The week of July 18–25, 2026 delivered what may be the most consequential seven days in China’s AI story since DeepSeek R1 shocked the world in January 2025. DeepSeek’s founder gave a four-hour manifesto on AGI that went viral — and then suspended the company’s second funding round. Moonshot AI’s Kimi K3 proved that China’s frontier model moment was not a one-off. Huawei demonstrated an exascale supercomputer built with zero American components. The world’s first AI agent regulations kicked in, erasing 345 million users’ AI companions overnight. And Hugging Face turned to a Chinese model for cybersecurity work that US models refused to do.

This is not another week of incremental progress. This is the week the fulcrum shifted.

Here is our professional analysis.

DeepSeek: The AGI Gambit and the Funding Freeze

A Founder’s Four-Hour Manifesto

On July 23–24, DeepSeek founder Liang Wenfeng held a four-hour meeting with investors that would ripple through the entire Chinese AI landscape. Reported by Yicai/China Business News, his message was uncompromising:

AGI over profit. DeepSeek’s top priority is developing artificial general intelligence, not maximizing revenue or racing for user acquisition. This is not marketing — Liang has consistently positioned DeepSeek as a research-first organization, and his willingness to pause a $7.4 billion funding round underscores that he means it.

Open-source as strategy, not charity. DeepSeek will keep its most advanced models open-source. Liang argued that open-source development and commercial monetization are not mutually exclusive — a position that aligns with Zhipu’s GLM-5.2 MIT license and Moonshot’s upcoming K3 weight release. The strategic logic is clear: open weights embed Chinese models into the global developer ecosystem in ways that proprietary models cannot, creating dependencies that export controls cannot sever.

Compute is the real gap. Liang’s most striking admission was about the US-China gap. It is not technical know-how — DeepSeek already knows the research directions it wants to pursue. The constraint is compute. With approximately 20,000 Nvidia H100 chips, DeepSeek is operating at a fraction of the scale available to OpenAI, Anthropic, or Google. Liang revealed that the company often must postpone research not because they lack ideas, but because they lack the hardware to execute them.

Nvidia’s grip is loosening. Liang said Nvidia’s AI chip ecosystem monopoly is gradually weakening, presenting an opportunity for China’s domestic chip industry. He hopes to buy chips at reasonable prices rather than build them himself — but DeepSeek has been reported to be developing its own inference chips, suggesting hedging on multiple fronts.

The Funding Round Suspension

On July 25 — just days after Liang’s comments went viral — DeepSeek told prospective investors it was suspending its second fundraising round. The round had been targeting a valuation of approximately 500 billion yuan ($74 billion), with plans to raise up to 50 billion yuan ($7.4 billion).

The timing is not coincidental. Liang’s viral comments about the US-China compute gap and his AGI-first posture created enough market noise to spook the fundraising process. But there may be a deeper logic. DeepSeek completed its first external round in June at a ~$52 billion valuation, with backing from Tencent (10 billion yuan), CATL (5 billion yuan), NetEase, JD.com, IDG Capital, and China’s national AI fund. Liang himself reportedly poured in an additional 20 billion yuan. The company is not desperate for capital — and pausing the second round may be a deliberate strategy to avoid over-diluting at a valuation that Liang may consider premature given his AGI timeline.

The IPO Path and the STAR Market

DeepSeek has begun early deliberations on an IPO on Shanghai’s Nasdaq-style STAR Market, with an internal target to complete a filing before end of 2026. This is a structural necessity, not a choice. China’s private capital pool is shallower than Silicon Valley’s, and frontier AI burn rates cannot be sustained through private rounds indefinitely. Going public earlier than Western counterparts is the only viable path to long-term capital independence.

The tension is obvious. Public markets demand quarterly results and revenue growth. Liang’s AGI-first philosophy is inherently patient. How DeepSeek navigates this collision between research culture and market discipline will be one of the defining corporate governance stories of 2026–2027.

The Quant Fund Drawdown

Adding to the week’s drama, Liang’s quant hedge fund Zhejiang High-Flyer Asset Management — which oversees over 70 billion yuan ($10 billion) — saw one of its funds slump 15.7% in the week ended July 17, amid a broader China quant rout tied to the global chip stock selloff. Liang’s personal financial situation is now entangled with market volatility in a way that could influence DeepSeek’s strategic decisions. The fund’s performance is a reminder that the personal wealth powering DeepSeek is not immune to the very market dynamics that export controls and chip competition have set in motion.

