't contain China's AI development and may be accelerating it by forcing superior domestic infrastructure investment [6][7]; Dario Amodei argues restriction is clearly in the US national security interest and counterarguments are 'fishy' [8]. [6][7][8]
  • Nathan Lambert argues GLM-5.2 is the first open-weight model to match closed frontier performance in coding agent harnesses [23]; Zvi Mowshowitz argues it is very likely heavily distilled from Claude Opus, trails frontier closed models substantially, and occupies an awkward commercial niche [24]. [23][24]
  • The US government believes a banned EUV lithography tool reached China; ASML categorically denies having ever shipped one, saying it tracks all ~314-340 EUV units worldwide [20][21][28]. [20][21][28]
  • Rohan Paul and UBS enterprise survey data treat Chinese model adoption as a structural competitive shift affecting corporate budgets [3][2][1]; developer-community commentators argue the gains reflect spot-market price routing and are conditional rather than lasting preference [29]. [3][2][1][29]
  • US hyperscalers are projected to spend approximately 8.3x more than Chinese hyperscalers on AI infrastructure by 2027 [30]; UBS enterprise migration data, J.P. Morgan's 50x price differential, and GLM-5.2's benchmark performance suggest the dollar gap is not translating proportionally into capability or market separation [3][4][1]. [30][3][4][1]
  • The Information's report that Anthropic's Mythos preview prompted DeepSeek's $7.4B raise [12] implies the US frontier still holds a lead Chinese players are spending to close; the same period's OpenRouter share data and J.P. Morgan pricing findings suggest Chinese models are already winning the cost-performance competition where most enterprise work happens [2][1][3]. [12][2][1][3]

Sources

  1. [1] 🇨🇳🇺🇸Chinese AI models are up to 50 times cheaper than their American counterparts on a per-token basis. — Rohan Paul Twitter (2026-06-27)
  2. [2] "the share of tokens used for US models on OpenRouter has collapsed" Bloomberg — Rohan Paul Twitter (2026-06-27)
  3. [3] 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)
  4. [4] GLM-5.2 is probably the most powerful text-only open weights LLM — Simon Willison (2026-06-17)
  5. [5] 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)
  6. [6] Jensen Huang explains how blocking China from Nvidia does not mean blocking China from AI. — Rohan Paul Twitter (2026-06-29)
  7. [7] Aravind Srinivas just explained why China’s open-source AI may become more powerful than ever. — Rohan Paul Twitter (2026-06-30)
  8. [8] Dario Amodei has a really hardline view that China shouldn’t have strong AI. — Rohan Paul Twitter (2026-06-29)
  9. [9] 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)
  10. [10] Opinion from a former Meta PM. — Rohan Paul Twitter (2026-06-30)
  11. [11] 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)
  12. [12] The Information reports that Anthropic’s Mythos preview spooked DeepSeek into fundraising. — Rohan Paul Twitter (2026-06-26)
  13. [13] Reuters: DeepSeek is going on a hiring sprint, aiming to double every department. — Rohan Paul Twitter (2026-06-26)
  14. [14] Fantastic, @deepseek_ai just published their new inference optimization method. — Rohan Paul Twitter (2026-06-27)
  15. [15] DeepSeek is going heavy-asset. — SemiAnalysis Twitter (2026-06-10)
  16. [16] China’s CXMT Is Set to Challenge DRAM Incumbents — SemiAnalysis Twitter (2026-06-23)
  17. [17] 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)
  18. [18] 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)
  19. [19] Is SMIC N+3’s Metal Pitch Smaller than Intel 18A’s? — SemiAnalysis Twitter (2026-06-14)
  20. [20] 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)
  21. [21] ASML denies US government report that its EUV chipmaking tool ... — reactive:chinese-ai-competitive-rise
  22. [22] American AI startups are routing far more app traffic to Chinese LLMs. — Rohan Paul Twitter (2026-06-08)
  23. [23] GLM-5.2 is the step change for open agents — Interconnects (2026-06-22)
  24. [24] GLM-5.2 Is The New Best Open Model — Zvi's AI Roundups (2026-06-22)
  25. [25] DAY 0 ALERT: @MiniMax_AI M3 is now available on HuggingFace & has been added to InferenceX. The M3 architecture has … — SemiAnalysis Twitter (2026-06-13)
  26. [26] Congrats to @vllm_project & @lmsysorg for releasing MiniMax M3 428B on both the CUDA & ROCm stack on day 0! Mini… — SemiAnalysis Twitter (2026-06-13)
  27. [27] Alibaba's core chip entity behind their PPUs, T-Head, just filed a business registration change lifting registered capit… — SemiAnalysis Twitter (2026-06-23)
  28. [28] If an EUV machine reached China, three years of export controls failed silently. ASML denies it - they track all 314 in ... — reactive:chinese-ai-competitive-rise (2026-06-21)
  29. [29] 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)
  30. [30] China is growing very quickly in AI, but the scale difference is brutal, spending gap is enormous. — Rohan Paul Twitter (2026-06-22)
  31. [31] Chinese AI models hit 61% market share on OpenRouter - LinkedIn — reactive:chinese-ai-competitive-rise
  32. [32] 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)
  33. [33] The US says ASML's top chip tool may be in China, but how? — reactive:chinese-ai-competitive-rise