The Information Machine

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

Version 18

2026-07-06 02:16 UTC · 317 items

What

Chinese AI models now account for 1.15 billion tracked open-source downloads versus 723 million for US models, with Qwen — not DeepSeek — as the primary driver through a broad range of model sizes, especially small ones [1]. At the pricing layer, Chinese models cost as low as $0.18 per million tokens against a $4 frontier average [2], and 60% of enterprises monitoring AI budgets have migrated to Chinese or open-source alternatives [4]. At the hardware layer, CXMT has entered the global RAM market at $138 module pricing [14] and Huawei is preparing to sell Ascend AI accelerator chips in South Korea in Q4 2026 — its first attempt to market AI compute outside China [17]. Apple is in talks to source memory from CXMT and YMTC for China-market devices [15], and the US export controls debate has accumulated prominent voices arguing the policy has not achieved its aims [19][20][21].

Why it matters

The open model download data adds academic grounding to the usage-share and enterprise-migration numbers: the shift in Chinese AI adoption is driven by a wide range of Qwen model sizes rather than a single flagship, which makes it harder to dismiss as a transient pricing arbitrage. The competition has also moved into hardware supply chains and international compute sales, with Chinese memory suppliers entering the global market and Chinese AI accelerators being evaluated for frontier model inference.

Open questions

  • Will South Korea — a US treaty ally hosting Samsung and SK Hynix — permit or block Huawei Ascend chip sales in Q4 2026, and how does the US respond to allied purchases of Chinese AI accelerators [17]?

  • Apple is seeking US government approval to source memory from blacklisted CXMT for China-market devices [15][16]; will the administration grant it, and does a yes signal a broader accommodation on Chinese chip supply chains?

  • Academic research finds China surpassed the US in open model downloads in summer 2025 and held a 1.15B-to-723M download lead by March 2026 [1]; does Qwen's dominance across model sizes, especially small ones, translate into durable enterprise default routing rather than benchmark-driven attention?

  • Eric Schmidt, ITIF, Jensen Huang, and Aravind Srinivas all argue export controls have not contained China's AI development [19][20][21][22]; what policy revision follows, and on what timeline?

Narrative

Chinese AI providers have gained structural ground on US rivals across pricing, model performance, enterprise adoption, and open-source download share over roughly 18 months. Academic research confirms that China surpassed the US in open-source model downloads in summer 2025; by March 2026, Chinese models had accumulated 1.15 billion tracked downloads against 723 million for US models [1]. The primary driver is Qwen, which dominates through a broad range of useful sizes — especially small, cheap-to-run models that developers reuse frequently — rather than a single flagship. DeepSeek leads only the very large segment above 250 billion parameters. Citibank Research separately finds Chinese models priced as low as $0.18 per million tokens against a $4 frontier average [2], consistent with J.P. Morgan's finding of up to a 50x price differential [3]. A UBS survey found 60% of enterprises monitoring AI budgets migrating to cheaper or open-source Chinese models [4], and Coinbase CEO Brian Armstrong disclosed his company is experimenting with Chinese open-weight models as defaults for execution tasks through its internal LLM gateway [5].

At the model layer, Z.ai's GLM-5.2 (753B parameters, MIT license) topped the Artificial Analysis Intelligence Index v4.1 [6], and Meituan's LongCat-2.0 (1.6T parameters) became the first large Chinese model trained from scratch on domestic hardware — 50,000 Huawei Atlas-950 chips — for both pre-training and inference [7]. DeepSeek V4 has since been deployed on InferenceX and evaluated across multiple accelerator types including Huawei Ascend NPUs [8], indicating the Ascend software ecosystem is gaining traction for frontier model inference. 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 [9][10].

At the hardware layer, China reclaimed the world's top-ranked supercomputer for the first time since 2018, built without US GPUs [11]. CXMT is approaching China's largest semiconductor IPO with DRAM technology scaled toward the 10nm class [12][13], and has entered the global RAM market at $138 module pricing, competing directly with Samsung and SK Hynix [14]. Apple is in talks to source memory from both CXMT and YMTC for China-market devices [15] while separately seeking US government approval to purchase from blacklisted CXMT [16]. Huawei is preparing to sell Ascend AI accelerator chips in South Korea in Q4 2026, its first attempt to market AI hardware outside China [17].

