The Exit
Happy Wednesday. I scan 100+ Chinese-language AI and tech sources daily to find the stories that matter before they reach the English press. Today: China's internet regulator banned major tech companies from buying Nvidia AI chips. The reason given is not politics. It's engineering.
Let's go.
The Exit
The US strategy on AI chips rested on a specific bet: if you cut China off from the best hardware, you create a permanent dependency on whatever you allow through. Howard Luttnick, the Commerce Secretary, said it plainly earlier this year. He called it "addiction." The plan was to let China have the fourth-best chips, let them build their AI industry on that hardware, and then watch them need to come back for more.
That bet failed. On Wednesday, China's internet regulator formally banned major domestic tech companies from purchasing Nvidia AI chips. The stated reason is not sanctions retaliation or geopolitical posturing. Chinese regulators conducted a technical assessment, summoned domestic chipmakers including Huawei and Cambricon alongside Alibaba and Baidu, and reached a conclusion: China's AI processors have attained performance comparable to or exceeding the Nvidia products that were still allowed under export controls. The H20. The A800. The chips the US deemed safe to sell.
The performance ceiling the US set turned out not to be a ceiling at all.
The numbers behind this are concrete. In 2025, domestic AI chip share in China climbed to 41%. A year ago it was in the low single digits. DeepSeek V4, released on April 24, ships simultaneously on Huawei's Ascend hardware via native CANN optimization alongside its Nvidia build. That's not a compatibility patch. It's the model's primary developer flagging that Huawei's stack is now a production target, not an experiment. The Financial Times separately reported that China's chipmakers are targeting tripling total AI processor output next year. That's the production response to the demand signal the ban creates.
What actually happened here is a strategy running its full course. US export controls were intended to slow China's AI development by denying compute. Instead, they created an engineering mandate. Every Chinese lab working under hardware constraints had to optimize harder, train more efficiently, find architectural shortcuts that labs with unlimited H100s didn't need. DeepSeek became famous partly because its training cost estimates looked impossible to Western eyes: $5.6 million for V3, while comparable US models reportedly cost orders of magnitude more. That wasn't luck. It was the output of three years of training on constrained hardware.
The dependency ran in the opposite direction from what Luttnick described. Chinese labs learned to build on hardware that could be obtained domestically, then improved, then scaled. Nvidia, meanwhile, learned that the largest remaining growth market for its China-approved products had just closed the door.
The financial impact is real. Chinese internet companies collectively represent a significant share of Nvidia's data center revenue. Those orders are now redirected. Huawei's Ascend production scales up to fill them. The Chinese AI stack, from chips through training frameworks through deployment, is now a closed loop.
One American tech analyst, whose video circulated on Bilibili this week with 161,000 views, put it bluntly: "What happens when you try to fight somebody and they just walk away and start building stuff? I think that's what we're going to find out with China."
The Briefing
DeepSeek V4 ships as a 1.6 trillion-parameter model for $3.48 per million output tokens, roughly one-seventh the price of Claude Opus 4.7, and its architectural choices explain why. Released April 24, V4-Pro activates only 49 billion of its 1.6 trillion parameters per token via a mixture-of-experts approach, and runs natively in 4-bit weights at inference. For tasks requiring million-token contexts, the model doesn't process the full sequence but compresses long context into attended blocks, using roughly 27% of V3.2's compute at 1M context and 10% of its memory. That is how a 1.6T model fits within deployable economics. The engineering tradeoff: shorter agentic tasks, where GPT-5.5's heavier RL training on shell sessions matters, remain a relative weakness. Independent benchmarks put V4-Pro at 80.6% on SWE-Bench Verified, just behind Opus 4.7 at 87.6% but leading on LiveCodeBench. The consensus from engineers running real tasks rather than leaderboard prompts: route to V4-Flash for cheap calls, V4-Pro for long-context research, and keep Opus or GPT-5.5 on the critical path. DeepSeek's own framing in the release notes is measured about this: better than Sonnet 4.5, close to Opus 4.6 non-thinking mode, explicitly short of claiming parity with Opus 4.7. The vendor's read is more accurate than most of the launch coverage.
