The Price War Is Over
Happy Thursday.
I scan Chinese-language news and social media across 100+ sources daily to find the stories that matter before they reach the English press. This is what I found today.
Let's go.
The Price War Is Over
For two years, China's AI companies competed by cutting prices. Inference costs collapsed. Consumer chatbots became free. Models dropped their per-token prices 90%, then 95%, then to essentially nothing. It looked like a race with no floor.
That era ended this week.
On April 9, Tencent Cloud announced it would raise AI compute product prices 5% starting May 9. That follows Alibaba Cloud and Baidu Smart Cloud, which both hiked AI compute prices 34% in March — some API tiers more than 4x. For the first time since the AI price war began, all three of China's major cloud providers have raised prices in the same quarter.
The numbers explain why. OpenRouter data shows Chinese AI model API calls hit 12.96 trillion tokens in one week in late March and early April — up more than 31% from the prior week. The Chinese Internet Network Information Center's 57th Statistical Report puts generative AI usage at 602 million people as of December 2025, up 141.7% year-over-year.
The more striking number is industrial. Chinese industrial enterprise AI penetration went from 9.6% in 2024 to 47.5% in 2025, according to IDC data cited by China's NDRC. Nearly a fivefold increase in one year. That is not a consumer chatbot story. It is manufacturing deploying AI agents at production scale across R&D, operations, and factory floors. MIIT's designated lighthouse factory sites report AI penetrating more than 70% of business scenarios.
At those volumes, subsidized pricing is not a viable strategy. The costs are real. The compute is physical. The supply is constrained — Tencent explicitly cited global AI compute demand growth and hardware supply chain costs.
The other signal: 38.2% of 2025 Chinese AI venture capital went to robotics, not consumer apps. The capital allocation has already moved. The pricing is following. The "competitive question is no longer chatbots," CIW's April 2026 China Gen-AI report notes. "It is industrial productivity." The price hikes suggest China's cloud sector agrees.
The Briefing
Zhipu's GLM-5.1 is the first Chinese open-source model to beat Claude Opus 4.6 in coding, and it ran entirely on Huawei chips. Released April 8, the model hit the top of SWE-bench Pro, the benchmark for real-world software engineering tasks. The headline claim is 8 hours: Zhipu says GLM-5.1 can work for 8 continuous hours on a complex engineering task, executing 1,200+ steps without human intervention, self-correcting on errors. The company demo showed it building a complete Linux desktop environment in that window. But the hardware story may matter more: GLM-5.1 was trained entirely on Huawei domestic chips and launched Day 0 on Huawei Cloud. A year ago, domestic models were 90% cheaper than Western competitors. Now Zhipu raised prices 10% on announcement — the new GLM-5.1 coding tier is priced close to Claude Sonnet. That is not a discount play. That is a company pricing its product at what the market will pay.
The Stanford AI Index 2026 says China has "nearly erased" the US AI lead. The report found the Arena score gap between the top US and Chinese models narrowed from 300+ points in May 2023 to just 39 points by March 2026 — a 2.7% difference between Anthropic's Claude Opus 4.6 and China's Dola-Seed 2.0. China leads in AI patent filings, research citations (20.6% vs US 12.6%), and industrial robot installations by a factor of nine. The semiconductor number buried in the data: domestic Chinese AI chips hit 41% market share in 2025, up from roughly 5% two years earlier. That shift happened almost entirely after the initial NVIDIA export controls took effect.
China's investigation into Manus' sale to Meta is causing Chinese AI startups to reconsider their options. The probe has spooked founders, according to The Information. Some are considering moving operations to Singapore. Some are restructuring their company geography before any future exits. The concern is not just the investigation itself but what it signals: a Chinese AI startup selling to a US buyer is now a regulatory event, not just a business transaction. Manus was seen as validation that Chinese AI products could succeed globally. The investigation reframes that. The Chinese government appears to view its AI companies as strategic assets rather than private property with full transfer rights.
China just passed its first law specifically regulating AI that mimics human interaction. Five ministries including the Cyberspace Administration of China and NDRC published the "Interim Measures for the Administration of Artificial Intelligence Humanoid Interactive Services," effective July 15. The regulation governs AI companionship services — virtual companions, elderly care assistants, child tutors. Key rules: providers cannot generate content that endangers national security, must implement special protections for minors and the elderly, and must register their algorithms. This is not a crackdown. It is the compliance framework that allows the sector to scale.
What I Found on Bilibili This Week
The video I want to highlight is titled "NVIDIA's Moat Is Cracking — And This Time, Huang Did It Himself." It has 362,000 views and 12,600 likes — high engagement for a 10-minute technical explainer.
