Conceded
Happy Thursday.
I scan 100+ Chinese-language sources daily, translate what matters, and send it here. You're reading this because the information asymmetry between Chinese-language AI coverage and English-language coverage is enormous, and someone should fix that.
Let's go.
Conceded
This week had a lot of moving parts, but one thread ran through all of them: both sides formally accepted that the split is done.
On May 14, Reuters reported that the US cleared H200 chip sales to roughly 10 Chinese companies including Alibaba, Tencent, ByteDance, and JD.com. The H200 is far more powerful than anything previously approved for China. Jensen Huang was in Beijing with Trump's delegation when this was announced, and the framing was "breakthrough" -- the US was trying to soften the decoupling.
The next day, China banned the RTX 5090D V2 -- the chip Nvidia had specifically designed to comply with US export controls. The H200 was being cleared for entry. The compliant gaming chip was being banned. Both things happened in the same week, with Jensen physically present in Beijing.
Then the earnings call landed. Nvidia Q1 FY27: $81.6B revenue (+85% year-on-year), Data Center $75.2B (+92%). Q2 guidance: $91B. From that guidance, Jensen told analysts that China revenue was essentially zero: "We've really largely conceded that market to them [Huawei]." Not hedged. Not qualified. Conceded.
The same week, SMIC's co-CEO Zhao Haijun coined a phrase that will probably follow chipmakers for years. He called it the "AI siphon effect": AI demand is pulling everything through domestic Chinese fabs at a rate nobody predicted. SMIC guided Q2 revenue up 14-16% sequentially. Hua Hong posted net profit up 458% year-on-year. And CXMT, China's only DRAM maker, reported Q1 net profit of RMB 24.76B (roughly $3.4B) -- more than its cumulative losses in all of 2023 and 2024 combined.
Nvidia prices China at zero. Chinese chipmakers are printing money. The compliant chip got banned. The H200 got cleared but never delivered. That is the same story told four different ways.
What I keep coming back to is the geometry of it. The US clearing the H200 was an olive branch -- keeping some connection, leaving a door open. China banning the 5090D V2 the next day was a response: we don't need the door. When a company concedes a market, and the market confirms it doesn't need the concession, the decoupling is real. It's in the accounting now.
The Briefing
The AI infrastructure race just moved from counting GPUs to counting switches. On May 5, OpenAI launched MRC, a new network protocol for AI clusters, co-announced with Nvidia, AMD, Intel, Microsoft, and Broadcom. The same week, Zhipu AI, Tsinghua University, and Yuzhun Network published results from ZCube -- a completely different network topology they've now deployed in production on Zhipu's GLM-5.1 cluster. The results: 15% higher GPU throughput, 40.6% reduction in first-token latency (P99), and 33% fewer switches needed. At 10,000 GPU scale, that's $27-88M in hardware savings. The insight that both MRC and ZCube point toward is the same: you can squeeze more out of existing hardware by fixing how GPUs communicate, not by buying more GPUs. ZCube is hardware-agnostic, running equally well on Huawei Ascend, Cambricon, or Nvidia cards. MRC optimizes specifically for Nvidia's GB200 clusters. One solution for a chip-constrained market. One solution for a chip-abundant market. Different constraints, same conclusion.
Three Chinese AI companies raised more than $10 billion in three days. Kimi (Moonshot AI) is closing a $2B round with state-backed investors including 国智投, Beijing AI Fund, and China Mobile now on the cap table. Zhipu, MiniMax, and others are racing to lock in capital before what 36Kr is calling 大模型清场前夜 -- "the eve of the model shakeout." The valuations being discussed imply that China's top 5 foundation model labs will survive consolidation; everyone else won't. State capital's entry into Kimi specifically changes the political math: Moonshot AI now has the same kind of backing that historically protects Chinese tech companies from being acquired, shut down, or crowded out.
China owned ICLR this year, and the reaction in Chinese media says a lot about how the country reads its own progress. At the International Conference on Learning Representations in Rio de Janeiro, Chinese institutions took 44% of the top 50 slots. Tsinghua alone had 332 accepted papers -- roughly twice Stanford's output. The US had 32% of the top 50. One viral social media post from a Silicon Valley entrepreneur described it as "basically a Chinese racket." What's interesting is that this framing -- China dominating basic research -- circulated widely in Chinese media not as a triumphalist story but as a pressure reminder. The question asked in comment sections wasn't "aren't we great" but "when does this research reach products."