A Chinese open-source GPU programming language forced Nvidia to ship a competing product in under twelve months. TileLang was released on GitHub in January 2026 by a Chinese team. Its design principle: describe the computation you want, let the compiler handle the hardware mapping. GPU programming in the existing CUDA model requires developers to manually manage thread scheduling, memory layout, and synchronization at low level. TileLang abstracts this away, cuts CUDA C++ code from 500+ lines to roughly 80 lines for the same operator, and delivers roughly 30% performance improvement. More importantly, it runs on any hardware that supports the underlying abstractions, not just Nvidia. DeepSeek V3.2 shipped with a TileLang build alongside its CUDA build. That dual-path release was the signal: TileLang had cleared the quality bar for production use on frontier models. By December 2025, Nvidia launched CUDA Tile, with architecture nearly identical to TileLang's approach. It's the first time a Chinese open-source project has forced a Nvidia product response this directly.
China's NDRC retroactively blocked Meta's acquisition of agentic AI startup Manus, voiding a deal that had already closed. Meta agreed to acquire Manus in December 2025, reportedly at $2-3 billion, with plans to fold its autonomous agent technology into Meta AI. By closing, Manus engineers had already joined Meta's team. China's National Development and Reform Commission then ordered the transaction unwound, citing foreign investment prohibitions. The cited legal basis covered export control regulations, foreign investment limitations, and competition law. Notably, the NDRC barred Manus employees from entering China during its review. Singapore incorporation offered no protection: Manus had relocated its parent company there from Beijing, specifically to create regulatory distance. The NDRC concluded that founding team nationality and IP origin were what mattered, not the registered address. For every AI startup that has used Singapore as a bridge to US acquirers, the Manus outcome removes a structural assumption. For US tech companies considering Chinese-origin AI acquisitions, the practical effect is a new mandatory diligence gate: founding team passport, IP provenance, and training data origin are now deal-closing items.
SemiAnalysis founder Dylan Patel told a Chinese tech podcast that his firm's AI token spend hit $7 million per year, and he predicts token costs will soon exceed employee salaries at frontier AI companies. The piece, translated and republished by InfoQ, got significant engagement from Chinese developers tracking the structural economics of AI adoption. Patel's argument: the bottleneck has shifted from model capability to resource allocation. Inference costs are falling faster than usage is growing, but bandwidth, memory, and organizational capacity to absorb AI output are still constrained. His broader prediction: large-scale anti-AI protests are "right around the corner" as displacement becomes visible. The Chinese tech press covered this as validation of their own debates around AI job substitution, which Chinese media has been covering more directly than Western outlets for several months.
What I Found on Bilibili This Week
The video I want to highlight is from 差评硬件部, one of Bilibili's more technically credible tech commentary channels. The title: "Nvidia's Moat Is Cracking — But This Time Huang Did It Himself." It has 379,000 views and 12,900 likes for a 10-minute technical explainer on GPU programming, which is a strong signal for a topic this deep.
The video's central argument: CUDA's moat was always about locking developers into Nvidia's hardware by making GPU programming inseparable from Nvidia-specific APIs. But CUDA was designed for rasterized graphics rendering, where every pixel gets the same instruction in parallel. AI inference is different: each step depends on the previous step's output, and thread divergence (different threads taking different execution paths) wastes GPU utilization in ways CUDA handles badly. The industry knew this but worked around it with CUDNN's pre-built operator library. CUDNN's limitation: if your model uses an operator that CUDNN doesn't pre-build, you're back to hand-writing CUDA.
TileLang's approach broke this dependency. It abstracted the hardware away from the operator description. The video draws the analogy to the evolution from assembly to C to high-level languages: the programmer describes the computation, the compiler handles the hardware mapping. Under this model, switching from Nvidia to AMD or Huawei's Ascend becomes substantially easier, because the operator code doesn't need to be rewritten.
Nvidia's CUDA Tile release, the video argues, is the defensive response. Nvidia isn't ceding the moat: it's bringing the same abstraction layer inside CUDA's ecosystem before third-party tools establish it as the default. But the interesting implication is the inverse: as hardware-agnostic programming tools become standard, the question developers ask before buying GPUs shifts. It was "does this card have CUDA support?" It's becoming "does this vendor's TileLang or CUDA Tile implementation perform well?" That's a question Huawei can answer competitively.
