The Information Machine

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

Version 21

2026-07-11 08:09 UTC · 333 items

What

Chinese AI providers have taken positions across the model, pricing, enterprise adoption, and hardware layers of the AI stack. Chinese open-source models account for 1.15 billion tracked downloads vs. 723 million for US models [1]; Chinese pricing runs as low as $0.18 per million tokens against a $4 frontier average [3]; and 60% of enterprises monitoring AI budgets have migrated to cheaper or open-source Chinese models [4]. At the hardware layer, CXMT has entered global RAM markets [7] with Chinese memory module vendors now switching supply away from Samsung, Micron, and SK Hynix toward CXMT chips [9]; Huawei is marketing Ascend accelerators internationally [12][13]; and DeepSeek has spent approximately one year developing its own AI inference chip [15].

Why it matters

The combination of software wins at the cost-performance tier and hardware moves toward self-sufficiency reflects vertical integration across the AI stack. Whether DeepSeek can solve not just chip design but fabrication access—the harder constraint under US export controls—determines whether the chip effort removes a real point of policy leverage or remains a design exercise.

Open questions

  • Hardware Busters argues silicon design was never the bottleneck for DeepSeek's chip development—manufacturing access under US export restrictions is the real constraint [19]; can DeepSeek secure advanced fabrication capacity, and does that determine whether the effort matters?

  • Huawei is seeking Ascend AI chip customers across the Middle East and Southeast Asia, many of them US security partners, in addition to planned South Korea sales in Q4 2026 [12][13]; what is the US response to allied purchases of Chinese AI accelerators?

  • Chinese memory module vendors are switching supply chains from Samsung/Micron/SK Hynix to CXMT [9] as Apple separately seeks US approval to source from blacklisted CXMT [10][11]; does the domestic supply chain shift give CXMT the scale to compete globally regardless of US export decisions?

  • Eric Schmidt, ITIF, Jensen Huang, and Aravind Srinivas all argue export controls have not contained China's AI development [21][24][22][23]; does the accumulated evidence of Chinese vertical integration across model downloads, pricing, and hardware layers change the policy calculus?

Narrative

Chinese AI providers have taken positions across the model, pricing, enterprise adoption, and hardware layers of the AI stack over roughly 18 months. Academic research confirms 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], with Qwen dominant across a broad range of useful sizes. Tencent's Hy3 (295B total parameters, 21B active, MIT license) adds another entry, claiming to rival open-source models with two to five times more parameters [2]. At the pricing layer, Chinese models cost as low as $0.18 per million tokens against a $4 frontier average [3], and 60% of enterprises monitoring AI budgets have migrated to cheaper or open-source Chinese models [4]. Coinbase CEO Brian Armstrong disclosed his company routes Chinese open-weight models as defaults for execution tasks through its internal LLM gateway [5].

At the hardware layer, China reclaimed the world's top-ranked supercomputer in June 2026, built without US GPUs [6]. CXMT entered the global RAM market at $138 module pricing [7] and secured a reported $3 billion memory supply deal with Tencent [8]; Chinese memory module vendors are separately switching their supply chains from Samsung, Micron, and SK Hynix to CXMT chips, a domestic consolidation that builds CXMT's scale independent of its export market ambitions [9]. Apple is in talks to source memory from CXMT and YMTC for China-market devices while seeking US government approval to buy from blacklisted CXMT [10][11]. Huawei is marketing Ascend AI accelerators internationally—planning South Korea sales in Q4 2026 and seeking customers across the Middle East and Southeast Asia [12][13]. Meituan's LongCat-2.0 (1.6 trillion parameters) was trained from scratch on 50,000 Huawei Atlas-950 chips, the first large Chinese model to use domestic hardware for both pre-training and inference [14].

Multiple outlets confirm DeepSeek has been developing its own AI inference chip for approximately one year, hiring engineers and meeting with hardware and silicon partners, with US export controls as the primary stated driver [15][16][17][18]. Hardware Busters argues that silicon design was never the harder constraint—manufacturing access under US export restrictions is—making the effort's significance contingent on whether DeepSeek can secure advanced fabrication capacity rather than simply complete a chip design [19]. That distinction matters: US hyperscalers face the design challenge; DeepSeek faces both design and fab access.

The debate over whether US export controls have achieved their stated aims has accumulated substantial evidence running against the policy. A research study found Chinese developer activity on open LLM projects grew more than US activity after major export controls took effect [20]. Eric Schmidt—whose SCSP think tank helped design the export control architecture—publicly acknowledged the controls have not achieved their intended effect [21]. Nvidia CEO Jensen Huang argues controls function as industrial stimulus for Huawei [22]; Perplexity CEO Aravind Srinivas argues controls compressed the frontier gap to roughly 12 months while forcing China to build superior domestic infrastructure [23]; ITIF argues overly stringent controls are eroding US AI leadership [24]. Anthropic CEO Dario Amodei holds the opposing view, calling restriction clearly in the US national security interest and dismissing counterarguments as 'fishy' [25].

