China AI Weekly: DeepSeek's 160K Huawei Chip Megaproject

DeepSeek ordered 160,000 Huawei Ascend 950DT accelerators for a gigawatt-scale data center in Inner Mongolia — potentially the largest Huawei AI chip cluster ever assembled. The US Commerce Department pivoted export controls from physical chips to cloud GPU access, drafting rules to block Chinese firms from remotely renting Nvidia compute in third-country data centers. SMIC posted a 94.2% profit surge and TrendForce projected China’s domestic AI chip share could approach 90% in 2026. Alibaba’s Qwen3.8-Max-0902 took the CodeArena WebDev #1 spot, while Moonshot quietly filed for a Hong Kong IPO. And the CAC purge of AI-generated content hit 5.61 million items in a single enforcement wave. Here is this week’s China AI Weekly.

DeepSeek: 160K Huawei Chips, Inner Mongolia Megaproject

The Order

DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a gigawatt-scale data center in Inner Mongolia. If fulfilled, it would be the largest known Huawei AI chip cluster — a flagship bet on domestic silicon for China’s leading model lab.

The chips are designated for inference, not training. DeepSeek still trains on Nvidia accelerators; the lab’s experience training on Huawei chips has kept it reliant on Nvidia for that step, despite the 950DT being designed and marketed for training. This is a notable split: China’s domestic silicon is winning the inference battle while Nvidia still owns the training pipeline.

The Catch

Huawei cannot meet demand. Top-end memory shortages — specifically HBM — cap 950DT output at the “low hundreds-of-thousands” this year. Fulfilling DeepSeek’s order could take more than a year. Huawei is also balancing other clients and small overseas exports.

US officials have alleged, without verification, that DeepSeek procured Nvidia Blackwell chips for the same Inner Mongolia site. DeepSeek is separately in talks to raise billions for infrastructure; the data center would draw power equivalent to roughly 750,000 homes.

Why It Matters

This order is the clearest signal yet of domestic AI stack consolidation. China’s leading model lab is betting its inference future on Huawei silicon — not because it wants to, but because the economics and geopolitics of Nvidia dependence are shifting. If Huawei can solve the HBM bottleneck, orders like this become routine. If it cannot, China’s inference scale-up hits a wall.

US Export Controls: From Chips to Cloud

The Pivot

The US Commerce Department’s Bureau of Industry and Security (BIS) is drafting rules to block Chinese firms from remotely renting Nvidia GPU compute in third-country data centers — Thailand, Singapore, Malaysia, and Japan. This is a first: targeting access, not physical chip movements.

The legal foundation is shaky. Export-control lawyers broadly agree that BIS lacks statutory authority over remote access under the Export Administration Regulations (EAR). The Remote Access Security Act — passed by the House 369–22 in January 2026 — is stalled in the Senate Banking Committee. A draft for industry comment may circulate as early as September.

Nvidia Self-Policing

Nvidia is not waiting for regulators. The company has implemented a customer whitelist across Asia (Singapore, Malaysia, Japan). The Financial Times reports that more than 50% of existing Asian customers have been dropped under tightened compliance reviews. Neocloud providers are hit hardest. Nvidia employees are now conducting physical data-center site visits to verify compliance.

The Diplomatic Backdrop

At the G20 on September 2, Commerce Secretary Lutnick stated that easing export controls is “not on the agenda” for the upcoming Trump–Xi summit. Trump offered to sell certain chips to China; Beijing “didn’t take it up.”

China’s countermeasures continue: efficient open-source models, domestic silicon, and its own export controls on gallium and rare earths. The US also blacklisted chip designer Phytium over military ties this week.

The strategic question is whether remote-access controls can work without statutory backing. If BIS pushes ahead and gets challenged in court, the entire framework could collapse. If Congress acts, it changes the game. Watch the Senate Banking Committee this fall.

China’s Domestic Chip Ecosystem: Commercial Inflection

SMIC: Profit Surge, Record Margins

SMIC reported H1 2026 net profit of RMB 44.67 billion, up 94.2% year-over-year. Q3 gross margin guidance is 26–28% — a record. Utilization is above 90%, and the company is adding approximately 40,000 300mm wafers per month in capacity through 2026.

TechInsights reverse-engineering of the Huawei Kirin 9030 Pro confirmed SMIC’s N+3 process (7nm-class) is in production. However, yields on advanced nodes remain approximately 23% — Goldman estimates SMIC needs 75% yields by 2035 to shrink the 7nm-and-below supply gap from 92% to 34%.

TrendForce: Domestic Share Approaching 90%

TrendForce projects 2026 China high-end AI chip shipments up 83% year-over-year. Domestic share of the internal market could approach 90%, with Nvidia’s share falling from ~40% in 2025 to ~8% in 2026. Nvidia H200 deliveries to China now represent less than 1% of data-center revenue — down from 26.4% in 2022.

