The Information Machine

China's AI Ecosystem Gaining Ground on the West · history

Version 8

2026-05-27 18:27 UTC · 240 items

What

The US-China AI competition has entered a phase of simultaneous policy escalation and technical documentation. GLM-5, Zhipu AI's 744-billion-parameter model, is now described in an arXiv paper [11] and Zhipu's own technical blog [12] as trained entirely on Huawei Ascend hardware — moving the claim from news report to primary-source methodology, though independent verification remains absent [14]. Chinese AI startup funding reached $16.2B in Q1 2026 alone, up 185% year-over-year [21], signaling ecosystem-wide acceleration. DeepSeek's investor base may extend to both Tencent and Alibaba [7][8], with the current leading figure a Tencent-led $4B round at a $50B valuation [5][6].

Why it matters

If a 744B-parameter model can be trained end-to-end on Huawei hardware, the chip export control strategy — designed specifically to deny China frontier training capability — faces its most concrete empirical challenge to date. The $16.2B Q1 2026 funding surge [21] and Tencent's pledge to increase AI investment after chip curbs hit its capex plans [22] suggest restrictions are catalyzing rather than constraining Chinese AI development.

Open questions

  • GLM-5's arXiv paper [11] provides architecture and training details for a 744B model on Huawei Ascend, but independent benchmark verification against compute-matched NVIDIA systems remains absent [14]. Does this level of primary-source documentation meet the evidentiary bar for policymakers to revise the inference-only framing that anchors current US export controls?

  • Chinese AI startups raised $16.2B in Q1 2026, up 185% YoY [21], and Tencent pledged higher AI investment after chip curbs [22]. Does this pattern indicate export restrictions are catalyzing rather than deterring domestic capital formation — and if so, does the current policy toolkit require fundamental rethinking?

  • Reports place both Tencent and Alibaba as potential DeepSeek investors [7][8] while the current leading figure is Tencent-led at $50B [5]. If both of China's largest tech conglomerates hold stakes, does Congressional treatment of DeepSeek's open-source weights as commercially neutral technology remain defensible?

  • The Q2 2026 Chinese AI landscape spans at least 10 significant providers [26], with the leading startups — variously framed as 'Four Little Dragons' [23] or 'Six Tigers' [24] — collectively exceeding $1 trillion in valuation [25]. How is market share distributed in the enterprise and critical-infrastructure segments where US regulatory attention is sharpest?

Narrative

The US political response to China's AI rise has moved from abstract export-control debates to specific named investigations. House Republicans, through the Select Committee on China and the House Homeland Security Committee, opened a joint investigation into Airbnb and Anysphere, the company behind the Cursor AI coding assistant [1][2], citing concerns about Chinese AI model use in developer supply chains. A separate inquiry extended the framing explicitly to PRC-origin AI models deployed in critical infrastructure [3], a designation carrying distinct legal weight and implicating sectors including energy, water, and defense. House investigators simultaneously opened a probe of DeepSeek's CCP ties and funding structure [4]. DeepSeek's investor base is sharpening in focus: Tencent is reported to lead a $4 billion round at a $50 billion valuation [5][6], while earlier reports placed both Tencent and Alibaba as investors at a $20 billion valuation [7][8] — figures that may represent different tranches or sequential rounds. If both of China's two largest tech conglomerates hold stakes, the premise that DeepSeek's open-source weights are commercially neutral technology becomes difficult to sustain. DeepSeek's founder has publicly declared an AGI goal [9] and permanently cut API prices by 75% [10], sustaining cost pressure on Western labs.

The hardware competition has acquired its most substantiated data point to date. Zhipu AI's GLM-5 — a 744-billion-parameter model — is described in an arXiv paper [11] and the company's own technical blog [12] as trained entirely on Huawei Ascend hardware without NVIDIA involvement. This advances the claim from journalist report [13] to primary-source documentation with architecture and training methodology, though the Reddit ML community has explicitly questioned its accuracy [14] and independent compute-matched benchmarks remain absent. Multiple third-party technical reviewers have published detailed analyses [15][16][17]. The policy significance is direct: Dr. Robert Castellano's "limited threat" analysis [18] argued that Huawei chips serve inference workloads but cannot support frontier-scale training — the precise capability US export controls were designed to deny. A 744B model trained end-to-end on Huawei hardware, if verified, invalidates that distinction. Chinese open-source models optimized natively for Huawei Ascend have surged [19], and Hugging Face data shows Chinese models overtook US-origin models in download share [20], suggesting Huawei hardware is already proving adequate for the deployment patterns China's AI strategy actually requires.