Moonshot AI: Kimi K3 and the “Kimi Moment”

2.8 Trillion Parameters — The Largest Open-Weight Model Ever

Beijing-based Moonshot AI (backed by Alibaba at 36% and Tencent) launched Kimi K3 on July 17, and it immediately became the story of the week. At 2.8 trillion parameters, K3 is the world’s largest open-weight AI model — the first to approach the 3 trillion parameter mark. It features a 1 million-token context window and native multimodal support (text and images).

The performance numbers are where this gets consequential. K3 approaches Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol on overall intelligence. It beats Claude Opus 4.8 and GPT-5.5 on coding and agentic tasks. On benchmark evaluations, it is not merely “competitive for a Chinese model” — it is competitive globally, full stop.

Moonshot has promised full open weights by July 27, 2026, with vLLM and SGLang providing Day-0 inference support. This matters enormously. When weights drop, developers worldwide can run K3 locally, fine-tune it, build on it, and integrate it into workflows without API dependency on Moonshot’s infrastructure. This is the open-source wedge strategy in action.

Premium Pricing — A Strategic Repositioning

K3’s API pricing is $3 per million input tokens and $15 per million output tokens — the most expensive Chinese model to date, but still approximately 60% of Claude Opus 4.8 pricing. This is not a commodity play. Moonshot is positioning K3 as a premium capability provider, competing directly with Anthropic and OpenAI for high-value enterprise workloads rather than competing on token price alone.

The strategic signal: Chinese AI is moving from cost disruption to capability leadership. The “third shockwave” — after DeepSeek R1 and Manus — is not about being cheaper. It is about being as good, and then being cheaper.

Demand Overwhelming Supply

K3’s reception was so intense that Moonshot had to halt new subscriptions within two days of launch due to GPU shortages. This is a demand signal that validates the premium pricing strategy — and a supply constraint that highlights exactly the compute gap Liang Wenfeng described. Even with Alibaba and Tencent backing, Moonshot cannot provision enough GPUs to meet demand for a model that proves China can compete at the frontier.

Distillation Accusations — and Why They Matter

US officials and Anthropic accused Moonshot of distillation — training K3 on outputs of American models. Moonshot denied the claims, and most independent analysts found the accusations implausible given K3’s architecture and training methodology. But the accusations themselves are a signal. When the narrative shifts from “Chinese models are cheap knockoffs” to “Chinese models are so good they must have stolen our techniques,” the underlying assumption has changed. The accusation is an implicit admission that the performance gap has closed.

Market Impact

Kimi K3 and Alibaba’s Qwen3.8 Max triggered a rotation from Asian semiconductor hardware into Chinese AI platform stocks. Alibaba rose 4–6% in Hong Kong; Tencent gained 4%; Baidu added 3.8% premarket. Goldman Sachs declared that China’s open-source AI systems have achieved a “critical point” enabling worldwide adoption. Bloomberg Intelligence reported that the US-China AI performance gap narrowed to a record-low 6% in June 2026, down from 9% in May.

A 6% gap — and closing. This is the number that should be on every enterprise AI strategy deck.

Alibaba: Qwen3.8 Max and the Full-Stack Play

Qwen3.8 Max Preview

Alibaba released a preview of Qwen3.8-Max on July 20. At 2.4 trillion parameters, it is positioned as “second only” to Anthropic’s Claude Fable 5. Alibaba claims it outperforms OpenAI’s GPT-4.1 and Google’s Gemini 2.5 Pro on several benchmarks — though these claims have not been independently verified.

The flagship model is not open-weight (smaller Qwen variants are), but it is available through Alibaba’s cloud platform and developer tools. This is a deliberate bifurcation: open-weight smaller models for ecosystem embedment, proprietary flagship for cloud revenue. It is the same strategy that Western cloud providers employ, and Alibaba is executing it with increasing sophistication.

The Apple Intelligence Deal

Apple Intelligence received Chinese regulatory clearance in mid-July 2026 with Alibaba’s Qwen as the underlying model. This cannot be overstated. Apple’s installed base in China numbers in the hundreds of millions of devices. Qwen will be the AI engine behind Siri and Apple Intelligence features for the world’s largest smartphone market. This is not a research partnership — it is a production deployment at consumer scale that gives Alibaba unparalleled distribution and data feedback loops.