The debate over US chip export controls has accumulated evidence running against the policy's stated aims. A research study found Chinese developer activity on open LLM projects grew significantly more than US developer activity after major export controls [18]. Eric Schmidt — whose SCSP think tank helped design the export control architecture — publicly acknowledged the controls have not achieved their intended effect [19]. ITIF published an analysis arguing overly stringent controls are eroding US AI leadership [20]. Nvidia CEO Jensen Huang argues controls function as industrial stimulus for Huawei and cede operating-layer standards to China [21]; Perplexity CEO Aravind Srinivas argues controls compressed the frontier gap to approximately 12 months while forcing China to build superior domestic infrastructure [22]. Anthropic CEO Dario Amodei holds the opposing view, calling restriction of China's AI access clearly in US national security interest and dismissing counterarguments as 'fishy' [23].

Timeline

  • 2026-06-16: DeepSeek raises $7.4B at a $50B valuation; founder Liang Wenfeng personally contributed approximately $3B. [9][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. [6]
  • 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. [31][32][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. [12][13]
  • 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. [10]
  • 2026-06-26: UBS survey finds 60% of enterprises monitoring AI budgets are migrating to cheaper or open-source Chinese models. [4]
  • 2026-06-27: J.P. Morgan report: Chinese models up to 50x cheaper per token; Chinese firms held over 45% of OpenRouter traffic by April 2026, up from under 2% in late 2024. [3]
  • 2026-06-28: China reclaims the world's top-ranked supercomputer for the first time since 2018, built in Shenzhen without US GPUs. [11][35]
  • 2026-06-29: Apple is in talks to source memory chips from CXMT and YMTC for China-market devices; separately seeks US government approval to purchase from blacklisted CXMT. [16][15]
  • 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. [21][23]
  • 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. [24][25][22]
  • 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. [7]
  • 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. [5]
  • 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 to 65% in June 2026. [2]
  • 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. [19]
  • 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. [18]
  • 2026-07-01: CXMT enters the global RAM market with module pricing around $138, competing directly with Samsung and SK Hynix. [14]
  • 2026-07-01: InferenceX deploys DeepSeek V4 and evaluates performance across multiple accelerator types including Huawei Ascend NPUs. [8]
  • 2026-07-03: Huawei preparing to sell Ascend AI accelerator chips in South Korea in Q4 2026 — the first time it has attempted to market AI hardware outside China. [17]
  • 2026-07-04: Academic research confirms China surpassed the US in open model downloads in summer 2025; by March 2026, Chinese models had 1.15B tracked downloads vs 723M for US models, with Qwen as the dominant family across model sizes. [1]

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.

[21]

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 long-run competitor at the physical layer.

Evolution: Consistent.

[24][25][22]

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.

[23][22]

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: Consistent.

[5]

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.

[26]

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.

[27]

SemiAnalysis

Covers Chinese AI and semiconductor infrastructure with technical grounding: CXMT's DRAM traces to Qimonda documentation and is approaching a major IPO; LongCat-2.0's domestic chip pre-training is the first for a large Chinese model; InferenceX has evaluated DeepSeek V4 on Huawei Ascend NPUs.

Evolution: Consistent.

[12][13][8]

ITIF (Information Technology and Innovation Foundation)

Overly stringent US export controls are eroding American AI leadership rather than protecting it, adding a policy-research institutional voice to the Huang/Srinivas/Schmidt position.

Evolution: Consistent.

[20]

Tensions

  • Jensen Huang, Aravind Srinivas, Eric Schmidt, and ITIF argue US chip export controls have not contained China's AI development and are actively harming US competitiveness [21][22][19][18][20]; Dario Amodei argues restriction is clearly in the US national security interest and counterarguments are 'fishy' [23]. [21][22][19][18][20][23]
  • Nathan Lambert argues GLM-5.2 is the first open-weight model to match closed frontier performance in coding agent harnesses [26]; Zvi Mowshowitz argues it is very likely heavily distilled from Claude Opus and trails frontier closed models substantially [27]. [26][27]
  • Enterprise migration data — UBS 60% migrating, Coinbase default routing, Citibank pricing, and academic download-share research — treats Chinese model adoption as structural [4][5][2][1]; developer-community commentators argue the gains reflect spot-market price routing conditional on investor tolerance for losses [28]. [4][5][2][1][28]
  • The Information's report that Anthropic's Mythos preview prompted DeepSeek's $7.4B raise implies the US frontier still holds a lead that Chinese players are spending to close [10]; the same period's OpenRouter share data and Citibank pricing show Chinese models are winning the cost-performance tier where most enterprise work runs [2][29]. [10][2][29]
  • CXMT is entering the global RAM market at $138 module pricing, signaling competitive intent against Samsung and SK Hynix [14]; an investment analyst argues the Apple-CXMT approval headline is 'almost pure FUD' and changes nothing for memory incumbents [30]. [14][30]
  • 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 [31][32]. [31][32]