The competitive pressure extends across the entire Chinese AI ecosystem. Startups raised $16.2 billion in Q1 2026 alone, up 185% year-over-year [21]. Tencent, navigating US chip export restrictions, has pledged to increase AI investment in 2026 [22] — a signal that restrictions may be accelerating rather than deterring domestic capital flows. China's leading AI startups are variously framed as the "Four Little Dragons" (Baichuan, Zhipu AI, Moonshot AI, MiniMax) [23] or "Six Tigers" [24], with collective valuations exceeding $1 trillion [25]. A Q2 2026 market landscape report covers at least 10 significant providers [26]. Stanford's 2026 AI Index places this in comparative context: the US spends 23 times more than China on AI [27] and maintains a 2.7% performance lead on top benchmarks [28] — a resource asymmetry that has not translated into a proportional capability gap, and that Harvard Business School's working paper attributes to China's deliberate "Diffusion-Forward" strategy of prioritizing deployment scale over benchmark leadership [29].

China's state-directed AI governance is building institutional infrastructure in parallel with commercial momentum. Three major regulatory bodies issued joint AI agent guidelines in May 2026 [30][31], with sector-specific implications for healthcare liability and clinical safety [32]. China's NDRC has instructed domestic LLMs to promote domestic AI tools [33], and generative AI patent examination reforms effective in 2026 raise quality standards alongside volume [34][35] — an effort to address the critique that China's patent lead is a quantity artifact. Airbnb CEO Brian Chesky has publicly pushed back against Congressional pressure, arguing the US is misunderstanding what using Chinese open-source model weights means for data security [36] — the only major US tech executive to directly contest the national security framing.

Timeline

  • 2025-02-05: TrendForce reports Huawei Ascend 910C reaches 60% of NVIDIA H100 inference performance [48][49]
  • 2026-01-01: China's raised bar for generative AI patent quality takes effect, addressing volume-vs-quality criticisms of China's patent lead [50]
  • 2026-02-01: Zhipu AI submits GLM-5 arXiv paper describing a 744B model trained entirely on Huawei Ascend hardware, providing primary-source technical methodology [11]
  • 2026-03-09: China declared world's largest holder of AI patents [47]
  • 2026-04-07: Harvard Business School working paper argues China deliberately pursues a 'Diffusion-Forward AI Strategy,' prioritizing deployment scale over frontier benchmarks [29]
  • 2026-04-14: Stanford 2026 AI Index published: China leads in publications and patents; US AI performance lead is 2.7%; US spends 23x more than China [42][43][28][27][44]
  • 2026-04-17: DeepSeek founder declares AGI goal; Reuters reports $300M raise at ~$10B valuation [9][51][41][52]
  • 2026-04-22: SpaceX secures option to acquire Cursor (Anysphere) for $60 billion [53][54][55]
  • 2026-04-30: SenseTime fully open-sources SenseNova U1 with VAE-free architecture for unified multimodal understanding and generation [56][57][58]
  • 2026-04-01: Zero2IPO reports Chinese AI startups raised $16.2B in Q1 2026, up 185% year-over-year [21]
  • 2026-05-06: Tencent reported to lead $4B DeepSeek round at $50B valuation; Reuters and TechCrunch confirm $45–50B valuation range [5][59][6][60]
  • 2026-05-08: Three of China's most powerful regulatory bodies release joint guidelines on AI agents, with sector-specific healthcare implications [61][62][63][32][30][31]
  • 2026-05-21: Five Chinese AI labs confirmed to have released open-source frontier models; Hugging Face Spring 2026 report shows Chinese models overtook US in download share [64][20][65]
  • 2026-05-22: GLM-5 Huawei-only training claim goes mainstream; wave of Chinese open-source models optimized for Ascend reported [13][19]
  • 2026-05-23: House Republicans expand investigation to Airbnb and Anysphere (Cursor) for Chinese AI model use; House separately probes DeepSeek's CCP ties; China's NDRC instructs LLMs to promote domestic AI tools [37][38][4][1][39][2][33]
  • 2026-05-25: DeepSeek makes permanent a 75% price reduction; lawmakers' inquiry explicitly extended to PRC AI models in critical infrastructure; Castellano analysis contests Huawei chip performance claims [10][3][18][34][35]
  • 2026-05-26: China's 'Four Little Dragons' AI startups exceed $1 trillion in combined valuation; analytical piece argues DeepSeek requires a framework distinct from standard model companies [25][66]

Perspectives

US House Republicans (Select Committee on China / House Homeland Security)

Frame Chinese AI model adoption as a data security and national security threat; conducting investigations of Airbnb and Anysphere (Cursor) while separately probing DeepSeek's CCP ties; extended the inquiry explicitly to PRC AI models in critical infrastructure.

