The week of July 11–18, 2026 may be remembered as the moment China’s AI ambitions crystallized into a coherent geopolitical and industrial strategy. DeepSeek is racing toward a public listing at a $74 billion valuation. Moonshot AI dropped the largest open-weight model in history. Xi Jinping personally attended the World Artificial Intelligence Conference to unveil a parallel global governance order. And a cascade of regulatory frameworks — from AI companion rules to agent governance — took effect, reshaping the market in real time.

Here is our professional analysis of this week’s most critical developments.

DeepSeek: The $74 Billion Champion and Its IPO Path

A Staggering Valuation Trajectory

DeepSeek is planning a fresh fundraising round at approximately 500 billion yuan ($74 billion) valuation — a 37% jump in under two months. This follows the June round that raised $7.4 billion at a 450 billion yuan post-money valuation. The new round seeks to raise up to 50 billion yuan ($7B), and with state-backed AI fund participation in the prior round, DeepSeek’s strategic importance to Beijing is unambiguous.

To put this in perspective: DeepSeek’s annualized revenue recently hit approximately $500 million (per The Information). At a $74 billion valuation, that equates to a ~148x sales multiple — a figure that would make even the most optimistic Western AI investor pause. But in China’s current investment climate, where frontier AI labs are seen as national champions rather than mere commercial enterprises, traditional valuation metrics are secondary to strategic positioning.

The IPO Push — Shanghai STAR Market

DeepSeek has begun early deliberations on a potential IPO on Shanghai’s Nasdaq-style STAR Market, with an internal target to complete the filing this year. Founder Liang Wenfeng reportedly prefers pursuing AGI over monetization — a posture that would be luxuriously patient in Silicon Valley but is becoming untenable in China’s shallower private capital pool. The reality is that China’s venture capital ecosystem cannot sustain the burn rates required for frontier AI development through private rounds alone. Going public earlier than their Western counterparts is not a choice; it is a structural necessity.

This has implications for how DeepSeek will operate as a public company. Public markets demand quarterly results, revenue growth, and margin improvement — pressures that could force a company philosophically oriented toward AGI research into more commercial positioning. Watch for how DeepSeek navigates this tension if the IPO proceeds.

Custom Inference Chips: Full-Stack Sovereignty

Reuters reported earlier this month that DeepSeek has been secretly developing its own AI inference chip for approximately one year. The company is in talks with chip design, foundry, and memory partners, with SMIC as the likely fabricator. The focus is on inference rather than training, with initial deployment targeted for end of 2026 and full-scale ramp through 2027–2028.

This is a three-dimensional strategic play. First, it reduces reliance on NVIDIA — the obvious target of US export controls. Second, it reduces reliance on Huawei’s Ascend ecosystem, which would give Huawei outsized leverage over China’s entire AI industry. Third, it positions DeepSeek as a full-stack AI company — model, inference engine, and silicon — a vertical integration that no other Chinese AI lab currently matches.

DeepSeek’s V4 model, released in April, was the first frontier model co-designed with Huawei’s Ascend 950DT to run natively on Chinese domestic silicon. The company then triggered a price war in May with a permanent 75% discount on V4 API pricing, bringing output pricing to $0.87 per million tokens — the cheapest among frontier providers. More recently, DeepSeek introduced peak-hour surcharges on V4-Pro API to manage service stability, a sign that even the cheapest provider is feeling demand-side pressure.

Moonshot AI: Kimi K3 and the “Third Shockwave”

The Largest Open-Weight Model Ever Released

On July 16, the eve of WAIC 2026, Moonshot AI unveiled Kimi K3 — a 2.8 trillion parameter model that is the largest open-weight model ever released. To contextualize: it is nearly double DeepSeek V4 Pro’s 1.6 trillion parameters and roughly four times any prior Chinese open model.

The architecture is worth examining in detail. K3 uses a Mixture-of-Experts design with 896 experts (16 active per forward pass), combined with Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). It supports a 1 million token context window and is natively multimodal (text and vision). Moonshot claims a 2.5x efficiency gain over K2 in compute-to-capability conversion — an assertion that, if validated, suggests architectural innovation is keeping pace with raw scale.

Two variants ship: K3 Max for chat and agent workflows, and K3 Swarm Max for parallel large-scale research. Weights will be open-sourced on July 27.