Qwen Office and the Agent Strategy

At WAIC, Alibaba launched Qwen Office (currently in internal testing, operated by the DingTalk team), integrating three agent products: QoderWork, Wukong, and MuleRun. The company also showcased its Agent Native Cloud with AgentTeams for multi-agent orchestration, and DAMO Lingshu — an AI research agent platform for scientific research. Notably, Alibaba’s Elements Claw became the first AI agent to discover superconductors, a milestone from DAMO Academy that demonstrates AI’s move from text generation to scientific discovery.

The AI Companion Shutdown

On July 15, Alibaba shut down its Qwen/Tongyi Qianwen AI companion and user-created AI agent features immediately. Configurations and conversation histories were deleted with no export option. This was not Alibaba’s choice — it was mandated by new regulations (detailed below). But the abruptness and the lack of data export options reveal how quickly compliance can override user experience when regulators move first and ask questions later.

Z.ai (Zhipu AI): GLM-5.2 and Infrastructure Independence

The Model That Silicon Valley Can’t Ignore

Zhipu AI — internationally rebranded as Z.ai — released GLM-5.2 in June 2026, and its impact has only intensified through July. At 744B total parameters with ~40B active (Mixture-of-Experts), GLM-5.2 is the top open-weight model on the Artificial Analysis Intelligence Index, surpassing GPT-5.5 on SWE-bench Pro (62.1 vs 58.6). It carries an MIT license with no regional restrictions — the most permissive licensing of any frontier-class model.

The pricing is devastating: approximately $1.40 per million input tokens and $4.40 per million output tokens — roughly one-sixth of GPT-5.5’s cost and one-tenth of Claude Fable 5. On OpenRouter, GLM-5.2 is the most-routed model, and approximately 60% of token usage attributed to US companies on OpenRouter now comes from Chinese-developed models.

The Hugging Face Cybersecurity Incident

On July 22, Hugging Face used GLM-5.2 to analyze a rogue AI agent built with OpenAI technology — after US AI models declined the task because their safety guardrails prevented them from distinguishing between defender and attacker. GLM-5.2 had no such restrictions and successfully completed the analysis.

This is a watershed moment. It is not about which model is “better” in the abstract. It is about the concrete trade-offs that guardrails impose. US AI companies have built increasingly restrictive safety layers — and those restrictions are now creating functional gaps that Chinese models are filling. When Marc Andreessen and Snowflake CEO Sridhar Ramaswamy publicly endorse GLM-5.2, the Silicon Valley consensus is fracturing.

The Reuters headline said it plainly: Chinese AI’s role in stopping a rogue OpenAI agent “shows the cost of US guardrails.” This story will be cited for years.

The 1-Gigawatt All-Domestic Data Center

On July 21, Z.ai completed construction of a 1-gigawatt data center powered entirely by Chinese chips — no NVIDIA silicon whatsoever. Several clusters with over 10,000 chips each are already partially operational. The scale is equivalent to the power consumption of approximately 750,000 households. It matches the scale of Musk’s Colossus v2.

Z.ai shares surged 37% in a single session on the Hong Kong exchange following the announcement. The market read this as proof that the export control workaround works: train frontier models on domestic chips at exascale, and the compute gap becomes closeable.

GLM-5.2 was trained entirely on Huawei Ascend accelerators. The 1GW data center is the infrastructure bet that this approach scales — not just for one model, but for the entire GLM family going forward.

The IPO and Financial Trajectory

Z.ai IPO’d in Hong Kong in January 2026, raising approximately $558 million as the first major listed Chinese LLM developer. At one point, the stock traded at roughly 13× its IPO price. JPMorgan projects over 500% revenue growth for 2026. The combination of frontier model capability, infrastructure independence, and public market access makes Z.ai the most structurally positioned Chinese AI company to watch — and it does not get the same Western media attention as DeepSeek.

Huawei and the Domestic Chip Revolution

Atlas 950 SuperPoD — Exascale Without America

At WAIC, Huawei physically displayed the Atlas 950 SuperPoD for the first time — and it is a statement piece. The system links 8,192 Ascend 950DT processors into a single computing node, delivering 1 exaflops at FP8 and 2 exaflops at FP4, with 256 TB of globally addressable memory. Huawei claims 6.7× the computing power of NVIDIA’s NVL144.

