China's Domestic AI Stack | AI Field Notes #70

AnalysisMiniMax rolled out M3, a multimodal model built for extended reasoning, at the Shanghai conference on July 17. Multimodal means it reads and connects text, images and other formats in one pass. The launch lands inside a run of capable Chinese models, many of them open-weight, that now includes GLM-5.2, Kimi K2.7 and DeepSeek V4 Pro. MiniMax is one of China's better-funded model startups, carrying a multibillion-dollar valuation and real revenue. A year ago the frontier was a short list of American labs. Now half the interesting releases carry a Chinese name and, often, a license to download and self-host.

AnalysisEight thousand one hundred and ninety-two Ascend processors wired into one machine is Huawei's answer to the export controls that keep its best Nvidia options out of reach. The Atlas 950 SuperPoD made its first physical appearance at WAIC on July 17, built around 64-card cabinets linked by an in-house optical interconnect Huawei calls UnifiedBus 2.0. Ascend is Huawei's own AI chip line, and the pitch is a full training and inference cluster with no American silicon inside. Raw performance still trails Nvidia's latest per chip. The bet is that scale and a supply you can actually get beat a faster part you cannot buy.

AnalysisXi Jinping walked onto the stage of the World Artificial Intelligence Conference in Shanghai on July 17, the first time China's leader has given the keynote in the event's eight-year history, and the signal was hard to miss. More than 1,100 companies packed over 100,000 square meters of floor with 3,000 exhibits and 300 global product debuts, while a parallel High-Level Meeting on Global AI Governance pressed China's case to help write the world's AI rules. Washington spent the spring restricting Chinese and, for a stretch, American frontier models. Beijing answered with a trade fair the size of a small city and an open invitation to everyone else.

AnalysisA rideable transforming mech suit priced near 3.9 million yuan, about 537,000 dollars, is the kind of product that used to stay in a concept video. Unitree, the Chinese robot maker, is putting its GD01 into what it calls mass production while it races toward a Shanghai STAR Market listing that regulators cleared on July 3 in a record 73 days. The raise targets roughly 4.2 billion yuan, near 610 million dollars, for at least a tenth of the company, implying a valuation around 42 billion yuan. Unitree already outsells Tesla on humanoid robots and turns a profit on its actuators, the motorized joints that let a robot move.

AnalysisNinety-five percent of enterprise AI spending, by Ode's own estimate, still runs raw frontier models with no routing, task tuning or workflow wiring behind them, and that gap is what it launched on July 15 to close. Backed by Anthropic, Blackstone and Hellman & Friedman at a 1.5-billion-dollar valuation, Ode drops experienced engineers straight into community banks, manufacturers and regional health systems to build and maintain Claude-based systems on site. The model borrows Palantir's forward-deployed playbook: send people in, wire the software into real operations, leave working infrastructure behind. The bottleneck was never the model. It was everything around it.

AnalysisRather than patch the problems, Google threw out the base model and started over. Testing on its Vertex AI platform flagged structural failures in Gemini 3.5 Pro around recursive tool-calling, drawing SVG graphics and mathematical reasoning, and engineers judged them baked into the architecture rather than fixable in later training. A July 17 launch was still the target as of July 16, yet Google had published no model card, pricing page, working API or benchmark by then, and its own site still listed the model as coming soon. Reported specs point to a 2-million-token context window and a Deep Think reasoning mode gated behind a 250-dollar monthly tier. Until real numbers land, it is a promise.

AnalysisThe word officials kept using at WAIC was agent, not chatbot. Shanghai's economy chief Tang Wenkan pointed to what he called the world's first AI agent phone, a handset built to run tasks on your behalf rather than answer questions, while the startup StepFun showed an operating system designed around autonomous agents. An agent here means software that plans and acts across apps without step-by-step prompts. China's AI-capable device shipments passed 100 million units in 2025 and are expected to overtake ordinary gadgets this year, per the National Development and Reform Commission. The demo era is closing. The shipping era is loud.

AnalysisVoice AI usually sells on how human it sounds. Rime is selling on whether it can pass a compliance review. The company raised 24 million dollars in a Series A on July 15, led by M13 with Twilio Ventures joining, to scale a speech-to-speech platform aimed at healthcare and finance, where a mispronounced drug name or a stalled response is a real liability. Rime says it already powers more than 100 million interactions a month, and its focus is the quiet work of linguistics, latency and rules, the last meaning audit trails and consent handling a hospital or bank actually needs. The demo-quality voice was solved a while ago. The paperwork around it was not.

AnalysisThe bill for AI compute is now big enough that managing it is its own business. Spectro Cloud pulled in more than 100 million dollars in an oversubscribed Series D on July 15, led by Growth Equity at Goldman Sachs Alternatives with AMD Ventures joining, taking its total raised to 260 million. Its PaletteAI platform is a control layer that runs AI infrastructure across regular, government, neocloud and sovereign clouds. Sovereign cloud means compute kept inside one country's borders for legal or security reasons. The selling point is unglamorous and real: watch what the graphics processors actually cost, and keep expensive hardware from sitting idle.

AnalysisWhen an autonomous agent can log into your systems and move money, the old question of who is allowed to do what stops being only about humans. Oak came out of stealth on July 15 with 60 million dollars in seed funding from Accel, Greylock and CRV to build what it calls an AI-native identity operating system, a single control plane for the permissions of people, machines and agents alike. Identity and access management, deciding who can reach what, has been enterprise plumbing for decades. The twist is agents that act on their own and need governing like employees. As companies wire agents into real workflows, nobody wants one holding a standing key to everything.

AnalysisInkling will not top any leaderboard, and Thinking Machines said so out loud when it released the model on July 15. The startup founded by former OpenAI chief technology officer Mira Murati built a 975-billion-parameter mixture-of-experts system that fires only about 41 billion parameters per task, trained on 45 trillion tokens of text, image, audio and video. Mixture-of-experts means the model activates a small slice of itself per request to cut cost. The weights are open for download, and on coding tests Inkling reportedly matches Nvidia's Nemotron 3 Ultra while burning a third of the tokens. It is also built to flag its own uncertainty instead of bluffing.