China: Building a Parallel AI Stack — And It's Working

The composite picture is a Chinese open-weight ecosystem that now holds four of the top five positions on independent open-weight leaderboards, trailing the best closed Western models by a narrowing single-digit margin on most benchmarks — and in some cases, SWE-Bench coding evaluations in particular, matching or beating them outright. The gap didn't close because China caught up on compute. It closed because export controls forced an entire industry to obsess over efficiency, and that constraint became a genuine competitive advantage.

But it would be a mistake to read this purely as triumph. The Council on Foreign Relations has documented a less flattering picture: Huawei's own roadmap shows next-generation chips arriving in 2026 with lower peak performance than its current best chip — a signal that yield and manufacturing constraints at SMIC may be biting harder than official statements suggest. Huawei is not expected to field a chip beating Nvidia's H200 on raw specs until 2027 at the earliest.

That distinction — quality gap widening, deployment gap narrowing — is the whole story in one sentence. Chinese labs cannot yet out-train the best American frontier models chip-for-chip. What they can do is stand up enormous fleets of somewhat-less-capable chips, tune model architectures specifically around those chips' constraints, and get to “good enough, at massive scale, at a fraction of the cost” faster than anyone expected.

Beijing's National Development and Reform Commission is drafting a plan — reported at roughly $295 billion — to build a nationwide AI computing grid running on at least 80% domestically produced chips, structurally locking out both Nvidia and AMD by design. Fold in the electrical grid buildout needed to power facilities at this scale, and the total projected investment climbs to an estimated 5 trillion yuan, or roughly $740 billion.

It tells you where China believes the real bottleneck sits over the next decade: not model architecture, now broadly available as open-weight software, but physical capacity — chips, power, and the data centers to house them.

The investor list looks nothing like Silicon Valley's. Instead of specialist venture funds, the money comes from platform giants — Tencent, Alibaba, Meituan, ByteDance, Baidu — alongside explicitly state-backed funds and China's sovereign “Big Fund” vehicle.

American venture capital has put roughly 90% of its recent robotics and AI capital into software. China has inverted that ratio, treating hard tech — the physical robots, the chip fabs, the power infrastructure — as the strategic priority. That's a direct consequence of the open-weight strategy: when Alibaba gives away a state-of-the-art model for free, the return on investment isn't captured by the model itself — it's captured downstream, in whichever company builds the most competitive robot, EV, or industrial application running on that free model.

DeepSeek's cheapest tier runs at roughly $0.14–$0.30 per million input tokens. Step 3.5 Flash, from Shanghai lab StepFun, ships at $0.10 per million tokens for comparable math reasoning — by one estimate, twenty-five times cheaper than GPT-4o for similar quality. Stanford and UC Berkeley researchers have reportedly trained competitive derivative models on top of Qwen for as little as $30 to $50.

This is not charity. It maximizes the reach of the open-weight ecosystem, making the Chinese stack the default starting point for cost-sensitive developers everywhere, while putting sustained margin pressure on every Western lab competing for the same enterprise customers. The bet: China cannot yet win the race to build the single most capable model in the world — but it can win the much larger race to become the substrate the rest of the world's AI applications are quietly built on top of.

Zhipu training a frontier-class model entirely on non-Nvidia silicon is a real engineering achievement, not a propaganda claim — it proves the software stack around Huawei's chips has matured enough to support serious training runs, not just inference. Kimi K2.6 coordinating hundreds of sub-agents across thousands of steps represents a genuine advance in long-horizon agentic reliability. And the sheer speed of the open-weight release cadence — five labs shipping frontier-adjacent models within months of each other, each with a genuinely different architectural bet — has produced more experimentation than the more concentrated closed-model ecosystem in the United States currently exhibits.

China's AI labs draw disproportionately from a small cluster of institutions — Tsinghua, Zhejiang University, Harbin Institute of Technology, Huazhong University of Science and Technology — that have spent two decades building deep pipelines in applied engineering rather than pure research. That pattern has produced a founder pool that skews heavily toward people who have already shipped hardware or infrastructure at scale, which shows up directly in how quickly Chinese robotics startups move from prototype to mass production.

On data, China's advantage is less about volume and more about the state's willingness to actively shape data infrastructure as industrial policy — establishing uniform national requirements for how training data must be sourced, labeled, and secured. Whether that produces better models long-run is debatable. What it reliably produces is faster, more predictable compliance cycles for domestic players.