Zero US-origin components. Full domestic stack — chips, interconnect, cabinets, software. Shipments are targeted for Q4 2026, and the system can scale to 500,000 accelerators.

This is the engineering answer to export controls. The Atlas 950 SuperPoD is not a prototype — it is a product with a shipment date. And it is built on the Ascend 950 PR processor, which delivers approximately 1.5 petaflops with 112 GB of HBM using Huawei’s proprietary “HBBl1.0” memory standard — entirely domestic, not subject to US export controls.

The Market Share Reversal

The numbers tell the story. Huawei’s Ascend line is projected to capture approximately 60% of the AI chip market inside China by end of 2026, with total revenue exceeding $12 billion (up from $7.5 billion in 2025). NVIDIA’s market share in China has dropped from roughly 95% pre-controls to approximately 40% in 2025, and is projected at roughly 8% in 2026. NVIDIA’s CFO confirmed zero H20 shipments to China this year.

Bloomberg Intelligence’s survey found that Huawei Ascend 910B/910C chips are in use or under evaluation in 65% of surveyed Chinese companies’ AI clusters — versus NVIDIA H20/L20 at approximately 47%. The export control strategy was designed to deny China advanced AI capability. The actual outcome has been to catalyze a domestic chip industry that is now displacing NVIDIA inside the world’s second-largest AI market.

The Broader Domestic Chip Ecosystem

WAIC showcased a maturing domestic chip stack beyond Huawei:

  • Dongfang Suanxin unveiled the DF1000 — 14nm process with performance claimed comparable to 4nm-class chips, scaling to 512-GPU clusters. A DF2000 is planned for later in 2026.
  • Moore Threads demonstrated domestic chips completing ultra-large-scale AI training, including full training of a MoE-236B foundation model from scratch.
  • MetaX showcased its Xijing S-series SuperNode and Xisuo X-series products.
  • Alibaba demonstrated its Zhenwu M890 × Panjiu AL128 SuperNode, following the Zhenwu 810E AI chip launch.

China plans approximately $300 billion in data center infrastructure over the next five years, with 80% of core technologies expected to come from domestic suppliers. This is not aspirational — it is budgeted.

China’s Own Export Controls

In a move that mirrors US policy, the Chinese government is considering measures to strengthen export controls on Chinese AI and semiconductor technologies. The symmetry is notable. China is not merely defending against US restrictions — it is beginning to treat its own AI technology as a strategic asset worth controlling. This will be worth watching closely, as it could restrict the very open-source flows that have been embedding Chinese models globally.

WAIC 2026: The Pivot from Models to Deployment

Scale and Political Signal

The World Artificial Intelligence Conference in Shanghai (July 17–21) hosted over 1,100 companies exhibiting 3,000+ technologies and products, including 300+ global launches. Xi Jinping keynoted for the first time — pressing for an open-source AI ecosystem and international collaboration for Global South integration, and warning that no single country should determine AI’s development course.

WAICO: The Alternative Governance Institution

29 countries — including Russia, Pakistan, and Kazakhstan — signed the founding agreement of the World AI Cooperation Organization (WAICO), headquartered in Shanghai. The Global AI Capacity Development Network (AICDN) was officially launched in Geneva on July 5 and has entered operational status. This is the institutional infrastructure for a parallel AI order — and it is being built while Western governance frameworks remain fragmented.

The Enterprise Agent Wave

WAIC 2026 marked a clear inflection: the industry is pivoting from model competition to commercial agent deployment. The evidence is overwhelming:

  • Ant Group launched Agentar 2.0 — a “commercial AI agent super factory” with 200 pre-configured digital expert templates.
  • Alibaba Cloud deployed Agent Native Cloud with AgentTeams for multi-agent orchestration.
  • Tencent Cloud launched WorkBuddy Enterprise AI Workspace as a standalone app on iOS, Android, and Hongmeng. Citi flagged WorkBuddy as an overlooked value driver.
  • Baidu rolled out DuMate, a general-purpose AI agent for data analysis, presentations, reports, and office tasks — alongside a full suite including Baidu Buddy, One Shot, General Agent GenFlow 4.0, Miaoda 3.5, and Famou 2.0.

The message is clear: China’s “AI plus” action plan has moved from models to workflow automation. The companies that win this phase will be the ones that can deploy agents into production at enterprise scale — not the ones with the highest benchmark scores.