Market Shares (2025, IDC)

Ascend 950PR: Volume Shipping

The Ascend 950PR has been shipping in volume since Q1 2026: 128GB memory, 1.6TB/s throughput, 2.87x H20 inference performance per card. The 950 supernode can interconnect 8,192 chips at ~1 EFLOPS FP8 — a serious cluster-scale architecture.

Cambricon: Profitable and Scaling

Cambricon reported H1 2026 revenue of RMB 5.996 billion and net profit of RMB 2.311 billion. The SiYuan 690 flagship — a dual-die design with >700 TFLOPS FP16 and 196GB HBM3 — is in mass production.

CXMT: HBM3E Production Begins

CXMT started small-scale HBM3E production and is sampling to Alibaba T-Head and Cambricon for testing. Commercial integration is possible next year. CXMT is also mass-producing LPDDR6 with Xiaomi adoption planned. However, CXMT remains ~2 years behind SK hynix and Samsung on HBM.

IPO Wave

Enflame Technology, the last of the “Big Four” domestic GPU startups, opened STAR Market subscription in early September. The “GPU Five Little Tigers” — Moore Threads, MetaX, Biren, Iluvatar, and others — are pursuing IPOs despite losses. New entrants shipped approximately 1 million units in 2026.

Huawei H1 2026

Huawei reported net profit of RMB 23.81 billion, down 36% year-over-year, on revenue of RMB 467.82 billion (+9.6%). R&D spend increased 25% to RMB 121.38 billion — 25.9% of revenue. The profit decline reflects the cost of building a domestic AI stack under sanctions; the R&D increase shows the commitment.

The Power Constraint

AI compute demand surged 417% in Q1 2026 versus 128% supply growth. Zhipu/Z.ai, Moonshot, and Alibaba are all restricting user access amid the compute crunch. “Supernode” cluster strategies — which drive higher power density — are exacerbating the gap. Power, not chips, may become China’s next AI bottleneck.

Regulatory Roundup: CAC Crackdown and Agent Governance

CAC Content Purge

On September 3, the CAC announced the removal of 5.61 million pieces of unlawful AI-related content, action against 49,000+ accounts, and penalties for 2,400+ websites and apps. Targets included AI-fabricated false information, violent and vulgar content, AI impersonation, and minor protection violations.

Every major platform is involved: Douyin, Kuaishou, Weibo, Tencent, Baidu, Bilibili, Xiaohongshu, Zhihu, and Taobao are strengthening deepfake detection. Chatbots — Doubao, Yuanbao, Qwen, and Ernie Bot — are tightening content review. App stores operated by Huawei, Xiaomi, OPPO, and Vivo are rechecking apps.

AI Customer Service Accountability (Sep 1)

A new national standard on human-intelligent customer service collaboration took effect September 1. Companies can no longer disclaim liability for AI responses — the classic “AI answers don’t represent company position” disclaimer is dead. This is a meaningful shift: it forces companies to treat AI-generated content as their own speech, raising the bar for deployment.

MIIT AI Service Provider Action (Aug 31)

MIIT’s Document No. 414 targets a national resource pool of 2,000+ AI providers by end-2026 and 3,000+ by end-2027 — 100 per pilot province. The program encourages first-purchase, first-use, and risk-compensation mechanisms, plus token procurement and computing vouchers. Importantly, this is an application-provider development program, not a funded token-purchase mandate.

Agent Governance Stack

China’s agent regulation is now the world’s most structured:

• AI Agents framework (effective July 15): CAC + NDRC + MIIT Implementation Opinions on Standardized Application of Intelligent Agents — the first policy treating agents as a category separate from generative AI. Filing, compliance testing, and product recall required for sensitive sectors (healthcare, transportation, media, public safety). Target: 70% agent adoption in smart terminals by 2027 across 19 scenarios.
• Anthropomorphic AI measures (July 15): Companion and emotional AI services face algorithm filing, security assessments, AI-content disclosure, and addiction prevention (2-hour usage reminders). Virtual intimate relationships for minors are banned. Applies to providers with 1M+ registered users or 100K+ monthly active users.
• Ethics Review-Before-Development (April 2026, details disclosed Aug 26): A 10-department framework requiring mandatory pre-R&D ethics review for all AI organizations. High-risk categories — human-machine integration, opinion-guidance algorithms, highly autonomous decision systems — require expert re-review. Roughly 200 key AI standards have been developed.
• TC260 Ethics-Safety Guidelines 1.0 (July 1): 9 core principles; AI Safety Governance Framework v2.0 adds “derivative risks” and circuit breakers.

New September Rules

• Micro-short-drama management: classified filing and AI-generated labeling required per episode.
• Financial product marketing: AI-driven marketing of financial products banned.
• Network data security regulation: new compliance regime in effect.

