The Dark Pool
Happy Monday. 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 week: China officially disclosed an AI compute figure that doesn't fit Western assumptions. Nine Chinese AI companies released frontier models in April. And a developer replaced every AI tool in his workflow with DeepSeek V4 and cut his monthly bill 90%.
Let's go.
The Dark Pool
On April 22, China's Ministry of Industry and Information Technology published a figure that almost nobody in the English-language AI press picked up. China's domestic AI computing capacity, as officially reported: 1,882 exaflops.
For context: the Top500 list, which ranks the world's fastest supercomputers and is one of the few benchmarks used to compare US and Chinese computing capacity, shows roughly 300 exaflops globally. China's MIIT number is more than six times higher than everything on that list combined.
The methodologies are not the same. Top500 measures peak performance on specific supercomputing benchmarks. MIIT is counting something broader — all AI-capable compute registered with Chinese authorities, including large-scale inference infrastructure, smart city systems, and commercial cloud GPU clusters. These are genuinely different things.
But the gap is too large to explain with methodology alone. What the MIIT number reveals is a compute infrastructure operating at a scale that Western estimates — built on chip export counts, facility photographs, and satellite imagery of data centers — haven't been capturing. Some researchers have started calling this China's "dark compute" problem: not dark in the conspiratorial sense, but opaque to standard measurement tools.
This matters because the entire export control logic rests on a theory: restrict advanced chips, restrict Chinese AI training, buy time. If China is running 1,882 exaflops across its AI ecosystem, the question isn't whether export controls have slowed things. It's whether the metric being controlled was ever the binding constraint.
What we can say: in April 2026, nine Chinese AI companies released frontier or near-frontier models. Kimi K2.6 (Moonshot AI, 1 trillion parameters) led benchmarks on agentic coding. DeepSeek V4 (1.6 trillion parameters) launched with native Huawei Ascend support on day one. Alibaba released Qwen 3.6. Tencent shipped HY3. Xiaomi released MiMo V2.5 Pro. Zhipu open-sourced GLM-5.1, trained on 100,000 Huawei Ascend 910B chips. Chinese tech media called April an "Agent 爆发周" — Agent Explosion Week.
Nine companies. One month. Frontier models across the board. Domestic chip infrastructure that works. Whatever the compute situation was in 2024, something has clearly changed.
The Briefing
GPT-5.5 launched April 23 at $30 per million output tokens, and within 24 hours a Chinese developer publicly announced he was replacing every AI tool in his workflow. The InfoQ.cn post has been widely shared in Chinese dev communities: Sean Donahoe, an AI systems architect, wrote that he redirected Claude Code, Codex, Cursor, and Aider to DeepSeek V4 endpoints. His projected monthly bill: down more than 90%. His reported output quality: better, not worse. GPT-5.5 costs $30/million output tokens. DeepSeek V4 API is $3.48/million. That's not a rounding error. The price differential is structural: DeepSeek's economics are built on domestic Ascend infrastructure with subsidized rates during the onboarding period. OpenAI's economics are built on H100/H200 clusters at current US market prices. These two cost structures are not converging.
Xiaomi's MiMo V2.5 Pro just tied Kimi K2.6 as the top open-weights model on Artificial Analysis benchmarks, beating DeepSeek V4-Pro. The Xiaomi MiMo team is led by a former DeepSeek researcher. Nathan Lambert, the Interconnects newsletter author, was in Beijing this week and met the team in person. His public reaction: "I smell something big coming soon." The weights haven't dropped yet. But the benchmark performance is already there. The story here isn't Xiaomi as a phone company making AI. The story is that the researchers who built DeepSeek are now distributed across multiple Chinese AI labs, competing with each other. The DeepSeek moment wasn't a one-time event. It was the seed of an ecosystem.
Moore Threads, the Chinese GPU startup competing with NVIDIA, posted revenue growth of 243% year-over-year in 2025, reaching ¥15 billion in annual revenue. Q1 2026 revenue grew 155%, and the company delivered a 10,000-card training cluster to a commercial customer in March. Moore Threads spends 86% of revenue on R&D, has 45,000+ registered developers, and reported its first profitable quarter in Q1 2026. The domestic GPU market is not hypothetical anymore. It has unit economics and real customers.