Tencent’s Quiet Ascension

Tencent’s Hunyuan Hy3 model saw usage jump 68× in one week, topping OpenRouter global rankings. The company is internally integrating WorkBuddy and QClaw teams — consolidating enterprise and consumer AI agent efforts. Marvis, an OS-level AI assistant coordinating tasks across applications, rounds out a portfolio that is deeper than the Western press typically acknowledges.

Other WAIC Highlights

  • Agibot’s humanoid robot “远征 A3 Ultra” was selected as a WAIC “treasure” and entered mass production delivery.
  • StepStar launched STEPX Neo — the first large-model-native agentic phone running Step AOS, partnered with Meituan, Alipay, and Didi via A2A protocols.
  • Kuaishou’s AI video model spin-off aims to start a Hong Kong listing process within a year, as recurring revenue soars.

Regulatory Firsts: The World’s Most Advanced AI Governance

AI Agent Regulations — Three-Tier Authorization (Effective July 15)

China implemented the world’s first binding regulatory framework dedicated entirely to AI agents — the “Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents,” issued jointly by the CAC, NDRC, and MIIT.

The framework establishes a three-tier decision authorization system:

  • Tier 1 — Fully autonomous: Routine, low-stakes actions (data retrieval, summarization, scheduling). No human approval required at execution time, but classification must be documented.
  • Tier 2 — Meaningful but reversible: Sending external communications, modifying records, initiating workflows. Requires documented authorization and logging.
  • Tier 3 — High-consequence or irreversible: Financial transactions, external data transfers, sensitive personal data, physical systems. Requires explicit human approval before execution.

Organizations in high-risk sectors — financial services, healthcare, critical infrastructure — must file their agent frameworks with regulators before deployment. The regulations also mandate traceability, version control, human override mechanisms, and recall mechanisms.

This is the playbook that other jurisdictions have not started writing. The EU AI Act addresses AI systems broadly; China’s framework is the first to address autonomous agents specifically — and it is already in effect.

AI Companion Regulations — 345 Million Users Displaced

Also effective July 15, the “Interim Measures for the Administration of AI Anthropomorphic Interactive Services” — co-issued by five agencies (CAC, NDRC, MIIT, MPS, SAMR) — became binding after a three-month grace period.

The regulations target AI services providing “sustained emotional interaction” simulating human personality. Key provisions include:

  • Minors banned from virtual companion/relative services entirely.
  • Platforms must prevent emotional dependence and addiction, with detection and intervention mechanisms.
  • Crisis intervention: alert emergency contacts when users show distress signals.
  • Mandatory AI disclosure at the start of every session.
  • Private user conversations cannot be used for model training.
  • Security assessment required for services with 100K+ monthly active users or 1M+ registered users.
  • The CAC can shut down any service deemed unsafe; explicit government approval is required before public release.

The immediate impact was dramatic. ByteDance’s Doubao and Alibaba’s Qwen shut down AI companion and agent features on July 15. An estimated 345 million users lost access to AI companions overnight. Qwen deleted data immediately with no export option. Doubao gave users until October 15 for read-only data access.

Exempted from the rules: customer service bots, knowledge Q&A, workplace productivity assistants, and educational/research tools — provided they avoid sustained emotional engagement. This carve-out is strategically significant. China is regulating the social/emotional layer of AI while leaving the productivity layer largely untouched. The message to the industry is clear: build tools, not friends.

International AI Governance Action Plan

At WAIC, the NDRC and other departments released an Action Plan for AI Cooperation and Development with eight practical initiatives: high-quality data supply, inclusive computing power, open-source ecosystems, AI-enabled industrial transformation, talent cultivation, international standards, AI safety governance, and ethical development. China also released the MAZU intelligent weather early-warning system — the world’s first nationwide AI solution for the UN Early Warnings for All initiative, expected to deploy in 30 countries.

These are not abstract declarations. They are concrete deliverables attached to institutional infrastructure (WAICO, AICDN) and backed by the resources of a state that has made AI a national priority at the highest level.

ByteDance: Tesla Integration and the Consumer AI Play

Tesla’s China AI Swap

On July 24, Tesla began rolling out a dual-AI voice assistant for vehicles in China that replaces SpaceXAI’s Grok model (used in North America):

  • ByteDance Doubao handles direct vehicle controls — climate, navigation, media, owner manual — and supports 18 regional Chinese dialects.
  • DeepSeek handles conversational queries — general questions, weather, news.