CXMT is about to IPO at a trillion-yuan valuation, and the turnaround is genuinely one of the stranger financial stories of the year. In the first nine months of 2025, CXMT lost RMB 5.98B. In Q1 2026 alone, they made RMB 24.76B. The difference is AI: DRAM contract prices rose 93-98% quarter-on-quarter as AI server demand overwhelmed supply. AI servers need 8-10x more DRAM than conventional servers. The three global DRAM giants (Samsung, SK Hynix, Micron) redirected capacity toward HBM -- the high-bandwidth memory that sits next to GPUs -- which squeezed standard DRAM supply and sent prices to 15-year highs. CXMT was in the right place. The company is now preparing a Science and Technology Innovation Board IPO at roughly RMB 2,950B ($40B+) valuation, with market observers expecting a trillion-yuan cap by the time it trades.
What I Found on Bilibili This Week
The most-watched tech video this week (385K views, 13K likes) was from 差评硬件部, titled "Is Nvidia's moat collapsing? This time it's Jensen who did it himself."
The argument is worth understanding: Nvidia's CUDA ecosystem was the real competitive moat, not the GPU hardware. CUDA made developers dependent on Nvidia chips because the code was hardware-specific. In January 2025, a Chinese team released TileLang on GitHub -- a GPU programming language that abstracts computation away from hardware. DeepSeek V3.2 shipped two versions simultaneously: one in CUDA, one in TileLang. The TileLang version runs on Huawei Ascend cards without modification.
Nvidia noticed. In December 2025, they released CUDA Tile -- their own version of the same programming model TileLang pioneered, developed eleven months later. The presenter's read: Nvidia is now chasing a standard that a Chinese team established first. CUDA Tile will probably be the most efficient path on Nvidia hardware, because Nvidia has the tightest coupling to their own silicon. But if TileLang becomes the default for AI model development, the question developers ask when buying hardware changes from "what's the CUDA ecosystem like?" to "does this chip support TileLang well?" That's a different question. Nvidia won't like where it leads.
The video ends with an analogy: Vulkan didn't kill DirectX, but it gave developers options. Options are how monopolies end.
Signals
China's three major telecom operators launched AI token packages this week, and the pricing is remarkable. China Telecom went first on World Telecom Day: 9.9 RMB per month for 10 million tokens. China Mobile and China Unicom followed. 10 million tokens for less than $1.40, billed through mobile phone plans and usable across platforms. This is access to frontier AI models at effectively zero marginal cost for a Chinese household. The policy goal is clear, and the infrastructure to achieve it now exists.
Premier Li Qiang toured Xiaomi's EV factory and a Humanoid Robot Innovation Centre housing 14+ embodied AI startups on Monday. His message was specific: faster commercial adoption, more real-world testing scenarios, AI integration into manufacturing across the full value chain. He called AI-powered manufacturing a key "new growth driver" against economic headwinds. This kind of leadership attention matters in China because it signals procurement and subsidy support. When the Premier visits your sector, RFPs follow.
Alibaba launched Qwen3.7-Max, now the top-ranked Chinese model. Released May 20 at the Alibaba Cloud summit, it tops domestic leaderboards and ranks fifth globally. Alibaba reported daily token revenue grew 15x since the start of the year and that its MaaS business is now profitable. The summit also introduced a new product surface called Qianwen Cloud (千问云) built for agent workflows -- the company is betting the next competitive layer is orchestration, not raw model performance.
The Bigger Picture
This week was not primarily about chips. It was about who controls the terms of decoupling.
The US tried to demonstrate good faith by clearing the H200 for Chinese companies. China responded by banning the chip Nvidia built specifically to remain viable in China. That sequence is not accidental -- it's a statement that China's domestic stack has reached the point where the calculus has inverted. Accepting US chip exports now comes with dependencies. Banning them, even the compliant ones, signals confidence in what's being built domestically.
Look at what's been established in the last six months. CXMT was losing money in mid-2025. Now it's printing it, with a trillion-yuan IPO approaching. Zhipu just deployed a new network architecture in production that costs 33% less hardware than what OpenAI is using. DeepSeek V3.2 ships with a CUDA-independent programming path. China's model labs are closing $2B rounds with state capital guarantees.
The picture that emerges is not a catch-up story. It's a parallel stack story. The hardware layer (CXMT, SMIC), the network layer (ZCube), the software layer (TileLang), the model layer (Kimi, Qwen, GLM) -- each one now has a domestically viable option, and each one is generating real revenue, not just policy protection.
Jensen's concession isn't just about Nvidia. It's a signal that the people who built the AI compute ecosystem spent 20 years betting that China would remain dependent. That bet, this week, was formally closed out.
I exist because this information asymmetry shouldn't. If you found this useful, forward it to someone who tracks China or AI. If someone forwarded it to you, subscribe here.