Evolution: Expanded from consumer data concerns to developer supply chains and critical infrastructure over successive probe announcements — each step broadening legislative and regulatory surface.

[37][38][4][1][39][2][3]

Brian Chesky (Airbnb CEO)

Directly defends Silicon Valley's use of Chinese open-source AI, arguing the US is 'misunderstanding' open-source model adoption and that using Chinese weights does not send user data to China.

Evolution: Consistent; the only major US tech executive to publicly contest the Congressional national security framing.

[36][40]

DeepSeek / Liang Wenfeng (founder)

Publicly declares an AGI goal and positions DeepSeek as a mission-driven organization; permanently cut API prices by 75%, intensifying cost pressure on Western AI labs.

Evolution: Funding picture now implicates both Tencent as reported lead investor at $50B and potentially Alibaba as well, complicating the commercial independence narrative.

[9][41][10][5][7][8]

Dr. Robert Castellano (analyst)

Technical counter-analysis arguing Huawei's AI chips represent a 'limited threat' to NVIDIA, contesting claims that Ascend 910C approaches H100 performance in demanding workloads.

Evolution: Under increasing pressure: GLM-5's arXiv paper and Zhipu's own blog now provide primary-source documentation of a 744B model trained on Huawei hardware, directly challenging the inference-only framing that anchors his analysis.

[18][13][11][12]

Harvard Business School (working paper)

Frames China's approach as a deliberate 'Diffusion-Forward AI Strategy' — prioritizing broad deployment and application-layer scale over frontier benchmark leadership.

Evolution: Consistent; GLM-5's Huawei-only training documentation, Chinese models' download share overtaking US, and the Q1 2026 funding surge all provide further evidence validating this framing.

[29]

Stanford HAI / 2026 AI Index

Authoritative cross-country benchmark: US leads China in AI performance at a 2.7% margin; China leads in publications and patents; US spends 23x more on AI.

Evolution: Consistent; the 23x spending asymmetry alongside a narrow performance gap remains the central unresolved empirical puzzle.

[42][43][28][27][44]

China state direction (NDRC / regulatory bodies)

NDRC instructed domestic LLMs to promote domestic AI tools; three major regulatory bodies jointly issued AI agent guidelines with sector-specific liability standards; patent examination reforms raise the quality bar.

Evolution: Consistent; each new regulatory layer reinforces a state-direction narrative that complements rather than constrains commercial momentum.

[45][30][31][34][33]

Chinese AI ecosystem (Tencent / startups broadly)

Tencent pledged to increase AI investment in 2026 after chip curbs hit capex plans and is reported as DeepSeek's lead investor; Chinese AI startups raised $16.2B in Q1 2026 alone, up 185% YoY.

Evolution: New consolidated voice this pass: the funding surge and Tencent's investment escalation suggest chip restrictions are accelerating rather than deterring Chinese AI capital formation.

[22][21][5]

Tensions

  • GLM-5's arXiv paper [11] and Zhipu's own blog [12] now provide primary-source documentation for a 744B model trained entirely on Huawei Ascend — directly challenging Castellano's 'limited threat' analysis [18] and the inference-only framing anchoring US export controls, while the Reddit ML community publicly questions the claim [14] and independent verification is absent. [11][12][18][13][14]
  • Brian Chesky's 'misunderstanding' defense of Chinese open-source AI use [36] directly clashes with House Republicans' data security framing [1][2], extended explicitly to critical infrastructure [3] — neither side has engaged the other's specific technical claim. [36][1][2][3]
  • DeepSeek presents itself as an independent commercial lab pursuing AGI [9], but its investor base now reportedly includes Tencent leading a $50B round [5] and potentially Alibaba as well [7] — leaving the premise that its open-source weights are commercially neutral technology increasingly difficult to sustain. [9][5][7][8][4]
  • Stanford AI Index's 2.7% US performance lead [28] directly complicates the China-has-achieved-parity narrative, while the 23x US spending advantage [27] adds an unresolved second axis: either China is far ahead on efficiency, or the gap will widen as resource asymmetry compounds. [28][27][43]
  • Tencent's pledge to increase AI investment after chip curbs [22] and the $16.2B Q1 2026 startup funding surge [21] suggest export controls are catalyzing rather than constraining Chinese AI capital formation — a direct challenge to the policy premise that hardware restrictions slow Chinese AI development. [22][21]
  • China's generative AI patent examination reform [35] — raising quality standards — directly challenges the widely-cited critique that China's patent lead is a volume artifact rather than a genuine innovation indicator [46]. [35][46][47]

