The Information Machine

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

Version 7

2026-05-26 09:39 UTC · 221 items

What

The US Congress has expanded its investigations of Chinese AI adoption from consumer apps to critical infrastructure, while simultaneously probing DeepSeek's CCP ties and funding structure [4][1][3]. DeepSeek's funding picture is consolidating: Tencent is reported to lead a $4 billion round valuing DeepSeek at $50 billion [9][11], distinct from an earlier-reported $300 million raise at a $10 billion valuation [12], with multiple outlets now confirming the $45–50 billion range [10][11]. A frontier model from Zhipu AI (GLM-5) was reportedly trained entirely on Huawei Ascend hardware, bypassing NVIDIA [18] — the most concrete claimed evidence yet that Chinese chips can handle frontier-scale training. China's 'Four Little Dragons' AI startups collectively exceed $1 trillion in combined valuation [17], indicating the competitive pressure extends far beyond DeepSeek alone.

Why it matters

The combination of Tencent's reported entry into DeepSeek at a $50 billion valuation and GLM-5's claimed Huawei-only training represents two simultaneous stress tests of Western assumptions: that DeepSeek is commercially independent rather than strategically backed by China's largest tech conglomerates, and that Huawei chips are adequate only for inference rather than frontier training. If either assumption proves wrong, the current US policy toolkit — export controls on chips, scrutiny of open-source weights — requires significant rethinking.

Open questions

  • Tencent is now reported to lead a $4B round valuing DeepSeek at $50B [9][11], while a separate $300M raise at $10B has also been reported [12][13]. Do these represent different share classes, different tranches of the same round, or conflicting reports — and does Tencent's strategic involvement change the Congressional calculus on treating DeepSeek's open-source weights as commercially neutral technology?

  • GLM-5 (Zhipu AI) was reportedly trained entirely on Huawei Ascend hardware without NVIDIA [18]. If verified, does this directly invalidate Castellano's 'limited threat' analysis [19] and the inference-only framing used to justify current export control strategy?

  • The Congressional probe now explicitly covers PRC AI models deployed in critical infrastructure [3]. Which sectors are under review, and does this trigger existing critical infrastructure law or require new legislative authority distinct from mechanisms used for consumer-app restrictions like the TikTok ban?

  • An analyst piece argues DeepSeek's valuation requires 'throwing out the model company playbook' [16]. If DeepSeek's value is structurally different from Western model companies — due to open-source strategy, Chinese market position, and state adjacency — how should Western policymakers classify and regulate it?

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]. Where Airbnb's case centered on consumer data exposure through a Chinese AI API, Cursor is embedded in the development workflows of software engineers at US companies and government contractors, raising supply-chain concerns about code and intellectual property. A separate lawmakers' inquiry has explicitly extended the framing to PRC-origin AI models deployed in critical infrastructure [3] — a designation that carries distinct legal weight and implicates sectors such as energy, water, and defense. House investigators have also opened a direct probe of DeepSeek's CCP ties and funding structure [4][5], establishing bipartisan, bicameral pressure on the executive branch. Adding a commercial wrinkle, SpaceX has secured an option to acquire Cursor at a $60 billion valuation [6][7][8], placing the investigation's highest-profile developer-tool target under Elon Musk's ownership.

DeepSeek's funding structure, which was a source of analytical confusion, is coming into sharper focus — while introducing new complications. Tencent, China's largest technology conglomerate, is reported to lead a $4 billion round valuing DeepSeek at $50 billion [9], with Reuters and TechCrunch independently confirming the $45–50 billion valuation range [10][11]. This sits alongside a separately reported $300 million raise at approximately $10 billion [12][13] — the two figures may represent different share classes or sequential tranches. Crucially, Tencent's entry means DeepSeek's backing is no longer confined to government-linked capital: it now includes one of China's most powerful private-sector technology players. DeepSeek's founder has publicly declared an AGI goal [14], and the company has made permanent a 75% reduction in its API pricing [15], while analysts argue its valuation defies standard model-company frameworks because of its open-source posture and structural proximity to the Chinese state [16]. China's 'Four Little Dragons' AI startups — including Zhipu AI — collectively exceed $1 trillion in combined valuation [17], signaling that the competitive dynamic extends well beyond DeepSeek.