This is a telling signal. Tesla, an American company, is swapping out an American AI model for Chinese models in its China-deployed products. The practical logic is straightforward: Chinese models perform better in Chinese-language contexts, support dialects that Western models cannot, and face no regulatory barriers in the Chinese market. But the strategic implication is larger — American companies operating in China are adapting to the Chinese AI stack rather than fighting it.

HMD Touch AI — Doubao at $70

HMD launched the Touch AI — a $70 Nokia Lumia-style phone with Doubao AI built in (China-exclusive). The device offers real-time Chinese-English translation, image generation, and text processing, with Doubao 2.0 running at approximately 10× lower cost than comparable Western models. This is the ultra-low-cost AI device play — bringing AI capabilities to price points that Western OEMs are not addressing.

The OpenRouter Revolution: Chinese Model Dominance

The top six models on OpenRouter by monthly usage are all Chinese: Xiaomi, DeepSeek, Tencent, MiniMax, and Zhipu AI. Chinese models dominate OpenRouter token usage, and approximately 60% of token usage attributed to US companies on OpenRouter comes from Chinese-developed models.

The cost differential is staggering. Comparing what might be called a “barrel of intelligence” for AI inference:

  • Anthropic: $56
  • OpenAI: $26
  • Meta: $1.50
  • xAI/Google: $1
  • Chinese models: $0.50

That is a 112× price gap between the most expensive and cheapest options. For any organization building AI-powered products at scale, this is not a marginal consideration. It is a fundamental architecture decision.

Washington has noticed. Axios reported that the US government has begun reviewing measures to bar US companies from using Chinese models — a direct response to Kimi K3’s launch. But restricting government use is one thing; restricting private enterprise from using the most cost-effective models available is another entirely, and the economic logic of the market will push hard against such restrictions.

Key Themes and What to Watch

  1. DeepSeek’s funding freeze is a feature, not a bug. Liang Wenfeng’s AGI-first posture and the decision to pause the second funding round suggest a company that is playing a longer game than the market expects. Watch for the IPO filing timeline and any shift in commercial positioning.

  2. The performance gap is functionally closed. At 6% and narrowing, the US-China AI performance gap is no longer a strategic moat for US companies. The competition is shifting from capability to cost, deployment, and ecosystem — and China has advantages in all three.

  3. Huawei’s chip sovereignty is validated. The Atlas 950 SuperPoD and Z.ai’s 1GW data center prove that the domestic chip stack works at exascale. NVIDIA’s China market share collapse is not temporary — it is structural.

  4. Regulation as competitive advantage. China’s AI agent and companion regulations are the world’s most advanced. They create compliance costs, but they also create clarity — and in global markets, regulatory clarity can be a competitive advantage over jurisdictions still debating.

  5. Open-source as geopolitical strategy is working. GLM-5.2’s MIT license, Kimi K3’s open weights, and DeepSeek’s open-source commitment are embedding Chinese models into the global developer ecosystem. The OpenRouter dominance and Hugging Face’s use of GLM-5.2 for cybersecurity are proof points.

  6. The enterprise agent wave is here. WAIC 2026 marked the pivot from model competition to commercial deployment. The winners of this phase will be determined by distribution, not benchmarks — and companies like Alibaba (via Apple Intelligence and DingTalk), Tencent (via WorkBuddy), and ByteDance (via Tesla integration) have massive distribution advantages.

  7. US guardrails are creating market opportunities for Chinese models. The Hugging Face cybersecurity incident is a case study. When US models refuse tasks due to safety restrictions, Chinese models fill the gap — and customers follow.

  8. China is building a parallel AI order. WAICO with 29 nations, the AICDN, Xi’s keynote, MAZU for the UN — these are the institutions of a systemic alternative, not just competition within an existing system.

The week of July 18–25, 2026 will be remembered as the moment China’s AI ecosystem matured from catching up to setting the pace. The models are at the frontier. The infrastructure is sovereign. The regulations are in force. The institutions are being built. And the market — from Tesla to Apple to Hugging Face to OpenRouter — is responding.

Follow @kkaminski on X for daily China AI analysis.

Big Hat Group Inc. provides AI consulting and early adoption strategy for enterprises. Contact us to learn how these developments affect your AI roadmap.