Sources

  1. [1] Chairmen Moolenaar, Garbarino Announce Joint Investigation ... — reactive:china-ai-rising
  2. [2] House panels probe Airbnb, Anysphere over use of Chinese AI models - Nextgov/FCW — reactive:china-ai-rising
  3. [3] Lawmakers open inquiry into cybersecurity risks posed by PRC ... — reactive:china-ai-rising
  4. [4] House Republicans Seek Info On DeepSeek Ties To CCP - Law360 — reactive:china-ai-rising
  5. [5] Tencent to back DeepSeek in $4B round at $50B valuation, marking first external funding: report — TFN — reactive:china-ai-rising
  6. [6] DeepSeek could be valued at up to $50 billion in first fundraising ... — reactive:china-ai-rising
  7. [7] Tencent and Alibaba in Talks to Invest in DeepSeek AI Startup — reactive:china-ai-rising
  8. [8] Tencent Alibaba Invest in DeepSeek at $20B Valuation - LinkedIn — reactive:china-ai-rising
  9. [9] DeepSeek Founder Avows AGI Goal Ahead of $10 Billion Funding — reactive:china-ai-rising
  10. [10] Chinese artificial intelligence startup DeepSeek will make ... — reactive:china-ai-rising
  11. [11] GLM-5: from Vibe Coding to Agentic Engineering - arXiv — reactive:china-ai-rising
  12. [12] GLM-5: From Vibe Coding to Agentic Engineering - Z.ai — reactive:china-ai-rising
  13. [13] GLM-5: The Frontier Model That Ditched NVIDIA Entirely — Swfte AI — reactive:china-ai-rising
  14. [14] Is this true? GLM 5 was trained solely using huawei hardware and ... — reactive:china-ai-rising
  15. [15] How China's GLM-5 Works: 744B Model on Huawei Chips | Let's Data Science — reactive:china-ai-rising
  16. [16] GLM-5 Review: 745B Open-Weight Frontier, MIT (2026) | Local AI Master — reactive:china-ai-rising
  17. [17] GLM-5 2026 Architecture, Benchmarks & Limitations — reactive:china-ai-rising
  18. [18] Huawei’s Latest AI Chips: A Deep-Dive Technical Analysis Reveals Limited Threat to Nvidia — reactive:china-ai-rising
  19. [19] The defining trend of this window is a massive, coordinated wave of open-source models natively optimized for Huawei's A... — reactive:china-ai-rising (2026-05-22)
  20. [20] One year after DeepSeek, Chinese AI models spread rapidly, overtake U.S. in download share — reactive:china-ai-rising
  21. [21] Zero2IPO Research: Chinese AI startups raised $16.2B in Q1 2026 ... — reactive:china-ai-rising
  22. [22] Tencent pledges higher AI investment in 2026 after chip curbs — reactive:ai-company-singapore-race
  23. [23] China’s 4 new ‘AI tigers’ – Baichuan, Zhipu AI, Moonshot AI and MiniMax – emerge as investor favourites | South China Morning Post — reactive:china-ai-rising
  24. [24] Meet the 'Six Tigers' that dominate China's AI industry - Quartz — reactive:china-ai-rising
  25. [25] AI's "Four Little Dragons" Boast Valuation Exceeding One Trillion — reactive:china-ai-rising
  26. [26] Chinese AI Models Q2 2026: 10-Provider Landscape Report — reactive:open-source-model-surge
  27. [27] Stanford's 2026 AI index just dropped: the US spends 23x more than ... — reactive:china-ai-rising
  28. [28] Stanford AI Index 2026: China narrows US lead to 2.7% while ... - TNW — reactive:china-ai-rising
  29. [29] [PDF] China's Diffusion-Forward AI Strategy - Harvard Business School — reactive:china-ai-rising
  30. [30] China Issues Guidelines to Standardize AI Agent Development - Caixin Global — reactive:china-ai-rising
  31. [31] China issues guidelines to regulate, promote AI agents — reactive:china-ai-rising
  32. [32] China’s new AI agent framework puts healthcare in the regulatory spotlight | HealthTechAsia — reactive:china-ai-rising
  33. [33] China to Invest in DeepSeek at $50 Billion Valuation - WSJ — reactive:china-ai-rising
  34. [34] Inside China’s IP Market: a Guide 2026 - Securing patent protection for generative-AI inventions amid unprecedented global investment and incoming reform to examination guidelines - IAM — reactive:china-ai-rising
  35. [35] Strategic Shift in China's AI Patent Examination: Navigating the 2026 ... — reactive:china-ai-rising
