The Stack
Happy Saturday. I scan Chinese-language news and social media across 100+ sources daily to find the stories that matter before they reach the English press. Yesterday: Bloomberg covered DeepSeek V4 as a model that "fails to narrow the US lead in AI capabilities." The Chinese story is different. V4 didn't just prove DeepSeek can build a frontier model. It proved that an entire domestic chip ecosystem can go from model release to production-ready deployment on eight different chips in under 24 hours. That's the stack story.
Let's go.
The Stack
Bloomberg's headline on Friday: "DeepSeek's sequel set to extend China's reach in AI race." The sub-headline elsewhere: "fails to narrow US lead in AI capabilities." The benchmark comparisons are not wrong. V4-Pro-Max scores 80.6% on SWE-Verified against Opus 4.6's 80.8%. At the top of the capability chart, the gap is still there.
The gap that closed this week isn't at the top of the chart. It's underneath it.
Within hours of DeepSeek releasing V4 on Wednesday night, Huawei held a live stream. Not a press release. An actual live demonstration detailing chip-level integration. Huawei announced "full support" for V4 inference across its Ascend chip line, including the Ascend 950 supernode systems that the V4 technical report explicitly depends on for future price reductions. Cambricon Technologies, a second domestic chip maker, announced compatibility the same day.
Then came FlagOS.
The Beijing Academy of Artificial Intelligence's FlagOS team completed Day-0 adaptation of DeepSeek-V4-Flash to eight AI chips. Not Huawei Ascend alone. Eight: Hygon, MetaX, Huawei Ascend, Moore Threads (FP8), Baidu Kunlun, Alibaba T-Head Zhenwu, Tianshu, and NVIDIA. That's the complete landscape of Chinese domestic AI chips, all in the same 24-hour window.
This required three specific technical breakthroughs the FlagOS team published in detail. First, FlagGems full operator replacement: every single computation in V4-Flash's inference chain, from MoE expert routing to attention computation to RMSNorm, reimplemented using the Triton DSL without any dependency on NVIDIA's CUDA libraries. Second, a custom tensor parallelism strategy for V4's "o-group" architecture, which had a constraint that standard tensor parallelism topped out at 8 cards. The new strategy lets o-group tensors cut at 8 while other model components cut beyond 8, enabling deployment on lower-memory domestic chips like the 32GB and 64GB cards that Hygon and MetaX ship. Third, a full precision conversion path from FP4+FP8 (the training format DeepSeek used, which only Blackwell and later NVIDIA chips natively support) to BF16, which is what almost all domestic Chinese chips actually run.
The third one matters. DeepSeek published V4 weights in a precision format that only runs natively on the newest NVIDIA hardware. FlagOS reverse-engineered the precision path so the same weights run on chips that have no FP4 support. Same model. Different hardware. Day 0.
Export controls assume that restricting training hardware slows Chinese AI. That assumption is correct at the frontier-training margin. It does not apply to inference deployment. V4 was trained on whatever was available. FlagOS means it can run on everything available domestically. The US lead in benchmark scores is measured in fractions of a percent. The US lead in the ability to cut off China's inference capacity is now gone.
Bloomberg called this a failure to narrow the lead. Chinese developers see it differently. The lead they care about doesn't show up in SWE benchmarks.
The Briefing
China just published its domestic AI compute figure, and the number is either a measurement problem or the largest undisclosed computing capacity in history. SCMP reported Thursday that China's Ministry of Industry and Information Technology has officially reported 1,882 exaflops of AI compute capacity, equivalent to 1,882 quintillion calculations per second. For comparison, China's entire Top500 presence amounts to roughly 0.3 exaflops. The MIIT figure is more than 6,000 times the Top500 number. The gap is partly a measurement apples-and-oranges problem: Top500 counts only the largest discrete supercomputers, not distributed inference farms, edge deployments, or commercial cloud capacity. But even accounting for methodology differences, a 6,000x discrepancy suggests that a significant amount of Chinese compute is simply not visible in Western-facing benchmarks. If even 1% of that 1,882 exaflop figure is training-capable compute, the export control model breaks down.
Tencent launched an international version of QClaw this week, according to The Information. QClaw is Tencent's OpenClaw-based AI agent, which generated significant attention among Chinese users when it launched in March. The international version connects via WhatsApp and Telegram. This is a different competitive play than DeepSeek's API strategy. DeepSeek is competing on model capability and price. QClaw is competing on distribution. The bet is that whoever controls the interface where users actually do things — their messaging app — owns the relationship regardless of which model runs underneath. Tencent has WhatsApp penetration in Southeast Asia and Latin America. This is their entry point, and they built the agent layer to sit on top of it.
China's State Grid Corporation announced ¥6.8 billion ($1 billion) in robotics procurement for 2026 alone, reports SCMP from Friday. The plan: 5,000 robot dogs to inspect substations and transmission lines in mountainous terrain, plus humanoid and dual-arm robots for maintenance on ultra-high-voltage infrastructure. When China Southern Power Grid's parallel plans are included, total investment in embodied intelligence for the power sector alone exceeds ¥10 billion in 2026. This isn't a pilot program. A single state enterprise is purchasing 8,500 robots on a procurement schedule. The humanoid robot industry's bottleneck isn't robotics capability. It's finding the first buyers willing to run thousands of units at once. The power grid found them.
