Closed in Code
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.
Closed in Code
Nvidia's competitive advantage is not the GPU. It's CUDA.
The software stack Nvidia built over fifteen years, including the libraries, the debugging tools, the community knowledge, the millions of solved engineering problems, is what makes Nvidia chips the default choice even when alternatives exist. You can design a competitive chip. You cannot, through engineering effort alone, replicate fifteen years of community polish.
This is the problem Chinese AI labs have been quietly trying to solve. Two developments this week suggest they have found a different approach.
The first is ForgeTrain, released by 面壁智能 (MiniCPM's research team). ForgeTrain is a production-grade training framework with one unusual property: every line of code was written by an AI, with no human intervention during the writing process. The team designed a testing environment (what they call a Harness), put an AI model into it, and let the model write a training framework that had to pass a suite of correctness and performance benchmarks. The result is already running on Huawei's Ascend chips. In head-to-head benchmarks against Nvidia Megatron, the industry standard training framework, ForgeTrain is 10% faster.
The reason it can be faster is also the reason AI-written software may close the CUDA gap at all. Megatron has to work for every model architecture: Qwen, Llama, DeepSeek, MiniCPM, and dozens more. It cannot be deeply optimized for any one of them without breaking compatibility with the others. ForgeTrain is written fresh for each specific model. The trade-off that made Megatron general-purpose is the same trade-off that left headroom for a purpose-built replacement to outperform it.
The team's stated goal is ambitious: rewrite all major domestic AI compute software by year-end. Not piece by piece, but by deploying AI as the engineering workforce for the entire stack. When a user needs to train a new model on a new chip, the system generates a custom framework for that specific combination rather than adapting a general one.
The second development comes from SLAI, a joint project between 深圳河套学院, Harbin Institute of Technology's Shenzhen campus, and Huawei's GTS division. SLAI completed what it claims is the first third-party, full-parameter post-training of DeepSeek V4 Pro on a domestic compute cluster. This matters because full-parameter post-training is the hardest part. Inference on Ascend has existed for some time. Lightweight fine-tuning has worked for a while. Full-parameter post-training, where all 1.6 trillion parameters are updated simultaneously, requiring terabytes of distributed memory management and sustained stability across thousands of chips, is the benchmark foreign companies have used to judge whether a domestic chip can genuinely replace Nvidia. SLAI ran 1,500+ training steps without a single NaN error or instability event, reaching MFU (model flops utilization) of 34.9% on Ascend SuperNode.
These two data points describe the same inflection from different angles. ForgeTrain is software writing itself faster than human engineers can. SLAI is that software being validated at production scale on domestic hardware. The Leiphone piece on Ascend's ecosystem maturation puts a third frame on it: Ascend's CANN operator coverage went from below 30% in early 2024 to 80-90% now. One training team describes the transition as "from infancy to adolescence." Migration that once took months now takes a week. The same team reached 65% MFU on Ascend, roughly double what they were achieving on that hardware eighteen months ago.
None of this means CUDA has been beaten. Megatron still runs on more configurations and has been stress-tested in more production environments. But the relevant question has changed. It is no longer whether domestic compute can run AI workloads. It is how fast the software can catch up. If AI writes the software stack, the answer may be: faster than the hardware side of the CUDA moat can compensate for.
The Briefing
MiniMax disclosed its fastest commercialization trajectory yet: ARR doubling every 60 days. As of this week, MiniMax serves 1 million enterprise and developer customers globally, a five-fold increase from six months ago, with approximately 300 million total users. The company's February ARR was above $150 million. The new disclosure implies that figure has roughly doubled again in the ten weeks since. MiniMax is not a lab that most Western AI analysts would name in a top-five list, but its commercialization rate is now inside the bracket of the fastest-growing AI companies globally.
StepFun released Step 3.7 Flash, an 11 billion parameter open-source model designed specifically for agent loops. The framing from the company is precise: the Agent era does not reward the model that answers best once. It rewards the model that can verify, search, retry, and recover reliably across fifty tool calls in a single task chain. Step 3.7 Flash (196B total parameters, 11B activated, up to 400 tokens per second) is designed as the execution layer for that loop, with native multimodal reasoning, search-augmented verification, and code generation. StepFun's prior Flash model hit number one on OpenRouter's trending chart within a month of availability. The bet: at the task-chain level, inference cost and latency are themselves a form of intelligence, because they determine how many verification loops you can afford.
Alibaba's voice AI now holds all three top spots on Artificial Analysis. Fun-Realtime-ASR topped speech recognition with a word error rate of 1.8%, the lowest globally. Fun-Realtime-AudioChat placed first in voice dialogue at 97.6% speech reasoning. Fun-Realtime-TTS-Preview placed first among Chinese models in text-to-speech. The announcement notes the underlying models run across Qwen app, Gaode Maps, DingTalk, and automotive and education partners. Realtime voice AI was considered OpenAI's moat after GPT-4o Voice. It has become a contested benchmark inside a year of that demo.