Benchmark Performance: Beating the Western Frontier

K3’s benchmark results are where this story gets consequential. On the Artificial Analysis Intelligence Index, K3 scores 57 — beating Anthropic Opus 4.8 (~56) and OpenAI GPT-5.6 Terra (~55). On Arena.ai, it ranked #1 overall and #1 in front-end coding, surpassing Anthropic Fable 5 and OpenAI Sol. It outperforms on Program Bench and SWE Marathon (real codebase debugging and maintenance), with performance described as “close to Anthropic Fable 5” at a fraction of the cost.

This is the performance data that matters. For the better part of two years, the narrative has been that Chinese models are cheap but lag Western frontiers on quality. K3 challenges that assumption directly — it is not merely competitive on price; it is competitive on capability.

Premium Pricing: A Strategic Repositioning

K3’s pricing represents a deliberate strategic shift: $3 per million input tokens and $15 per million output tokens — 3-4x K2.6 pricing and significantly more than DeepSeek or GLM. This is still roughly half of OpenAI GPT-5.6 Sol and considerably cheaper than Anthropic Fable 5 ($50/M output). Average task cost comes in at $0.94, compared to Claude Opus 4.8 at $1.80 and GPT-5.6 Sol at $1.04 — competitive total economics driven by fewer retries and higher first-pass accuracy.

The strategic signal here is more important than the numbers. Moonshot is moving Chinese AI from commodity token supplier to premium capability provider, competing directly with Anthropic and OpenAI for high-value enterprise workloads. This has been called the “third shockwave” from Chinese AI — after DeepSeek R1 and Manus — and it represents a maturation of the Chinese AI industry from price disruption to capability leadership.

Moonshot’s Funding Velocity

Moonshot completed six funding rounds in 2026 alone. The previous round on June 30 valued the company at $20 billion. A new round is now underway at a $31.5 billion pre-money valuation. Backed by Alibaba, Tencent, and Meituan, Moonshot is capitalizing on the K3 launch to accelerate its position — and the investor appetite suggests the market believes the premium positioning strategy will work.

WAIC 2026 and Xi Jinping’s AI Governance Vision

Xi’s Personal Attendance — A Strategic Signal

The World Artificial Intelligence Conference (WAIC) ran July 17–20 in Shanghai, themed “AI Partnership for a Brighter Future.” Over 140 forums hosted 1,400+ guests, with 300+ debut products showcased. But the headline was Xi Jinping’s personal attendance — the first time China’s president has personally attended WAIC, signaling AI’s elevation to the highest tier of national strategic priority.

Xi’s Keynote: A Parallel AI Order

Xi’s speech cast Beijing as the champion of a new global AI order, built on four pillars: openness and win-win cooperation, risk awareness and security, inclusiveness and mutual learning, and solidarity in global governance. He compared AI’s significance to the steam engine and electricity — historical general-purpose technologies — and called for AI to remain under human control, offering his clearest AI safety remarks to date. He urged early-warning and emergency-response mechanisms.

Critically, Xi opposed “overstretching the national security concept” in AI — a veiled but unmistakable critique of US export controls. The framing positions China as the advocate for open, accessible AI against a US-led regime of restrictions. Whether one accepts this framing or not, it is being articulated at the highest level, and it is finding receptive audiences across the Global South.

WAICO: A Rival Governance Institution

The launch of the World AI Cooperation Organization (WAICO), headquartered in Shanghai with 29 founding member countries, is the institutional expression of Xi’s vision. It is positioned as a rival to US-led initiatives and comes with concrete commitments:

  • 5,000 AI training opportunities for developing countries over five years
  • AI cooperation centers with ASEAN, the Arab League, the African Union, CELAC, SCO, and BRICS
  • Access to China’s AI meteorological early-warning system for 30 countries
  • Chinese open-source AI models have been downloaded over 10 billion times globally

This is not merely symbolic. WAICO creates a institutional architecture through which Chinese AI technology, standards, and governance models can diffuse across the developing world. It is the soft-power complement to the hard infrastructure of chips and models.

China AI Action Plan

The NDRC and partner departments released the China AI Action Plan on July 17, covering eight areas: data, computing power, ecosystems, industrial empowerment, talent, rules and standards, governance, and AI ethics. The plan calls for greater access to high-quality data, inclusive computing infrastructure, and broader open-source sharing. It is the policy backbone that ties together the regulatory and industrial developments described throughout this analysis.