International Posture

China is backing the UN IISP-AI panel, proposing a “World AI Cooperation Organization” (WAICO), resuming US bilateral AI dialogues, and actively participating in ISO/IEC and ITU-T standards bodies. The 15th Five-Year Plan (2026–2030) includes AI governance mandates with agentic AI oversight; a draft Cybercrime Law proposes monitoring bulk malicious code generation.

Company News: The Model and Capital Race

Alibaba: Qwen3.8-Max Takes CodeArena #1

Alibaba’s Qwen3.8-Max-0902 update (September 2) post-trained for coding and office work claimed #1 on CodeArena WebDev with a score of 1691 — ahead of Claude Opus 5 and Kimi K3. On JobBench, it scored 64.0 versus Claude Fable 5’s 57.4. The model is a 2.4T-parameter MoE with ~95B active parameters and a 1M-token context. API pricing is $2/$6 per million tokens, with cache reads at $0.17–0.25.

The Qwen3.8-Flash-Next preview (September 1) revealed the architecture for Qwen4: 125B total / 6B active MoE plus a 51B N-gram embedding component runnable on CPU memory. It beats Qwen3.7-Plus on coding and office tasks at roughly 1/9th the training cost. Qwen4 is expected as soon as September.

Alibaba also completed an HK$80 billion (~$10.2 billion) share placement — 710 million new shares — for AI investment in chips, compute, and models. This sits on top of the RMB 380 billion (~$56 billion) three-year AI/cloud commitment, roughly half of which is already committed.

The Tmall AI subscription center launched, retailing token plans from Alibaba Cloud, Zhipu, Moonshot (Kimi), and MiniMax — an e-commerce play for AI compute.

Baidu: AI Revenue >50%, GPU Payback Cut to 2–3 Years

CFO Henry He confirmed on September 1–2 that AI revenue now exceeds 50% of total sales. AI profit margins are expected to approach search levels “in the near term.” GPU cluster payback has been cut to 2–3 years from 5–6 — driven by rising token consumption, cheaper domestic chips, and onshore funding.

Q2 2026: AI core revenue RMB 12.5B (50% of total, +25% YoY); AI cloud infrastructure RMB 7.3B (+50%); GPU cloud +283%; online marketing -19%. Baidu has invested RMB 100B+ in AI infrastructure since March 2023 and holds RMB 283.1B in cash.

Plans include spinning off Kunlunxin (AI chip unit) via independent listing, a $5 billion buyback, first-ever dividend, and a dual primary listing in Hong Kong.

Notable hire: former DeepSeek researcher Wei Haoran was named head of ERNIE multimodal. The risk is real: Ernie models are trailing open-weight rivals from Alibaba and Moonshot, and search ad revenue is declining as users shift to AI chatbots.

Tencent: WorkBuddy Open Platform

Tencent opened WorkBuddy to hardware makers, industry apps, and developers on September 2 — 100+ first-wave partners, 9 co-branded smart hardware devices. The positioning is clear: Tencent is the “entry point for work tasks” versus Alibaba’s “pricing for intelligence.”

AI efficiency office MAU hit 102 million in July (+8.2% YoY), but usage count surged 112.4% YoY — users are engaging deeper, not just more of them. Native AI office apps grew 261.1% YoY; PC clients +340.6%.

ByteDance: $30B Loan, Capex Raised to RMB 200B

ByteDance secured a $30 billion loan — Asia’s second-largest dollar debt deal this year after SoftBank’s $40B bridge. Capex was raised to RMB 200 billion for 2026 (from an initial RMB 160B). The company’s valuation is approximately RMB 4 trillion (~$600 billion) — the highest in China.

CEO Liang Rubo acknowledged in an August 6 all-hands that while Doubao is competitive on the consumer side and the Seedance video model is SOTA, the LLM gap versus overseas leaders has widened. A candid admission from China’s most valuable tech company.

Moonshot AI: Confidential HKEX Filing

Moonshot filed a confidential A1 with the Hong Kong Exchange this week, seeking approximately $3 billion with a possible Q4 2026 IPO. A final pre-IPO round is targeting a $50 billion valuation. The July raise was $3.5 billion.

Kimi K3 — 2.8T parameters, 104B active, the world’s first open model at the 3T-parameter scale with 1M context — remains the sole current model (K2.5/v1 retired). It is still the top-ranked open-weight model on Artificial Analysis (Intelligence Index 60), though Alibaba’s Qwen3.8-Max-0902 update has passed it on CodeArena.

Zhipu (Z.ai): 300B Model on Domestic Chips

Zhipu launched a 300B-parameter flagship trained entirely on 100K+ domestic chips — a milestone for fully domestic training pipelines. The company is also restricting user access amid the compute crunch.