Domestic AI chip market share in China has crossed 41%, per a widely-shared Bilibili video with nearly 50,000 views. The shift is concentrated in inference workloads, where domestic alternatives have the strongest cost advantage over Nvidia. Training has been slower to change. The SLAI result above is one data point on whether that is changing too. Three years ago the domestic share was effectively zero.
What I Found on Bilibili This Week
The video I want to flag this week has 77,000 views and explains Huawei's τ (Tau) law, their published architectural roadmap written by HiSilicon president He Tingbo.
The τ law starts from a different premise than Moore's Law. Moore's Law measured transistor density per unit area. The τ law measures computation latency, end-to-end time to complete an AI workload, and asks what can be compressed at every level: chip, interconnect, and system architecture.
Huawei's answer unfolds across three layers. The first is Logic Folding. Rather than scaling to smaller process nodes (which requires TSMC's 2nm or 3nm capabilities Chinese fabs cannot access), Huawei is stacking compute in two physical layers on the same die, connected through hybrid bonding. A two-story structure achieves 55% more transistor density at the same process node. The first-generation Logic Folding chip, the new Kirin used in the Huawei GX7 vehicle, is already shipping. This is not a theoretical capability.
The second layer is Unified Bus. Huawei proposes replacing the current mix of interconnect protocols (PCIe, NVLink, Ethernet) with a single protocol across the full system. The τ reduction from this change, per Huawei's calculation, is 500x, because you eliminate the translation layers and synchronization overhead that accumulate at every protocol boundary.
The third is 3D packaging. Rather than routing connections around the edge of a chip (output scales with perimeter), route them through the die face (output scales with area). More bandwidth, shorter wire runs, lower latency per operation.
The analyst's translation: Huawei is not trying to beat TSMC at lithography. They are changing the rules of what matters. If system-level architecture can achieve 1.4nm-equivalent performance at a coarser process node, access to leading-edge fabs becomes a smaller part of the total competitive equation. The domestic supply chain beneficiaries listed in the video are the hybrid bonding equipment makers whose precision requirements match what Chinese fabs can already deliver.
Signals
六张网 (Six Networks) makes compute infrastructure a national grid. The NDRC announced that China's 十五五 infrastructure investment plan will concentrate more than 10 trillion yuan across six national network types: water, new power grid, computing network (算力网), next-generation communications, urban underground pipelines, and logistics. Compute is now categorized alongside electricity and water supply as foundational national infrastructure. This means compute buildout receives the same state coordination, financing access, and policy prioritization as power grid expansion.
LimX Luna launched with 1.72 million Bilibili views and 16,350 likes on the first day. The new full-size humanoid from 逐际动力 (LimX Dynamics) is positioned for interactive commercial deployment rather than pure locomotion, and is now accepting preorders. Chinese humanoid launches are generating more organic social engagement than most consumer electronics product reveals.
China is working on an AI token futures market, per a Reuters exclusive. The concept circulating at the policy level is a derivatives market for AI compute tokens, analogous to electricity futures. The logic: compute is now a commodity with volatile spot pricing, and enterprises planning deployments need long-term price certainty the spot market cannot provide.
Kimi K3 is reportedly launching in Q3 2026 at 2.5 trillion parameters, per a circulating Bilibili summary of leaked specifications. If accurate, K3 would be 56% larger than DeepSeek V4 Pro at 1.6 trillion parameters and more than double Kimi K2.6 at 1 trillion. The parameter scale-up race among Chinese frontier labs has not plateaued.
The Bigger Picture
The CUDA gap is real. Fifteen years of ecosystem knowledge, community debugging, and production hardening cannot be manufactured in a quarter. The argument against China ever closing it was always more about time than technology: even if the hardware reaches parity, the software takes a generation of developers to catch up.
ForgeTrain is a bet that this premise is wrong. If AI can write production-quality training frameworks faster than human engineers can, and if those frameworks can be tuned specifically for each model-chip combination rather than engineered for generality, the time constant changes. The catch-up is not ten years. It is however long it takes to design and run the evaluation loop.
Three data points are describing the same inflection today. SLAI running full-parameter DeepSeek V4 training on Ascend without instability. ForgeTrain writing software to make that possible at 10% above Megatron. Ascend's CANN coverage going from 30% to 90% in eight months.
The 六张网 policy announcement adds context to what this is building toward. China is not just trying to catch Nvidia. It is trying to build a compute infrastructure that, like its power grid, is domestic, state-coordinated, and not dependent on supply chains it cannot control. AI that writes the software stack is one part of that. A compute network as national infrastructure is another.
The pieces are in motion. The question is whether they connect into something that works at production scale. SLAI's 1,500 training steps and ForgeTrain's 10% Megatron advantage are the first operational answers.
I exist because this information asymmetry shouldn't.
If this was useful, subscribe. If you know someone who tracks China AI professionally, forward it. The more people reading this, the more sustainable it becomes.