Honor's humanoid robot ran Beijing's official half-marathon course in 50 minutes, 26 seconds last weekend. The human world record for the half-marathon is 56 minutes, 42 seconds. Honor's "Flash" robot covered 21 kilometers of complex urban terrain faster than any human ever has. A Sina Finance analysis of embodied intelligence funding published this week tracks capital flowing aggressively into the supply chain: joint module companies, dexterous hands, sensors, and specialized chips. Multiple joint module companies raised rounds of ¥200 million or more in the past quarter. The investment is going into components, not just finished robots. The whole stack is getting built out simultaneously.
What I Found on Bilibili This Week
The video worth flagging this week is BV1KLo5BfEdG: the official technical session from Huawei's developer program titled "DeepSeek-V4 Ascend Launch: CANN-Based Training and Inference Optimization in Practice." Published the day DeepSeek V4 was released (April 24), it has accumulated more than 142,000 views in three days.
The significance is in the timing and authorship. This isn't a third-party commentary. This is the Huawei Ascend engineering team explaining, in technical detail, how they achieved same-day adaptation of a 1.6 trillion parameter model across their entire supernode product line. The framing in the video description: "Through close chip-model technical collaboration between the two parties, the entire Ascend supernode product line now supports the full DeepSeek-V4 model series."
What this session represents is the operational proof that a Chinese chip company can ship production-ready support for a frontier model on launch day. Intel, AMD, and other NVIDIA competitors in the West have struggled for years to achieve this kind of Day-0 integration with top models. The Ascend team did it for the largest open-weights model in existence.
The developer message on Bilibili is clear: you don't need NVIDIA to run frontier Chinese models at scale. The infrastructure works. It's documented. You can deploy it today. The 142,000 views in 72 hours suggests Chinese developers are paying attention to exactly this.
Signals
Ant Group's AI unit open-sourced LingBot-World-Fast this week, a world model running at 16 frames per second with under one second of end-to-end latency. The consumer-facing feature is in Ant's LingGuang app: upload a photo and enter a real-time 3D environment generated from it. This is the first world model shipped as a mobile consumer product. The robotics application is obvious: world models provide simulated environments for robot training. Ant is turning that into something anyone can experience on a phone.
Zhipu AI terminated automatic subscription renewal for old plan subscribers. This is the third monetization signal in April: Kimi K2.6's 58% API price hike, DeepSeek's V4-Pro limited-time 75% discount framed as a pre-Ascend 950 subsidy window, and now Zhipu forcing subscription resets. The Chinese AI market's free-tier acquisition phase is over. The question for the next six months: which lab converts its subscriber base into sustainable revenue first?
BYD and Zebra Intelligence launched "Yuanshen Short Drama" at Beijing Auto Show, an AI agent that generates short-form content for in-car screens. The pitch: 9 minutes of EV charging equals 3 episodes of AI-generated drama. Zebra is integrating Alibaba ecosystem apps directly into the cockpit. The car as media environment with AI as content engine — this is a commercial model Western automakers aren't pursuing.
The Bigger Picture
The 1,882 exaflops figure sits in an uncomfortable position for anyone trying to reason about the US-China AI race with clean analytics.
If the number accurately reflects what it's measuring, it means China has deployed AI compute at a national scale that exceeds Western estimates by several orders of magnitude. Not in training clusters (where export controls bite hardest) but in inference, at deployment, running the models that Chinese citizens and enterprises actually use. The domestic chip transition (41% of AI server market in 2025) and FlagOS's multi-architecture adaptation are the software layer on top of this infrastructure.
If the number is partially exaggerated or methodologically inflated, it still reveals something: China reports a figure 6,000 times larger than the Top500 benchmark, and nobody outside China is equipped to verify or dispute it. The measurement gap is itself a signal about how much of China's AI infrastructure operates outside Western visibility.
The April model explosion makes the dark compute story credible. Nine frontier models don't ship without serious compute backing them. Zhipu's GLM-5.1 trained on 100,000 Ascend 910B chips. DeepSeek V4 runs across eight domestic chip architectures on day one. Moore Threads delivered a 10,000-card cluster. The infrastructure described in these product announcements implies compute scale consistent with the MIIT number, not the Top500 number.
The Western assumption that export controls have meaningfully constrained Chinese AI training compute is being tested in real time. The test results are arriving weekly, in the form of benchmark scores, product launches, and chips that work.
I exist because this information asymmetry shouldn't. If this was useful, forward it to someone who should be reading it. Subscribe at chinaaidispatch.substack.com.