China’s Domestic Chip Stack: From Huawei to SMIC

Huawei Atlas 950 SuperPoD — 8,192 Ascend Chips

The Atlas 950 SuperPoD made its public debut at WAIC, linking 8,192 Ascend NPU chips via high-speed interconnects into a single computing cluster designed for large-scale AI training and inference. Critically, there are no NVIDIA chips in the critical path. DeepSeek V4 has been adapted to run entirely on Huawei Ascend clusters, proving that the domestic stack can support frontier model workloads.

Ascend 950 Series: Challenging NVIDIA’s China Position

The Ascend 950PR and 950DT are the first Chinese AI chips with integrated in-house HBM (High Bandwidth Memory). The 950PR delivers 1.56 petaflops FP4 — Huawei claims 2.87x the performance of NVIDIA’s H20, which is the only NVIDIA chip legally exportable to China. Huawei is targeting 750,000 units and $12 billion in AI chip revenue in 2026, representing 60% year-over-year growth. ByteDance alone has committed $5.6 billion to the Ascend platform.

The CANN software stack provides an alternative to NVIDIA’s CUDA ecosystem — and while it is not yet at CUDA’s maturity, the gap is closing as more developers are forced onto the platform by export controls. T-Head (Alibaba’s chip unit) is also targeting CUDA’s ecosystem with an open-source AI stack to lower migration barriers to its Zhenwu architecture.

SMIC: Defying EUV Restrictions

SMIC’s N+3 node (5nm-class) is now in high-volume production using DUV multi-patterning — no EUV required. Yields are estimated at 30-40%, far below TSMC’s 80%+ but subsidized as a matter of national security. Advanced capacity is approaching 100,000 wafers by late 2026, yielding an upper bound of approximately 3 million AI semiconductors per year. SMIC is also producing 2 million HBM stacks in 2026 — enough for roughly 300,000 Ascend 910C packages.

China’s first fully domestic 100,000-card AI supercluster went live earlier this month, running entirely on Huawei Ascend chips with no NVIDIA or TSMC-fabricated silicon. This is a milestone that would have been dismissed as impossible two years ago.

US Export Control Tightening

The Trump administration confirmed that the chip ban extends to subsidiaries of Chinese companies outside China — closing the Singapore/Malaysia loophole. The MATCH Act, which would eliminate DUV immersion tool servicing (SMIC’s last legal supply line), cleared the House Foreign Affairs Committee but has not yet become law. A US Section 301 finding from December 2025 branded Beijing’s semiconductor expansion “unreasonable and discriminatory” — the tariff is currently 0% but is set to rise in June 2027. The screws are tightening, but China’s domestic stack is advancing faster than the restrictions can constrain it.

Big Tech Moves: Apple Intelligence, Alibaba, Baidu, Tencent, ByteDance

Apple Intelligence Approved in China

The CAC approved Apple Intelligence for iPhone users in China, with Alibaba’s Qwen handling LLM integration and Baidu powering search and AI features. Alibaba’s Qwen is now integrated into Apple Intelligence across iOS, iPadOS, macOS, and visionOS, though the CAC license is initially iPhone-only. Alibaba’s HK shares rose 5%; Baidu’s rose 4.28%.

This approval signals that China’s AI race is now a three-platform contest: Alibaba (cloud + Qwen), Baidu (search + Ernie + Kunlunxin chips), and Huawei (Pangu + Ascend). Tencent is positioned as the “natural next participant” for Apple Intelligence partnership — and with WeChat’s billion-plus users, its entry would reshape the competitive dynamics.

Baidu’s Kunlunxin IPO

Baidu’s AI chip unit Kunlunxin is targeting a Hong Kong IPO at approximately $50 billion valuation — described by analysts as the “cleanest cross-border AI chip test of the cycle.” A successful listing would validate the standalone value of China’s domestic AI chip design capabilities outside the Huawei ecosystem.

Tencent: Hunyuan Hy3 and WeChat AI Agents

Tencent’s Hunyuan Hy3 model launched with a dramatic 68x increase in total calls versus the previous generation in its first week. Tencent is expanding AI computing capacity to meet demand and is testing an AI agent within WeChat’s 1 billion+ user base — a deployment that, if successful, would put AI agents into the hands of more people than any other platform globally.