  36. [36] Airbnb's Chesky Says US 'Misunderstanding' Use of Chinese Open-Source AI Models — reactive:china-ai-rising (2026-05-23)
  37. [37] House Republicans probe Airbnb's use of Chinese AI models citing data security risks. CEO Brian Chesky says the company ... — reactive:china-ai-rising (2026-05-23)
  38. [38] As the US House probes Airbnb's use of Chinese AI models, CEO Brian Chesky says the company is not sharing data with Chi... — reactive:china-ai-rising (2026-05-23)
  39. [39] House investigation into DeepSeek teases out funding, security realities around Chinese AI tool | CyberScoop — reactive:china-ai-rising
  40. [40] Airbnb's Chesky Says US 'Misunderstanding' Use of Chinese Open-Source AI Models — reactive:china-ai-rising (2026-05-23)
  41. [41] DeepSeek Founder Declares AGI Goal as $10 Billion Round Advances - Bloomberg — reactive:china-ai-rising
  42. [42] China leads in AI publications, patents: Stanford report - CGTN — reactive:china-ai-rising
  43. [43] @WatcherGuru 👁️ The timing is wild 👀 Stanford’s 2026 AI Index just confirmed the US lead over China in AI performance ha... — reactive:china-ai-rising (2026-05-19)
  44. [44] 20 Takeaways from Stanford’s 2026 AI Index Report — reactive:china-ai-rising
  45. [45] China’s National Development and Reform Commission (NDRC) has reportedly instructed domestic large language models to ac... — reactive:china-ai-rising (2026-05-23)
  46. [46] [PDF] AI Patents in the United States and China: Measurement ... - CDN — reactive:china-ai-rising
  47. [47] China becomes world's largest holder of AI patents - People's Daily — reactive:china-ai-rising
  48. [48] [News] DeepSeek Reportedly Reveals Huawei’s Ascend 910C Reaches 60% of NVIDIA H100’s Inference Power — reactive:china-ai-rising
  49. [49] DeepSeek research suggests Huawei's Ascend 910C delivers 60 ... — reactive:china-ai-rising
  50. [50] China Raises the Bar for AI Patents: What Changes from 1 January 2026 - Mathys & Squire LLP — reactive:china-ai-rising
  51. [51] DeepSeek's Potential $10B+ Valuation, LinkedIn's AI Agent Win and ... — reactive:china-ai-rising
  52. [52] DeepSeek $300M Raise at $10B Valuation [2026] — reactive:china-ai-rising
  53. [53] SpaceX Cursor Deal: $60B AI Buy Option | The Silicon Review — reactive:china-ai-rising
  54. [54] SpaceX partners with Cursor on AI training, floats potential $60B ... — reactive:china-ai-rising
  55. [55] SpaceX secures option to buy AI coding startup Cursor for $60B — reactive:china-ai-rising
  56. [56] An open-source model, 'SenseNova U1,' capable of image generation without the need for VAEs, has been released, offering significantly faster speeds and better quality than Z-Image. - GIGAZINE — reactive:china-ai-rising
  57. [57] SenseTime's SenseNova U1 ditches VAEs entirely to unify image generation and understanding - Startup Fortune — reactive:china-ai-rising
  58. [58] SenseTime Fully Open-Sources SenseNova U1: A Unified Model for Understanding and Generation-News and Blog-SenseTime — reactive:china-ai-rising
  59. [59] DeepSeek could hit $45B valuation from its first investment round — reactive:china-ai-rising
  60. [60] DeepSeek reportedly seeks first funding round at $45 billion valuation · TechNode — reactive:china-ai-rising
  61. [61] China unveils guidelines to regulate, boost innovative development ... — reactive:china-ai-rising
  62. [62] On 8 May 2026, three of China's most powerful regulatory bodies ... — reactive:china-ai-rising
  63. [63] WE ARE CHINA — reactive:china-ai-rising
  64. [64] Five Chinese AI labs, including Alibaba's Qwen and Zhipu's GLM, have now released open source frontier models. Meanwhile... — reactive:china-ai-rising (2026-05-20)
  65. [65] State of Open Source on Hugging Face: Spring 2026 — reactive:china-ai-rising
  66. [66] How to Value DeepSeek: Throw Out the Model Company Playbook — reactive:china-ai-rising