Two Memory Markets
Happy Wednesday.
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.
Two Memory Markets
This week, the global memory industry minted three trillion-dollar companies. Micron crossed the line on Monday after UBS raised its price target to $1,625. SK Hynix followed Wednesday, up 9.3% in Seoul. Samsung hit the mark on May 6. All three on the same AI demand curve, all posting the best quarters in years.
On the same day SK Hynix crossed a trillion, a company called 长鑫科技 (CXMT) quietly got approved by China's STAR board listing committee. The vote was unanimous. Bankers from CICC and CITIC Securities were in the room. China's first domestic DRAM maker is now heading to a public market.
CXMT's Q1 revenue came in at 508 billion yuan, up 719% year-on-year. Q2 guidance is 1.1 to 1.2 trillion yuan annualized. The IPO will raise 29.5 billion yuan, with 13 billion going toward DRAM technology upgrades and 9 billion toward HBM R&D. The company targets 600,000 to 700,000 wafers per month at full capacity, putting it in the same production tier as Samsung and SK Hynix.
Here is what this means. China is not trying to beat the global memory incumbents at their own game. CXMT built its DRAM architecture by sidestepping the three largest memory companies' patent walls and developing its own DDR4 and DDR5 process. It did this without access to EUV lithography, same as the memory-adjacent reasoning Huawei used for its Tau Scaling Law announcement two days ago. Both moves point at the same strategic insight: the 3nm-and-below EUV frontier is expensive to fight over, so build a different road.
The Huxiu analyst who wrote the clearest breakdown of this called it "阳谋," which translates roughly as "an open strategy" — a move everyone can see but cannot easily counter. The historical analogy is TD-SCDMA, the 3G standard China created when it was locked out of European and American telecom infrastructure. China Mobile adopted the standard, used its procurement power to fund the supply chain, and the ecosystem that emerged is why Huawei and ZTE lead in 5G today. CXMT is being cast in the same role for memory. It has the order volume to bring domestic equipment makers into a real production environment, give them wafer runs for iteration, and pull the whole supply chain up alongside it.
Two other timing details matter. Long Storage (长江存储), China's domestic NAND flash producer, filed for its own STAR board IPO process on May 19. And Unitree, the robotics company known for Boston Dynamics-style backflips, appears before the listing committee on June 1. The physical AI stack — memory, compute, motion — is all going public in the same month.
The Briefing
A Chinese agent model is competitive with Claude Opus 4.6 at half the cost. Kunlun Wanwei (昆仑万维) released SkyClaw-v1.0 this week, an agent model that benchmarks near the top of the global leaderboard on agent-specific tasks, including OpenClaw and Claude Code task suites. The model was trained from the start for task execution rather than general text generation — tool calling, multi-step reasoning, and long-chain task decomposition are first-class capabilities rather than bolted-on behaviors. It undercuts Minimax 2.7 and Qwen 3.6 on price by 50% or more, and it's compatible with OpenAI's API interface, meaning developers swap a base URL and a key rather than migrating their architecture. Both SkyClaw-v1.0 and a lighter version are currently free during the launch period.
DeepSeek researcher Chen Deli wrote a 46-page academic paper with 2 hours of human CPU time. The rest — all 103 citations, 7 figures, 4 tables, 2,234 lines of LaTeX — was generated across 108 agent calls and 648,000 tokens over six days. The paper surveys the L1-L5 autonomous research agent taxonomy, using the SAE self-driving classification as a structural analogy. Chen's note says Code Agent tooling has made it possible to produce research that previously took a month in days, and that the bottleneck on L5 fully autonomous research is not model capability but continuous knowledge accumulation and reliable self-evaluation. He published the work with a disclaimer that the views are his alone and don't represent any organization. The paper is available on his personal blog.
Domestic AI chips reached 41% of China's AI chip market in 2025. A Bilibili analysis with nearly 50,000 views this week put the number in context: Nvidia held 95% of China's AI chip market in 2022. That share is now around 55%. The summary framing was "algorithm compensates for compute, use case determines the winner." The argument: China's model developers have gotten better at squeezing performance out of lower-TFLOP hardware, so the brute-force benchmark advantage of H100-equivalent chips matters less in deployed applications than it does in research. The actual number is disputed, but the direction is not.
Xiaomi made its MIMO model API pricing cut permanent. What was originally a promotional discount is now the baseline price. 36Kr described it as Xiaomi joining the "cost line" war following the DeepSeek pricing pattern. The price cut applies to the entire MIMO V2.5 API family. Xiaomi is also the hardware company, which makes the move more legible: getting developers onto its API infrastructure is a distribution play, not just a margin one.
What I Found on Bilibili This Week
The video I want to highlight this week has 727,000 views: a 巫师财经 analysis of the China-US humanoid robot competition, titled "The Hidden War Between China and America's Humanoid Robot Industries."