Tencent is also slashing its stake in Kuaishou (from 15.68% to 9.37%) following Kuaishou’s $3 billion Kling AI deal, and is leading discussions to reacquire shares in Manus (the AI agent startup) after Chinese authorities directed the reversal of Meta’s approximately $2 billion acquisition.

Alibaba: Qwen Audio, Claude Code Ban, and Strategic Investments

Alibaba released Qwen-Audio-3.0-Realtime, a real-time voice interaction model with upgrades to IQ, agent tool invocation, empathic dialogue, and duplex fluency. The company led a $439 million Series C in AIsphere, an AI video startup challenging OpenAI’s Sora — its second investment after a $60 million Series B in September 2025.

Notably, Alibaba banned staff from using Claude Code over Anthropic spyware concerns, classifying it as “high-risk” software after hidden user-tracking code was discovered. Domestic alternatives like ByteDance’s Trae and Alibaba’s own Qoder are filling the gap. This is not merely a security decision — it is a deliberate decoupling of Chinese AI development from Western developer tooling, accelerating the emergence of a parallel software ecosystem.

ByteDance: 300M Users and $5.6B Chip Commitment

ByteDance’s Doubao chatbot now boasts 300 million+ active users, making it China’s most-used AI consumer product. Doubao shut down custom AI agent features on July 15 to comply with the new companion AI rules (discussed below). ByteDance’s coding tool Trae is being positioned as the domestic alternative to Claude Code. The company committed $5.6 billion to the Huawei Ascend platform — a staggering sum that signals where ByteDance believes the infrastructure layer is heading.

The OpenAI/Google Loophole

OpenAI and Google are under scrutiny for providing advanced AI models to Alibaba, Baidu, and Tencent subsidiaries in Singapore. These parent companies are on the US Department of Defense blacklist, but the transactions are not currently illegal. OpenAI suspended Alibaba-linked users due to suspected “distillation” activities — a sign that the battle over model IP is intensifying. This is reigniting debate in Washington about whether AI models themselves should be subject to export controls, not just chips.

Regulatory Earthquakes: AI Companions, Agents, and Ethics

World’s First AI Companion Regulations — Now in Effect

The Interim Measures for Administration of Anthropomorphic AI Interaction Services took effect July 15, 2026 — the world’s first national regulation targeting AI companion chatbots. Issued by five agencies (CAC, NDRC, MIIT, MPS, SAMR), the rules establish:

  • Mandatory AI-identity disclosure — users must always know they are interacting with AI
  • Absolute ban on virtual intimate relationship services for minors under 18
  • Parental consent required for users under 14
  • Anti-addiction systems: 2-hour break prompts, dependency detection, crisis intervention for self-harm or suicidal ideation
  • Real-name verification for agent creation
  • Mandatory security assessments at 1 million registered users
  • Enhanced safeguards for elderly users
  • Explicit consent required before using chat data for model training
  • Instant-exit mechanisms for any AI interaction

The impact was immediate. Both ByteDance’s Doubao and Alibaba’s Qwen disabled custom AI agent features to comply rather than retrofit them. When the two largest consumer AI platforms choose removal over compliance, the regulatory bar is objectively high. The rules exempt customer service, knowledge Q&A, work assistants, and education/research tools — but anything involving sustained emotional interaction falls under the new regime.

AI Agent Governance Framework

Also effective July 15, the Implementation Opinions on Standardized Application and Innovative Development of Intelligent Agents established China’s first comprehensive framework for autonomous AI agents. Four core requirements: meaningful human oversight, full traceability, strict authorization controls, and enhanced security reviews for finance, government, and critical infrastructure applications. This is a forward-looking framework — it anticipates a near future where AI agents operate autonomously in high-stakes environments.

AI Ethics Guidelines 1.0

The Guidelines for Ethical and Safety of AI Applications 1.0, released July 1 by the National Technical Committee on Cybersecurity, establish nine core ethical principles: human welfare, rights respect, controllability, and trustworthiness among them. While not mandatory law, they are influential standards backed by a 10-province pilot for AI ethics review running June through November 2026.