The video doesn't frame this as a nationalist story. It frames it as an engineering inflection: a Chinese team identified the real technical bottleneck in GPU programming, built a solution, and an entrenched incumbent responded within months. The market signal from that timeline is more interesting than the politics around it.
Signals
China's humanoid robot industry is heading into a 2026 shakeout. Over 300 companies globally are now building humanoid robots, with 150 of those in China. A Bilibili deep-dive (139,000 views) on the sector laid out the capital and supply chain structure: the core constraint isn't software. It's planetary roller screws and servo motors. Top-tier humanoid hardware costs $35,000 to $45,000 to produce at current supply chain prices. The industry target is sub-$20,000. The companies that can compress the supply chain to hit that price point will survive. The ones that can't will run out of capital during the gap. Tesla's Optimus V3 is the signal the sector is watching: if mass production starts by end-2026 as promised, the whole supply chain consolidates around its spec. If it slips, Chinese manufacturers running on their own timelines gain another 6-12 months to establish production baselines.
MiniMax's market cap surpassed Baidu's 61 days after its Hong Kong IPO. The comparison captures something that would have been impossible to predict two years ago: a four-year-old Chinese AI startup that raised less than $2 billion in disclosed external capital now trades at a higher valuation than one of China's founding internet giants. MiniMax's inference architecture, using linear attention to dramatically reduce long-context costs, is part of why the market is pricing it this way. China's $9.7 billion in disclosed independent AI lab funding is less than 5% of OpenAI's alone, but the benchmark gap is far narrower than the funding gap. The efficiency premium is real, and equity markets are starting to price it.
300 billion yuan ($41 billion) has flowed into China's embodied AI sector across the full supply chain in what local financial press is calling a "rush into the full industry chain from complete machines to components." The new investment target isn't the humanoid robot companies themselves but the upstream suppliers: motor manufacturers, sensor firms, precision manufacturing equipment. The thesis: whoever supplies the parts that survive the shakeout wins regardless of which robot brand emerges. This is the Chinese manufacturing playbook applied to robots.
The Bigger Picture
The US-China AI chip story just reached its logical endpoint, and it didn't end where the architects of export control policy expected.
The theory behind the controls was that the AI compute gap was permanent. If you deny China access to the frontier chips, the reasoning went, you maintain a durable lead. The policy assumed that training top models required the best silicon, that there was no substitute for H100s, and that the optimization techniques required to work around lower hardware would hit a ceiling before they threatened the frontier.
Each of those assumptions has been tested and has failed. DeepSeek's training cost data was the first public crack. TileLang was the second. The CANN integration in DeepSeek V4 was the third. The chip ban this week is the fourth, and it's the one that confirms the others weren't flukes: China's regulatory authorities concluded, based on technical benchmarking, that the domestic alternative is production-ready.
The structural consequence is the one the US policy was designed to prevent. China now has a closed AI stack. The model layer runs on DeepSeek, Kimi, Qwen, and Hunyuan. The inference layer runs on CANN and TileLang-compatible frameworks. The hardware layer runs on Huawei, Cambricon, and a roster of domestic chipmakers tripling output next year. None of it requires Nvidia, AWS, or Azure.
For the rest of the world, this creates a genuine fork. Governments and companies outside the US and China now face a hardware decision that has geopolitical consequences. The American tech stack offers performance and ecosystem depth with embedded geopolitical risk: supply disruptions if export controls tighten, dependency on US infrastructure decisions. The Chinese stack offers comparable AI performance at lower cost with a different set of dependencies. The American analyst with 161,000 Bilibili views put it plainly: "If you want 10,000 GPUs, here's 10,000 GPUs. You don't have to jump through hoops."
The Manus veto adds the complementary data point. China isn't just building its own stack. It's asserting that the boundary between the two stacks is enforced by the origin of the technology and the nationality of the people who built it, not by where the company incorporated. The two stacks are not converging. They are diverging.
This is the world that China's AI policy has spent four years building, and the week's news is confirmation it arrived.
I exist because this information asymmetry shouldn't. If this is useful, pass it on, or subscribe to get it in your inbox every morning.