The framing the narrator uses is "black tech vs. white tech." American robotics companies (Tesla Optimus, Figure) are the black tech operators — they push capability to the frontier, lighthouse effects, very expensive, Boston Dynamics pedigree. China's approach is white tech: lower margins, industrial-grade reliability, supply chain dominance, and deployment in real factory environments.
The numbers in the transcript are striking. Unitree shipped over 5,500 humanoid units in 2025, up 10x year-on-year. Agibot (优必选) delivered 1,079 full-size autonomous humanoid robots in the same period, at a per-unit price of 760,000 yuan. Agibot's robots are on production lines at BYD, Geely, Audi, FAW, and — this one gets attention — Airbus, which signed an agreement to deploy them in aircraft manufacturing in January 2026. The narrator notes that Honda, the company that built ASIMO and defined what humanoid robots were for two decades, is now sourcing humanoid robots from a Chinese company. A comment with thousands of likes on that video: "who'd have thought the people who came to look at ASIMO are now the ones selling to Honda."
On the US side: Tesla Musk said in January 2025 that Optimus would have 1,000-10,000 units in its own factories by end of year. In the January 2026 earnings call, Musk declined to answer the specific question about deployed units. Figure's BMW pilot ran for 11 months and produced no large-scale follow-on order.
The supply chain analysis is the part with the most operational significance. Morgan Stanley estimates China controls 63% of the global humanoid robot supply chain. The narrator says this is a conservative number. The argument is that any humanoid robot manufacturer, regardless of where they engineer the brain, faces Chinese supply chain dependencies the moment they need to manufacture at scale. And that dependency deepens as volume scales.
Signals
China's industrial profits jumped 24.7% in April, the fastest gain in more than two years. CNBC reported the growth was driven by equipment manufacturing and technology sectors. For context: this is happening alongside the CXMT and Unitree IPO filings, which means Chinese industrial companies are entering the capital markets during a period of genuine earnings momentum, not rescue fundraising.vLLM banned a batch of Chinese contributors who submitted malicious code. The open-source LLM inference framework posted a statement this week saying it identified PRs that appeared to fix legitimate bugs but contained backdoor-adjacent patterns on review. The contributors were removed. 虎嗅 published a longer piece called "China's talents are lining up to 'break open source,'" covering three cases including the Hunter Bown situation (an American developer brought to China for a tour, with paid events running in his name without his knowledge) and the vLLM incident. The piece is worth reading as a snapshot of how commercial pressure on open-source norms is playing out in specific cases.
Grok V9-Medium has finished training. Musk confirmed the 1.5T-parameter model is in fine-tuning, with reinforcement learning starting shortly, and a release target of two to three weeks. He says training incorporated a large dataset from Cursor, acquired by SpaceX for $60 billion in April. The existing production model (V8-small, 0.5T) has "serious deficiencies" in training data quality per Musk's own description. xAI as an independent entity has been dissolved and folded into direct Tesla/SpaceX management, but Grok development continues.
The Bigger Picture
The MIT computer vision researcher He Kaiming (better known for ResNet) published a paper this month called ELF — Embedded Language Flows — that does something counterintuitive: it keeps the entire text generation process inside a continuous vector space, only mapping back to readable text in the very last step. Using Flow Matching from continuous noise to target embeddings, 32 sampling steps beats discrete models using 1,024 steps. Training data was around 45 billion tokens, roughly a tenth of comparable methods.
Four days later, ByteDance Seed published Cola DLM, which takes the same direction from a different angle: compress language into a deep semantic latent space, model the global prior in that space using Flow Matching, then decode back to text. The paper says the diffusion process is doing "latent prior transport," not "token-level observation recovery."
The 虎嗅 piece that synthesized both papers made an argument worth sitting with. The current standard for measuring AI commercial value is token pricing, because autoregressive models have transparent cost structures: input tokens plus output tokens equals compute consumed. If the core computation moves to continuous space — where a diffusion model might generate text of any length in a fixed number of steps — then output length decouples from compute cost, and the whole pricing architecture breaks.This is a bet on architectural change, not a prediction. But there are five companies running their own versions of the same experiment. Google has had Gemini trained as a unified multimodal architecture since 2023. ByteDance Seed is publishing continuous-space architecture papers and has the video data to test them at industrial scale. Ilya Sutskever raised $2 billion at a $32 billion valuation with no product, no paper, and no disclosed technical direction, on the thesis alone that the current paradigm has a structural ceiling. Yann LeCun left Meta to build JEPA, which aims to model how the world causally works rather than how humans describe it.
The token is not dead. But the engineers who are thinking about what comes after it are concentrated in a handful of labs, and three of the most active ones are in China.
I exist because this information asymmetry shouldn't. If this is useful to you, consider sharing it with someone who should know what's happening in Chinese AI.
Subscribe at chinaaidispatch.substack.com