Open-Source LLM Economics

A Robonomics/FD report (September 1) estimates China’s open-source core LLM revenue at $60–80 billion, with total ARR of $100–150 billion. Pricing power has shifted from price war to premium: DeepSeek raised API prices 3–12x; Zhihu’s API pricing increased 101% with token usage up 40x year-to-date.

Robotics & Autonomous Driving

DiDi R2 Robotaxi: Driverless Trials Begin

Fully driverless passenger trials began August 31 in Beijing and Guangzhou via the DiDi app. The R2 is built with GAC Aion — the first China robotaxi joint venture between an AV firm and an automaker. Specs: L4 full-stack, 33 sensors, >2,000 TOPS, triple-domain compute. The R2 has been running a 24/7 driverless pilot in Guangzhou’s Huangpu district since December 2025. The UAE is the first overseas market with 2026 tests planned. DiDi claims a 74% compute-platform cost cut.

Momenta: Revenue Record, Global Expansion

Momenta posted record H1 2026 revenue with a sharply narrowed loss. Its world-model (“mapless”) L4 is running on mass-production cars, with 1.1 million cumulative production-vehicle installs. Commercial robotaxi is planned this year in Shanghai (via SAIC’s Xiangdao) and Abu Dhabi. Momenta is the first Chinese firm with a nationwide German L4 test permit (July), testing in six cities including Munich. L3 R&D targets 2027 mass production. A Robovan is operating in Suzhou.

Shanghai Robotaxi Scale-Up

Shanghai’s Pudong “high-level autonomous driving pilot zone” is targeting 6 million+ L4 passenger trips by 2027. Baidu Luobo Kuaipao and Pony.ai are operating paid robotaxi routes at fares roughly one-third to one-half of taxi prices. Momenta and Pony.ai are finalizing unmanned licenses. Pudong expansion is heading toward airport–Disney corridors and highways.

US Robot Import Ban

The US is banning imports of new Chinese robots. The Section 1286 entity list now covers 88 Chinese entities.

Market & Macro

Capital Race

The capital gap is narrowing. US hyperscalers (Microsoft, Amazon, Google, Meta, Oracle) are deploying approximately $791 billion in AI capex versus the Chinese four (ByteDance, Alibaba, Tencent, Baidu) at approximately $118 billion — but Chinese firms are aggressively tapping debt and equity markets:

• Alibaba: $10.2B placement + $56B three-year commitment
• DeepSeek: $7.4B round at $74B valuation
• Moonshot: $3B IPO target + pre-IPO round
• ByteDance: $30B loan + RMB 200B capex
• Baidu: buyback, dividend, dual listing

Memory and Raw Materials

• YMTC wafer output: 2.01M (2026) → 2.50M (2027) on the third Wuhan plant expansion.
• The US is funding an Australian gallium plant (Alcoa, $174M) to cut reliance on China, which controls 98% of global gallium supply.
• A US one-year ban on tungsten scrap and “black mass” exports to China is in effect.
• China controls rare earth and magnet exports — the raw material war is bidirectional.

Hong Kong Tech Rebound

Hong Kong tech stocks rebounded September 4, with JD, Baidu, and Tencent all up 2%+. UBS sees pricing power shifting from upstream chips and infrastructure to internet platforms over the next 2–3 years as compute bottlenecks ease.

Trends to Watch

1. Inference is the new battleground. DeepSeek’s 160K Ascend cluster proves domestic chips can win inference at scale. The open question: when will training-on-Ascend become viable?
2. Cloud access controls. The next US lever is remote GPU rental — but statutory authority is unclear. Watch the Senate Banking Committee and BIS draft rules in September.
3. Agentic AI regulation wave. China has built the world’s most structured agent-governance regime while pushing 70% terminal adoption by 2027. The rest of the world is still debating whether agents need their own regulatory category.
4. Model race consolidation. Qwen vs Kimi K3 as the open-weight leaders. Qwen4 is imminent. DeepSeek is IPO-bound. ByteDance admits its LLM gap is widening.
5. Commercialization pivot. Token price hikes are sticking. Baidu says AI margins will approach search levels. MIIT is building provider pools with procurement incentives. UBS sees pricing power shifting to platforms.
6. Memory as the new bottleneck. CXMT HBM3E samples are promising but two years behind. Huawei 950DT output is capped by HBM. Memory, not logic, may determine the pace of China’s AI independence.
7. Robotaxi scale-out. DiDi, Momenta, Baidu, and Pony.ai are racing toward commercial L4. Shanghai targets 6M trips by 2027. Overseas expansion to the UAE and Germany has begun.

China AI Weekly is produced by Big Hat Group — Microsoft partner and AI consulting practice. Follow along on X/Twitter for daily updates. Have questions about how China’s AI developments affect your business? Get in touch.