Timeline

  • 2026-06-16: DeepSeek raises $7.4B at a $50B valuation; The Information reports Anthropic's Mythos preview prompted the fundraise. [37][34]
  • 2026-06-16: Z.ai releases GLM-5.2 (753B parameters, MIT license), topping the Artificial Analysis Intelligence Index v4.1. [38]
  • 2026-06-23: SemiAnalysis reports CXMT is approaching China's largest semiconductor IPO with DRAM technology scaled toward the 10nm class. [30][31]
  • 2026-06-26: UBS survey finds 60% of enterprises monitoring AI budgets 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. [39]
  • 2026-06-28: China reclaims the world's top-ranked supercomputer for the first time since 2018, built without US GPUs. [6][40]
  • 2026-06-29: Apple in talks to source memory from CXMT and YMTC for China-market devices; separately seeks US government approval to purchase from blacklisted CXMT. [10][11]
  • 2026-06-29: CXMT secures a reported $3 billion memory supply deal with Tencent. [8]
  • 2026-06-30: Meituan's LongCat-2.0 (1.6T parameters) trained from scratch on 50,000 Huawei Atlas-950 chips—first large Chinese model to use domestic hardware for both pre-training and inference. [14]
  • 2026-06-30: Citibank Research: Chinese models as cheap as $0.18 per million tokens vs. $4 frontier average; OpenRouter open-source processing share grew from 34% in January to 65% in June 2026. [3]
  • 2026-06-30: Eric Schmidt publicly acknowledges the chip export controls his SCSP think tank helped design have not achieved their intended effect. [21]
  • 2026-07-01: Research study finds US chip restrictions accelerated China's open AI ecosystem: Chinese developer activity on open LLM projects grew more than US activity after major export controls. [20]
  • 2026-07-01: CXMT enters the global RAM market with module pricing around $138, competing with Samsung and SK Hynix. [7]
  • 2026-07-03: Huawei preparing to sell Ascend AI accelerator chips in South Korea in Q4 2026—its first international AI hardware sales effort. [12]
  • 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 dominant across model sizes. [1]
  • 2026-07-06: Tencent releases Hy3 (295B parameters, 21B active, MIT license), claiming to rival open-source models with 2-5x more parameters. [2]
  • 2026-07-06: Bloomberg reports Huawei seeking Ascend AI chip customers across the Middle East and Southeast Asia, expanding beyond its previously reported South Korea effort. [13]
  • 2026-07-07: Reuters reports DeepSeek has been developing its own AI inference chip for approximately one year; confirmed by multiple outlets. [15][16][17][18][41]
  • 2026-07-08: Tom's Hardware reports Chinese memory module vendors switching supply from Samsung, Micron, and SK Hynix to CXMT chips, accelerating CXMT's domestic market consolidation. [9]
  • 2026-07-08: Hardware Busters argues silicon design was never the bottleneck for DeepSeek's chip effort; manufacturing access under US export restrictions is the harder constraint. [19]

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 to China; the long-term risk is a world where US technology is absent from systems America most wants to influence.

Evolution: Consistent.

[22]

Aravind Srinivas (Perplexity)

Export controls created a roughly 12-month frontier gap, but by forcing China to build domestic infrastructure they are converting China into a more capable long-run competitor at the physical layer.

Evolution: Consistent.

[26][27][23]

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.

[25][23]

Eric Schmidt (SCSP) / ITIF

Schmidt publicly acknowledged the controls his think tank helped design have not achieved their intended effect; ITIF argues overly stringent controls are eroding US AI leadership, adding institutional weight to the same position.

Evolution: Consistent.

[21][24]

Nathan Lambert (Interconnects) vs. Zvi Mowshowitz

Lambert argues GLM-5.2 is the first open-weight model to perform credibly as a general coding agent, comparable in significance to DeepSeek R1; Mowshowitz argues it is very likely heavily distilled from Claude Opus and trails frontier closed models substantially.

Evolution: Consistent disagreement.

[28][29]

Hardware Busters

Silicon design was never the bottleneck for DeepSeek's chip effort; manufacturing access under US export restrictions is the harder constraint, making the effort's significance contingent on fab access rather than design completion.

Evolution: First cited in this thread.

[19]

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; DeepSeek V4 has been evaluated on Huawei Ascend NPUs.

Evolution: Consistent.