The hardware competition has a new and consequential data point. Zhipu AI's GLM-5 was reportedly trained entirely on Huawei Ascend hardware, without any NVIDIA involvement [18]. This is the most concrete claimed evidence that Chinese chips can support not merely inference workloads but frontier-scale training — the precise capability that Dr. Robert Castellano's analysis argued Huawei lacked [19], and the capability that US export controls were specifically designed to deny. Whether the claim survives independent technical scrutiny remains open; the item carries no published methodology. But even as a contested claim, it shifts the burden: the inference-only framing that has anchored the 'limited threat' analysis now has a named counterexample to explain away. Chinese open-source models natively optimized for Huawei Ascend hardware have surged [20], and Hugging Face data shows Chinese models overtook US-origin models in download share [21] — suggesting that even at inference scale, Huawei hardware is proving adequate for the deployment patterns China's diffusion strategy actually requires.

China's state-directed AI strategy continues to build regulatory and institutional infrastructure. Three major regulatory bodies issued joint AI agent guidelines in May 2026 [22][23], with sector-specific implications for healthcare liability and clinical safety standards [24]. China's NDRC has instructed domestic LLMs to promote domestic AI tools [25], and generative AI patent examination reforms that took effect in 2026 raise quality standards alongside volume [26][27] — an effort to address the longstanding critique that China's patent lead is a quantity artifact rather than a genuine innovation indicator. Stanford's 2026 AI Index places the aggregate picture in relief: the US spends 23 times more than China on AI [28] and maintains a 2.7% performance lead on top benchmarks [29] — a resource asymmetry that has not yet translated into a proportional capability gap.

Timeline

  • 2025-02-05: TrendForce reports Huawei Ascend 910C reaches 60% of NVIDIA H100 inference performance [44][40]
  • 2026-01-01: China's raised bar for AI patent quality takes effect, addressing volume-vs-quality criticisms of China's patent lead [45]
  • 2026-03-09: China declared world's largest holder of AI patents [43]
  • 2026-04-07: Harvard Business School working paper on China's 'Diffusion-Forward AI Strategy' published, arguing China deliberately prioritizes deployment scale over frontier benchmarks [32]
  • 2026-04-14: Stanford 2026 AI Index published: China leads in publications and patents; US AI performance lead quantified at 2.7%; US spends 23x more than China on AI [33][34][29][28][35]
  • 2026-04-17: DeepSeek founder declares AGI goal ahead of fundraise; Reuters reports $300M raise at ~$10B valuation [14][46][38][12]
  • 2026-04-22: SpaceX partners with Cursor on AI training and secures option to acquire Cursor for $60 billion [47][6][7][8]
  • 2026-04-30: SenseTime fully open-sources SenseNova U1 with VAE-free NEO-Unify architecture for unified multimodal understanding and generation [48][49][50]
  • 2026-05-06: Tencent reported to lead $4B DeepSeek round at $50B valuation; Reuters and TechCrunch confirm $45–50B valuation range [9][10][11][51]
  • 2026-05-08: Three of China's most powerful regulatory bodies release joint guidelines to regulate AI agents, with sector-specific healthcare implications [52][53][54][24][22][23]
  • 2026-05-21: Five Chinese AI labs confirmed to have released open-source frontier models; Hugging Face Spring 2026 report signals Chinese models overtook US in download share [55][21][56]
  • 2026-05-22: GLM-5 (Zhipu AI) reported trained entirely on Huawei Ascend hardware, bypassing NVIDIA; coordinated wave of Chinese open-source models optimized for Ascend also reported [18][20]
  • 2026-05-23: House Republicans expand investigation to Airbnb and Anysphere (Cursor) for Chinese AI model use; House separately probes DeepSeek's CCP ties; WSJ reports Chinese government investing in DeepSeek at $50B valuation; China's NDRC instructs domestic LLMs to promote domestic AI tools [30][31][4][1][5][2][25]
  • 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; Lexology publishes practitioner analysis of China's 2026 AI patent examination reform [15][3][19][26][27]
  • 2026-05-26: China's 'Four Little Dragons' AI startups collectively exceed $1 trillion in combined valuation; analytical piece argues DeepSeek valuation requires a framework distinct from standard model companies [17][16]

Perspectives

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

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

Evolution: Expanded from Airbnb-only to include Anysphere, pivoted from scrutinizing US adopters to investigating Chinese companies directly, and extended the framing from consumer data to critical infrastructure — each step broadening scope and legal implications.

[30][31][4][1][5][2][3]

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, implying US benchmark-centric comparisons systematically undercount China's strategic progress.