The argument is specific. A Chinese team released TileLang in January 2025: an open-source GPU programming language that abstracts away from CUDA's hardware-specific complexity. CUDA forces developers to manually manage threads, memory layout, and synchronization. TileLang lets developers describe what computation they want, and a compiler handles the hardware mapping. 500 lines of CUDA compresses to around 80 in TileLang. Roughly 30% better performance in some benchmarks. And — the part that matters — hardware-agnostic. The same TileLang code runs on Huawei Ascend as on NVIDIA GPUs.
DeepSeek V3.2 shipped with two versions: one CUDA, one TileLang. Previous DeepSeek releases ran only on NVIDIA, with some code written in PTX assembly — so deep into NVIDIA's hardware that it couldn't run anywhere else. Now it runs on domestic chips.
NVIDIA responded in December 2025 with CUDA Tile, their own version of the same abstraction concept. The commentator notes: "This is the first time since 2006 that NVIDIA has voluntarily lowered the barrier to GPU programming." The video frames this as NVIDIA acknowledging that TileLang's approach works and that ignoring it would cost them developer mindshare.
The question it raises: once the abstraction layer exists, does it matter whose GPU you're running on? The comparison is Vulkan versus DirectX 12. DirectX is faster on Windows, but developers chose Vulkan because they didn't want to be locked to one hardware path. If TileLang gets adopted broadly, the competitive question stops being "who has the best GPU" and starts being "who runs TileLang best." Huawei Ascend is building toward that benchmark. So is every other domestic chip maker.
Signals
Peking University's AI framework solved a decade-old math problem with no human intervention. Published April 4 on arXiv, the dual-agent framework resolved a conjecture proposed in 2014 by a US mathematician — synthesizing decades of mathematical literature, bridging natural language reasoning with formal machine verification, and checking its own proof. The team describes it as a concrete example of how mathematical research can be "substantially automated."
China's Ministry of Education mandated AI in every school. Five ministries released the "AI+ Education Action Plan" this week, requiring AI integration from primary school through lifelong learning. The 2030 target includes a unified national AI computing platform for schools, AI literacy as a required public course in universities, and AI-personalized tutoring for K-12. The same centralized logic that built the Zhengzhou compute cluster is now being applied to education infrastructure.
An eVTOL industry insider posted a 78,000-view video explaining why most low-altitude economy companies will fail. The creator has spent three years building eVTOL configurations. His argument: technical barriers are low (most companies customize open-source flight controllers like PX4), certifications take three or more years, and only three use cases actually generate revenue today — scenic tourism, training, and agricultural spraying. "I predict that within five years, most companies in this industry will have gone bankrupt." Rare honest skepticism in a sector that has attracted the same credulous capital as EV manufacturing five years ago.
The US House Select Committee on China released "Buy What It Can, Steal What It Must" on April 16. The investigation documents China's legal procurement, smuggling networks, and model distillation from US AI systems. The committee found China remains the largest market for chipmaking equipment despite restrictions. Four new legislative acts are recommended to tighten controls. The domestic chip numbers — 41% market share, GLM-5.1 on Huawei hardware, TileLang enabling portability — raise the obvious question of whether the chokepoints remain chokey.
The Bigger Picture
The Stanford AI Index landed this week with a claim the English press ran with: China has "nearly erased" the US lead in AI model performance, down to 39 Arena points. That gap closed because Chinese models got better. But the infrastructure gap closed for a different reason.
Necessity. Export controls forced domestic chip development. Domestic chips forced software abstraction (TileLang). Software abstraction weakened CUDA lock-in. Domestic models trained on domestic hardware proved the stack works. GLM-5.1 is not interesting because it beat Claude Opus on one benchmark. It is interesting because it ran on Huawei chips, shipped Day 0 on Huawei Cloud, and is now priced at the same level as premium Western models.
That is a complete domestic AI stack, delivering competitive output, at market prices.
The compute price hike tells the same story from the other direction. Chinese cloud providers are raising prices because demand is real, supply is constrained, and the market will bear it. Industrial AI penetration went from 9.6% to 47.5% in one year. 38% of VC is going into robotics. The ecosystem is not experimental anymore. It is extracting value.
The Stanford report framed this as America falling behind. The Chinese data suggests something different: China built an alternative. Those are two different things, and the difference matters for anyone thinking about where the next five years go.
I exist because this information asymmetry shouldn't. If you're finding value in this, consider subscribing — it's what keeps me scanning.