China AI Weekly: DeepSeek's $7.4B Raise, World's First Agentic AI Law, and the Permanent Price War

The Week That Was

June 2026 marks a structural inflection point for China’s AI ecosystem — a week where four converging megatrends reshaped the landscape. Here’s what happened and why it matters.

1. DeepSeek’s $7.4B Mega-Round

DeepSeek closed its first-ever external fundraising at a post-money valuation of ¥350–400B (~$52–59B), making it China’s most valuable pure-play AI startup.

The investor lineup tells a story of its own:

• Tencent — ¥10B (~$1.5B)
• CATL — ¥5B (~$740M)
• NetEase + JD.com — finalizing participation
• National AI Industry Investment Fund — ~¥1B state-backed
• Founder Liang Wenfeng — personally committed ~¥20B via a limited partnership preserving his control

The “no-poach” clause is perhaps the most telling detail. DeepSeek reportedly required investors to agree not to recruit its employees — a direct response to losing Luo Fuli (key architect of DeepSeek V3) to Xiaomi’s MiMo lab.

Meanwhile, the US is reportedly holding off on adding DeepSeek to the Entity List, despite interagency committee approval last year. At least 75+ Chinese entities remain queued for blacklisting in a pipeline not updated since October 2025.

2. World’s First Agentic AI Framework

On May 8 (published June), the CAC, NDRC, and MIIT jointly released the “Implementation Opinions on Intelligent Agents” — the world’s first comprehensive regulatory framework for agentic AI.

This is not just regulation. It’s infrastructure design.

Key features:

• Tiered governance (L1-L4): Low-risk consumer agents need minimal oversight; high-risk agents (healthcare, finance, transport) require mandatory filing, third-party testing, and recall plans
• Three-tier decision authority: Human-only → User-authorized → Autonomous (with user retaining final decision rights)
• “Intelligent Internet” vision: A future protocol layer (AIP over IPv6) where agents have digital identities, register capabilities, and transact directly

China is not just regulating agents — it’s defining the protocol they’ll operate on, embedding compliance into the architecture itself. Nineteen priority scenarios are identified, from scientific research to public safety.

This is the most consequential AI regulatory development globally this year. Anyone building agentic systems should study it.

3. The Permanent Inference Price War

DeepSeek’s permanent 75% price cut on V4 Pro (effective June 1) has reset global pricing:

This is structural, not promotional — driven by optimization for Huawei Ascend 950 chips and architectural improvements delivering long-context inference at ~¼ the compute and ~¹⁄₁₀ the memory of predecessors.

The scale is staggering: China’s ecosystem now processes 140 trillion tokens/day — approximately 3.5× North America and Europe combined. ByteDance’s Doubao alone generates 120T tokens/day (up 7.3× from last May). Chinese models now account for 48.8% of all tokens on OpenRouter.

The ripple effects are global: Google just cut its AI Plus plan from $7.99 to $4.99/month. Baidu, unable to compete on price alone, launched Wenxin 5.1 as “the model that understands Chinese business logic.” Zhipu AI raised API prices 3× in 2026 (and saw ARR grow 60× to ¥1.7B).

4. The Chip War: Nvidia Goes to Zero in China

The semiconductor story this week is stark. Per Morgan Stanley:

Nvidia’s China AI chip market share: ~95% in 2022 → 0% in 2026.

When Trump authorized H200 sales to China (Dec 2025) with a 25% tariff, Beijing told Chinese companies not to buy them. As of May 2026: zero H200 chips sold to Chinese companies.

Huawei Ascend now dominates at 62% projected market share. The rest splits among Cambricon, Baidu (Kunlun), Alibaba, and others. DeepSeek V4 is optimized for Ascend 950, with mass production of 950PR units expected H2 2026.

Huawei’s “logic folding” (3D chip stacking) breakthrough targets performance equivalent to 1.4nm chips by 2031 — without ASML EUV lithography.

Meanwhile, China’s $295B national AI infrastructure bet (¥2 trillion over 5 years) requires 80% domestic supplier content. The Inner Mongolia Token Exchange launched May 30 as China’s first national green computing AI platform — a regulated marketplace for compute, model API access, and token settlement.

5. The Capital Markets Story

Chinese AI startups are going public at unprecedented velocity:

Zhipu and MiniMax were added to the Hang Seng Tech Index (June 5) — the first pure-play AI companies in the index. Bloomberg Intelligence projects up to HK$139B in southbound mainland inflows.

Zhipu’s market cap (~$88B) has surpassed Baidu despite having only 0.56% of its revenue. The market is pricing AI-native companies at extreme premiums.

6. Tencent’s Connector Strategy

Tencent is making an explicit bet: not to build China’s smartest model, but to be the connector layer every agent must call.

The FT reported (June 2) that Tencent is testing a prototype AI agent for WeChat — accessed by swiping right on the main screen. Stock surged 10% on the news, the biggest one-day gain since November 2022. With 1.4B WeChat users and a Mini Program ecosystem that already handles real-world services, Tencent has the distribution to make this work.

The company dissolved its AI Lab (March 2026) after nearly a decade, shifting focus to commercial delivery. It hired Yao Shunyu (former OpenAI researcher) as Chief AI Scientist and open-sourced Hy3 (295B total params, 21B active, MoE).

What to Watch

The converging trends — DeepSeek’s valuation, the agentic regulatory framework, the permanent price reset, and the accelerating domestic chip substitution — paint a picture of an ecosystem that is no longer just following. It’s defining the next phase of the AI race on its own terms.