Evolution: Consistent; China's AI agent guidelines, healthcare regulatory framework, and GLM-5's Huawei-only training claim all provide concrete evidence validating this framing.

[32]

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.

[33][34][29][28][35]

Brian Chesky (Airbnb CEO)

Directly defends Silicon Valley's use of Chinese open-source AI, arguing the US is 'misunderstanding' what open-source model adoption means for data security and denying that using Chinese model weights sends user data to China.

Evolution: Consistent; the only major US tech executive to publicly push back against Congressional pressure.

[36][37]

DeepSeek / Liang Wenfeng (founder)

Publicly declares an AGI goal ahead of a major fundraise, positioning DeepSeek as a mission-driven organization; has permanently cut API prices by 75%, intensifying cost pressure on Western AI labs.

Evolution: Consistent on mission framing; funding picture now includes Tencent as a reported major backer at $50B, complicating the independence narrative.

[14][38][15][9]

Dr. Robert Castellano (analyst)

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

Evolution: Under pressure from GLM-5's reported Huawei-only frontier training [18], which, if verified, directly challenges the inference-only framing that anchors his analysis.

[19][18]

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 and safety standards — framing Chinese AI governance as an active policy instrument.

Evolution: Consistent; each regulatory layer (agent guidelines, healthcare focus, patent reform) reinforces state-direction narrative.

[39][22][23][26]

GenAI Assembling / independent analysts

Argues DeepSeek's valuation cannot be assessed using standard model-company frameworks because its open-source posture, Chinese market position, and state adjacency create a structurally different entity.

Evolution: New voice this pass; adds an analytical layer to the DeepSeek independence debate.

[16]

Tensions

  • GLM-5 (Zhipu AI) reportedly trained entirely on Huawei Ascend without NVIDIA [18] directly challenges Dr. Robert Castellano's 'limited threat' analysis [19] and the inference-only framing used to justify current US export controls — a dispute whose resolution is central to whether chip restrictions can actually constrain China's frontier AI development. [18][19][40]
  • Brian Chesky's 'misunderstanding' defense of Chinese open-source AI use [36] directly clashes with House Republicans' data security framing [1][2], which has expanded to Anysphere (Cursor) and 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 [14], but Tencent's reported lead in a $4B round at $50B [9] places a Chinese tech giant at the center of its cap table, alongside earlier government-linked capital [41] — complicating any claim that open-source DeepSeek weights are commercially neutral technology. [14][9][41][4]
  • Stanford AI Index's 2.7% US performance lead [29] directly complicates the China-has-achieved-parity narrative, while the 23x US spending advantage [28] adds an unresolved second axis: either China is far ahead on efficiency, or the gap will widen as resource asymmetry compounds. [29][28][34]
  • China's generative AI patent examination reform [27] — raising quality standards — directly challenges the widely-cited critique (anchored in NBER analysis [42]) that China's patent lead is a volume artifact rather than a genuine innovation indicator. [27][42][43]
  • The $300M raise at ~$10B [12] and the Tencent-led $4B round at $50B [9][11] remain publicly unreconciled — they may be different tranches or conflicting reports, but the five-fold valuation gap means the answer materially changes assessments of DeepSeek's independence, capitalization, and strategic intent. [12][9][11][10]