Sources

  1. [1] A paper on open model adoption and finds Chinese models, led by Qwen, now dominate. — Rohan Paul Twitter (2026-07-04)
  2. [2] 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)
  3. [3] 🇨🇳🇺🇸Chinese AI models are up to 50 times cheaper than their American counterparts on a per-token basis. — Rohan Paul Twitter (2026-06-27)
  4. [4] 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)
  5. [5] 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)
  6. [6] GLM-5.2 is probably the most powerful text-only open weights LLM — Simon Willison (2026-06-17)
  7. [7] 🇨🇳China claims a new milestone in locally trained AI, as Meituan rolls out LongCat-2.0. — Rohan Paul Twitter (2026-06-30)
  8. [8] This week the InferenceX team discusses what it took to get DeepSeek V4 on InferenceX, changes in the model architecture… — SemiAnalysis Twitter (2026-07-01)
  9. [9] 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)
  10. [10] The Information reports that Anthropic’s Mythos preview spooked DeepSeek into fundraising. — Rohan Paul Twitter (2026-06-26)
  11. [11] 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)
  12. [12] China’s CXMT Is Set to Challenge DRAM Incumbents — SemiAnalysis Twitter (2026-06-23)
  13. [13] 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)
  14. [14] 7.2K views · 178 reactions | CHINA’S CXMT ENTERS GLOBAL RAM MARKET Chinese semiconductor firm ChangXin Memory Technologies (CXMT) has entered the global RAM market, reportedly offering memory modules priced around $138 amid higher global pricing. As China seeks semiconductor independence, CXMT’s expansion signals intensified competition in DRAM manufacturing, traditionally dominated by companies like Samsung and SK Hynix. Important Details: • CXMT entering global DRAM market • Competitive $138 pricing benchmark • Aims to challenge dominant suppliers • Supports China’s chip self-sufficiency • Expands semiconductor supply chain • Impacts global memory pricing • Strategic tech competition factor Memory chips are critical for AI systems, data centers, and consumer electronics. The semiconductor battleground is no longer just about processors — memory now shapes global tech power. #Unboxfactory #SemiconductorRace #ChinaTech #GlobalMarkets #RAMIndustry | Unbox Factory — reactive:chinese-ai-competitive-rise
  15. [15] Apple is in talks to source memory chips from China’s CXMT and YMTC for China-market devices, putting $AAPL at the cente... — reactive:chinese-ai-competitive-rise (2026-07-02)
  16. [16] 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)
  17. [17] Huawei is preparing to sell its Ascend AI accelerator chips in South Korea in Q4 2026 — the first time it has attempted ... — reactive:chinese-ai-competitive-rise (2026-07-03)
  18. [18] U.S. chip restrictions helped push China to build and spread open AI models. — Rohan Paul Twitter (2026-07-01)
  19. [19] 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)
  20. [20] Overly Stringent Export Controls Chip Away at American AI Leadership | Blogs | May 5, 2025 | ITIF — reactive:chinese-ai-competitive-rise
  21. [21] Jensen Huang explains how blocking China from Nvidia does not mean blocking China from AI. — Rohan Paul Twitter (2026-06-29)
  22. [22] Aravind Srinivas just explained why China’s open-source AI may become more powerful than ever. — Rohan Paul Twitter (2026-06-30)
  23. [23] Dario Amodei has a really hardline view that China shouldn’t have strong AI. — Rohan Paul Twitter (2026-06-29)
  24. [24] Opinion from a former Meta PM. — Rohan Paul Twitter (2026-06-30)
  25. [25] 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)
  26. [26] GLM-5.2 is the step change for open agents — Interconnects (2026-06-22)
  27. [27] GLM-5.2 Is The New Best Open Model — Zvi's AI Roundups (2026-06-22)
  28. [28] 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)
  29. [29] "the share of tokens used for US models on OpenRouter has collapsed" Bloomberg — Rohan Paul Twitter (2026-06-27)
  30. [30] Memory investors, don't panic the Apple-CXMT headline is almost pure FUD and here is why it changes nothing for memory s... — reactive:chinese-ai-competitive-rise (2026-07-01)
  31. [31] 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)
  32. [32] ASML denies US government report that its EUV chipmaking tool ... — reactive:chinese-ai-competitive-rise
  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
  35. [35] The world's fastest supercomputer just went live in Shenzhen — and it contains zero US GPUs. — reactive:chinese-ai-competitive-rise (2026-06-28)