[30][31][32]

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]

Tensions

  • Jensen Huang, Aravind Srinivas, Eric Schmidt, and ITIF argue US chip export controls have not contained China's AI development and are harming US competitiveness [22][23][21][24]; Dario Amodei argues restriction is clearly in the US national security interest and counterarguments are 'fishy' [25]. [22][23][21][24][25]
  • Multiple outlets confirmed DeepSeek is designing its own AI inference chip driven by export controls [15][17]; Hardware Busters argues silicon design was never the bottleneck and manufacturing/fabrication access under US restrictions is the real constraint [19]. [15][17][19]
  • Nathan Lambert argues GLM-5.2 is the first open-weight model to match closed frontier performance in coding agent harnesses [28]; Zvi Mowshowitz argues it is very likely heavily distilled from Claude Opus and trails frontier closed models substantially [29]. [28][29]
  • Enterprise migration data—UBS 60% migrating, Coinbase default routing, Citibank pricing, academic download-share research—treats Chinese model adoption as structural [4][5][3][1]; developer-community commentators argue the gains reflect spot-market price routing conditional on investor tolerance for losses [33]. [4][5][3][1][33]
  • 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 [34]; the same period's OpenRouter share data and Citibank pricing show Chinese models are winning the cost-performance tier where most enterprise work runs [3]. [34][3]
  • 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 [35][36]. [35][36]

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] tencent/Hy3 — Simon Willison (2026-07-06)
  3. [3] 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)
  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] 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)
  7. [7] 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
  8. [8] China's CXMT wins $3 billion memory supply deal with Tencent ... — reactive:chinese-ai-competitive-rise
  9. [9] Chinese memory brands ditch Samsung and Micron for ... — reactive:chinese-ai-competitive-rise
  10. [10] 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)
  11. [11] 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)
  12. [12] 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)
  13. [13] Jukan @ ICML on X: "Report: Huawei Seeks AI Chip Customers in Middle East, Southeast Asia- BBG Huawei Technologies is endeavoring to establish a foothold in the AI chip markets of the Middle East and Southeast Asia, areas currently dominated by Nvidia Corp. These efforts are ongoing despite https://t.co/0IrpdJH1K9" / X — reactive:chinese-ai-competitive-rise
  14. [14] 🇨🇳China claims a new milestone in locally trained AI, as Meituan rolls out LongCat-2.0. — Rohan Paul Twitter (2026-06-30)
  15. [15] Facing US export controls, China's DeepSeek plans to make its own chips — Ars Technica AI (2026-07-07)
  16. [16] 🟡 The sobering of AI — Semafor Technology (2026-07-08)
  17. [17] DeepSeek is reportedly designing its own AI chip to ... — reactive:chinese-ai-competitive-rise
  18. [18] Reuters: DeepSeek Is Developing Its Own AI Chips - Engadget — reactive:chinese-ai-competitive-rise
  19. [19] DeepSeek Is Building Its Own AI Chip — But Designing Silicon Was Never the Bottleneck - Hardware Busters — reactive:chinese-ai-competitive-rise
  20. [20] U.S. chip restrictions helped push China to build and spread open AI models. — Rohan Paul Twitter (2026-07-01)
  21. [21] 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)
  22. [22] Jensen Huang explains how blocking China from Nvidia does not mean blocking China from AI. — Rohan Paul Twitter (2026-06-29)
  23. [23] Aravind Srinivas just explained why China’s open-source AI may become more powerful than ever. — Rohan Paul Twitter (2026-06-30)
  24. [24] Overly Stringent Export Controls Chip Away at American AI Leadership | Blogs | May 5, 2025 | ITIF — reactive:chinese-ai-competitive-rise
  25. [25] Dario Amodei has a really hardline view that China shouldn’t have strong AI. — Rohan Paul Twitter (2026-06-29)
  26. [26] Opinion from a former Meta PM. — Rohan Paul Twitter (2026-06-30)
  27. [27] 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)
  28. [28] GLM-5.2 is the step change for open agents — Interconnects (2026-06-22)
  29. [29] GLM-5.2 Is The New Best Open Model — Zvi's AI Roundups (2026-06-22)
  30. [30] China’s CXMT Is Set to Challenge DRAM Incumbents — SemiAnalysis Twitter (2026-06-23)
  31. [31] 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)
  32. [32] 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)
  33. [33] 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)
  34. [34] The Information reports that Anthropic’s Mythos preview spooked DeepSeek into fundraising. — Rohan Paul Twitter (2026-06-26)
  35. [35] 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)
  36. [36] ASML denies US government report that its EUV chipmaking tool ... — reactive:chinese-ai-competitive-rise
  37. [37] 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)
  38. [38] GLM-5.2 is probably the most powerful text-only open weights LLM — Simon Willison (2026-06-17)
  39. [39] 🇨🇳🇺🇸Chinese AI models are up to 50 times cheaper than their American counterparts on a per-token basis. — Rohan Paul Twitter (2026-06-27)
  40. [40] The world's fastest supercomputer just went live in Shenzhen — and it contains zero US GPUs. — reactive:chinese-ai-competitive-rise (2026-06-28)
  41. [41] DeepSeek is reportedly designing its own AI chip to... — reactive:chinese-ai-competitive-rise