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] House investigation into DeepSeek teases out funding, security realities around Chinese AI tool | CyberScoop — reactive:china-ai-rising
  6. [6] SpaceX Cursor Deal: $60B AI Buy Option | The Silicon Review — reactive:china-ai-rising
  7. [7] SpaceX partners with Cursor on AI training, floats potential $60B ... — reactive:china-ai-rising
  8. [8] SpaceX secures option to buy AI coding startup Cursor for $60B — reactive:china-ai-rising
  9. [9] Tencent to back DeepSeek in $4B round at $50B valuation, marking first external funding: report — TFN — reactive:china-ai-rising
  10. [10] DeepSeek could hit $45B valuation from its first investment round — reactive:china-ai-rising
  11. [11] DeepSeek could be valued at up to $50 billion in first fundraising ... — reactive:china-ai-rising
  12. [12] DeepSeek $300M Raise at $10B Valuation [2026] — reactive:china-ai-rising
  13. [13] DeepSeek's $10 Billion Round Could Redefine China's AI Race — reactive:china-ai-rising
  14. [14] DeepSeek Founder Avows AGI Goal Ahead of $10 Billion Funding — reactive:china-ai-rising
  15. [15] Chinese artificial intelligence startup DeepSeek will make ... — reactive:china-ai-rising
  16. [16] How to Value DeepSeek: Throw Out the Model Company Playbook — reactive:china-ai-rising
  17. [17] AI's "Four Little Dragons" Boast Valuation Exceeding One Trillion — reactive:china-ai-rising
  18. [18] GLM-5: The Frontier Model That Ditched NVIDIA Entirely — Swfte AI — reactive:china-ai-rising
  19. [19] Huawei’s Latest AI Chips: A Deep-Dive Technical Analysis Reveals Limited Threat to Nvidia — reactive:china-ai-rising
  20. [20] 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)
  21. [21] One year after DeepSeek, Chinese AI models spread rapidly, overtake U.S. in download share — reactive:china-ai-rising
  22. [22] China Issues Guidelines to Standardize AI Agent Development - Caixin Global — reactive:china-ai-rising
  23. [23] China issues guidelines to regulate, promote AI agents — reactive:china-ai-rising
  24. [24] China’s new AI agent framework puts healthcare in the regulatory spotlight | HealthTechAsia — reactive:china-ai-rising
  25. [25] China to Invest in DeepSeek at $50 Billion Valuation - WSJ — reactive:china-ai-rising
  26. [26] 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
  27. [27] Strategic Shift in China's AI Patent Examination: Navigating the 2026 ... — reactive:china-ai-rising
  28. [28] Stanford's 2026 AI index just dropped: the US spends 23x more than ... — reactive:china-ai-rising
  29. [29] Stanford AI Index 2026: China narrows US lead to 2.7% while ... - TNW — reactive:china-ai-rising
  30. [30] 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)
  31. [31] 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)
  32. [32] [PDF] China's Diffusion-Forward AI Strategy - Harvard Business School — reactive:china-ai-rising
  33. [33] China leads in AI publications, patents: Stanford report - CGTN — reactive:china-ai-rising
  34. [34] @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)
  35. [35] 20 Takeaways from Stanford’s 2026 AI Index Report — 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] Airbnb's Chesky Says US 'Misunderstanding' Use of Chinese Open-Source AI Models — reactive:china-ai-rising (2026-05-23)
  38. [38] DeepSeek Founder Declares AGI Goal as $10 Billion Round Advances - Bloomberg — reactive:china-ai-rising
  39. [39] China’s National Development and Reform Commission (NDRC) has reportedly instructed domestic large language models to ac... — reactive:china-ai-rising (2026-05-23)
  40. [40] DeepSeek research suggests Huawei's Ascend 910C delivers 60 ... — reactive:china-ai-rising
  41. [41] DeepSeek value could be up to $50 billion in first fundraising ... — reactive:china-ai-rising
  42. [42] [PDF] AI Patents in the United States and China: Measurement ... - CDN — reactive:china-ai-rising
  43. [43] China becomes world's largest holder of AI patents - People's Daily — reactive:china-ai-rising
  44. [44] [News] DeepSeek Reportedly Reveals Huawei’s Ascend 910C Reaches 60% of NVIDIA H100’s Inference Power — reactive:china-ai-rising
  45. [45] China Raises the Bar for AI Patents: What Changes from 1 January 2026 - Mathys & Squire LLP — reactive:china-ai-rising
  46. [46] DeepSeek's Potential $10B+ Valuation, LinkedIn's AI Agent Win and ... — reactive:china-ai-rising
  47. [47] SpaceX Secures $60 Billion Option to Acquire AI Coding Giant Cursor — reactive:china-ai-rising
  48. [48] 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
  49. [49] SenseTime's SenseNova U1 ditches VAEs entirely to unify image generation and understanding - Startup Fortune — reactive:china-ai-rising
  50. [50] SenseTime Fully Open-Sources SenseNova U1: A Unified Model for Understanding and Generation-News and Blog-SenseTime — reactive:china-ai-rising
  51. [51] DeepSeek reportedly seeks first funding round at $45 billion valuation · TechNode — reactive:china-ai-rising
  52. [52] China unveils guidelines to regulate, boost innovative development ... — reactive:china-ai-rising
  53. [53] On 8 May 2026, three of China's most powerful regulatory bodies ... — reactive:china-ai-rising
  54. [54] WE ARE CHINA — reactive:china-ai-rising
  55. [55] 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)
  56. [56] State of Open Source on Hugging Face: Spring 2026 — reactive:china-ai-rising