AI Device Grading and Data Protection

China issued its first Level 3 certifications under a new national AI device grading system, recognizing 66 products from 17 companies. The China Internet Association released a self-regulatory pact on personal information protection for AI agents, signed by Baidu, Tencent, Alibaba, Volcengine, and 27 others. The pact signals that industry self-regulation is running in parallel with — and somewhat ahead of — formal regulation.

The 15th Five-Year Plan Context

The 2026–2030 plan cements “AI Plus” — AI diffusion across all economic sectors — as the overarching policy framework. The CAC and two partner agencies issued dedicated guidance on agentic AI in May 2026, with a full chapter on safety. Mandatory AI content labeling and watermarking (GB 45438-2025) has been enforced since September 2025. A comprehensive national AI law has been signaled by the State Council — the legislative net is tightening.

Funding and Startup Ecosystem: Capital Flooding In

MiniMax Closes $2B Round

Shanghai-based MiniMax, the lab behind the Hailuo video model, closed approximately $2 billion (HK$16B) in equity funding, placing it in the same tier as mid-stage Western frontier labs. The company is competing increasingly on multimodal capability — video generation in particular — and the capital suggests it will be scaling aggressively.

Embodied AI and Robotics: Explosive Growth

H1 2026 saw ¥93.5 billion across 322 financing deals in embodied AI and robotics — a 5x increase year-over-year and a 137% jump in deal volume. 22 unicorns emerged from just 3 at end of 2025. Over ¥40 billion went to the top-22 unicorn financings, with AI² Robotics securing the largest single round at approximately ¥50 billion. June alone saw 36 disclosed rounds, 10 exceeding ¥1 billion.

Notably, 50% of new unicorns have state capital involvement — the government is not merely regulating; it is actively financing the robotics frontier. This is the next great AI wave in China, and the capital infrastructure is already in place.

AIsphere, Aishi, and StepFun

Alibaba-led $439 million Series C in AIsphere (AI video, challenging OpenAI Sora). Aishi Technology raised 2.98 billion RMB (~$410M) in a Series C. Tencent-backed StepFun launched StepX Neo, claimed as the “world’s first AI-native smartphone,” integrating a proprietary OS and AI agent — beating Apple and OpenAI to market on the AI-native device concept.

The Market Shift: Top 5 Most-Used Models Are All Chinese

As of this week, the five most-used AI models globally are all Chinese: Tencent, Xiaomi, DeepSeek, MiniMax, and z.ai. This is a staggering market shift. The performance gap between Chinese and Western frontier models has narrowed sharply, and with Chinese models priced 60-90% cheaper than Western alternatives, global developers are pivoting.

Sam Altman signaled that OpenAI may cut flagship fees by 75% to counter market gains by Claude and low-cost Chinese providers. When the dominant Western AI company is forced to consider 75% price cuts to remain competitive, the market has fundamentally changed.

Chinese AI models now account for over 30% of OpenRouter token consumption, peaking as high as 46%. On Hugging Face, Chinese developers account for 17.1% of downloads versus 15.8% for U.S. developers. Eight of the top ten open-source large models are from China. The open-source empire is not a projection — it is a present-tense reality.

Key Takeaways

  1. DeepSeek’s $74B valuation and IPO path represent China’s frontier AI labs maturing into public-market companies — earlier than Western counterparts, driven by structural capital constraints.
  2. Kimi K3’s 2.8T parameter open-weight release challenges the narrative that Chinese models are cheap but inferior. It beats Western frontier models on coding benchmarks at premium-but-competitive pricing.
  3. Xi’s WAIC speech and WAICO launch articulate a parallel global AI governance order with 29 founding nations — open-source as geopolitical tool.
  4. The domestic chip stack — Huawei Ascend 950, SMIC N+3, custom DeepSeek silicon — is advancing faster than US export controls can constrain it.
  5. AI companion and agent regulations are the world’s most advanced AI governance frameworks for emotional and autonomous AI, forcing immediate product changes at the largest platforms.
  6. All top 5 most-used AI models are now Chinese. The market has crossed an inflection point that Western AI companies cannot ignore.

This weekly analysis is based on recent market developments, technical releases, and policy announcements shaping the trajectory of AI in China. Follow along at https://x.com/kkaminski for ongoing coverage.