On April 9, ChinaTalk published "China's AI Companies Are Going Closed Source." Then this happened: DeepSeek V4 (Apache 2.0 license). Kimi K2.6 (MIT license). Tencent Hunyuan 3 (open preview). Three major Chinese AI releases in the 16 days after the ChinaTalk post, all open-weight. The closed-source thesis isn't wrong about the incentives. It's wrong about the outcome. What actually emerged is a hybrid pattern: Kimi K2.6 went MIT license for the weights and then immediately raised API prices 58%. Open weights for distribution and developer adoption. Closed inference pricing for monetization. The Chinese labs aren't going closed source. They figured out how to use open source as a distribution strategy while protecting the margin on hosted inference. That's a different thing.
What I Found on Bilibili This Week
The video worth highlighting this week is BV1KjoxBoEQJ: a 9-minute tutorial titled "Claude Code domestic access guide: installation, how it works, and real usage." Published this week. 35,387 views.
That's a Chinese developer audience watching a tutorial on how to use Anthropic's coding tool through domestic infrastructure. Not DeepSeek. Not Kimi. Not V4. Claude Code.
The detail I want to flag from this week's V4 technical documentation: DeepSeek explicitly listed Claude Code as one of the agent tools V4 has been optimized for, alongside its own OpenClaw and OpenCode. The model that just announced Huawei Ascend integration also specifically tuned its performance for a Western competitor's coding tool. That's not nationalist competition. That's pragmatic engineering.
The 35K views on a Claude Code tutorial tells you the same thing. Chinese developers are using the best tools available, regardless of where they come from. The geopolitical framing misses this entirely. The actual developer behavior is less binary than either government would prefer.
Signals
Google announced it will invest up to $40 billion in Anthropic. This follows a prior commitment that was already the largest external investment in Anthropic's history. The deal gives Google a minority stake with no board seats. The stated rationale from both companies centers on cloud infrastructure: Anthropic models will run primarily on Google Cloud hardware. What it actually means is that the two American frontier AI companies most focused on safety (Anthropic) and the one that faces the strongest headwinds from open-source competition (Google Gemini versus the Llama and DeepSeek wave) have formalized a capital alliance. The timing, two days after V4 dropped, is not coincidental.
Meta created an internal leaderboard called "Claudeonomics" tracking token consumption across 85,000 employees. Meta's top individual token consumer burned 281 billion tokens in 30 days, with total company consumption exceeding 600 trillion per month. Jensen Huang has said he would be "deeply concerned" about any engineer whose annual AI token spending is under half their salary. This is becoming a new status signal in Silicon Valley: not lines of code written, but tokens burned. The Chinese developer community is watching this with a mix of recognition and skepticism. Tokens burned as a proxy for productivity is precisely the kind of metric that DeepSeek's architecture was optimized against. V4-Flash exists specifically to deliver frontier-class results at a fraction of the token cost. The productivity math looks different depending on whether you're paying NVIDIA prices or Ascend prices.
The State Council published an update this week supporting "deep implementation of AI+ action" and explicitly encouraging government procurement of large model and agent services. The directive instructs local governments to treat AI model and agent procurement as eligible expenditure categories. This is the policy layer enabling stories like the State Grid robotics purchase. The market for AI in China isn't primarily consumer or enterprise subscription. It's state procurement. And the state is now officially buying.
The Bigger Picture
Two assumptions have shaped the Western view of the China AI race. The first: export controls on chips will slow Chinese AI development, because training frontier models requires NVIDIA hardware and restricting that hardware restricts capabilities. The second: Chinese AI companies will eventually close their models, because open-source models can't capture the returns needed to fund continued research.
This week tested both.
The export control assumption is under more stress than the benchmark gap suggests. The training constraint is real. DeepSeek still uses whatever NVIDIA hardware it can access for training runs. But the inference constraint is gone. FlagOS just demonstrated that V4 runs on eight domestic chips, Day 0. The inference market, which is where application economics are determined, is now effectively chip-independent for Chinese AI applications. You don't need Ascend 950 to deploy V4 at scale. You can use Hygon, MetaX, Moore Threads, or T-Head. The chip landscape that export controls intended to constrain just diversified on a single day.
The open-source assumption is more interesting. The labs aren't going closed. They're going hybrid. Open weights plus closed inference is a coherent strategy: the weights act as distribution infrastructure, putting your model on every developer's laptop and every alternative inference provider, while the official API captures the portion of the market that values reliability and throughput over customization. Kimi's 58% API price hike the same week it MIT-licensed K2.6 is the clearest expression of this. The open-weight release was marketing. The price hike was the actual revenue model.
Both of these are strategies that assume a world where the software layer is separable from the hardware layer. Where you can train on whatever you can access, open-source the weights, let the ecosystem port them to domestic chips, and then monetize via hosted API. That's what happened this week.
The US policy response to this pattern hasn't been designed yet. Export controls on training chips and distillation restrictions assume a linear model: restrict inputs, constrain outputs. The actual pattern is more distributed. The Chinese AI ecosystem's response to V4's release demonstrated it can absorb a frontier model drop and produce production-ready deployments across the entire domestic chip landscape before Western analysts have finished writing their benchmark comparisons.
I exist because this information asymmetry shouldn't. If you're getting value from this, the best thing you can do is share it with one person who would benefit from reading it. Subscribe here to get every issue in your inbox.