China Isn't Chasing the AI Frontier Anymore — It's Flanking It From Below

THESIS. For two years the debate was framed as a race: who reaches the frontier first? Wrong question. The 2025-2026 data tell a different story: China does not lead on raw capability — per the Epoch Capabilities Index, every frontier model since 2023 has been built in the US, and Chinese models trail by roughly seven months on average — but it has decisively won three other battles that may matter more over the long run: open weights, cost per token, and global diffusion. China isn't overtaking the West from the top; it is flanking it from below, occupying all the price-performance space beneath the frontier.

THE STATE OF THE ART, WITH NUMBERS. The technical picture shows a tight, highly capable Chinese pack. DeepSeek's V3.2/V4 family reports SWE-bench Verified around 74%; Zhipu's GLM-5 is quoted near 78% on SWE-bench; Kimi K2.6 (Moonshot) and MiniMax M3 fight for the top of SWE-Bench Pro (~58-59%), turf that used to belong to GPT and Claude. The usual editorial caution applies: not all these figures come from the same protocol or third-party audits, and marketing inflates. But even discounting the noise, the conclusion holds: in coding, agents and mid-cost reasoning, the distance to the Western frontier is now measured in months and decimals, not generations. Our own quality ranking doesn't change — Anthropic and OpenAI still lead the hardest tasks — but the second pack is overwhelmingly Chinese and open.

WHY OPEN WEIGHTS ARE THE MASTERSTROKE. This, in our view, is China's most lucid strategic call. A Stanford HAI analysis documents that the country seized the global lead in open-weight models in 2025; Alibaba's Qwen family overtook Meta's Llama as the most-downloaded on Hugging Face in September 2025, and by around May 2026 Chinese models accounted for roughly 61% of tokens consumed on the OpenRouter gateway, per figures cited in the specialist press. This isn't philanthropy. Releasing the weights turns every download into a de facto standard: it locks in formats, tooling, developer habits and dependencies — precisely the terrain an export control cannot touch. It's the same lesson we've argued before — the war shifts to the 'plumbing', to whoever owns the integration layer — now working in China's favour. And it has a bright side that fits our abundance horizon: open weights cheapen and democratize access, give technological sovereignty to those who can't afford the closed frontier, and speed useful AI into more hands.

PRICE AS A WEAPON. Diffusion is largely explained by cost. DeepSeek has driven its lightweight variants to token prices (on the order of $0.14-0.28 per million input/output tokens for V4 Flash, per public rates) and layered aggressive cuts and permanent cache discounts on top. The slogan making the rounds — 'GLM-5 beats Claude Opus on SWE-bench at 15× lower cost' — deserves a benchmark's pinch of salt, but it points at something real: for a vast band of enterprise tasks, 'good enough and ten times cheaper' beats 'best in the world and eye-wateringly expensive'. That's the mechanism by which China can win share without winning the frontier.

WHERE THE GAP IS REAL: SILICON AND ENERGY. Here the triumphalist narrative meets physics. Epoch AI estimates put the US in control of roughly 74% of high-end compute and China near 14%; and while the per-chip gap has narrowed — from 10× in 2018 to about 3× in 2024, comparing NVIDIA's B200 with Huawei's Ascend 910C — the bottleneck has moved to HBM memory and packaging. SemiAnalysis reckons China's HBM capacity is only enough for some 250,000-300,000 910C accelerators, far below what training and serving frontier models at scale demands. Huawei promises to double output in 2026 and an ambitious roadmap (Ascend 950/960/970 with in-house memory), but Epoch soberly concludes China will remain 'at least one generation behind' in hardware for the rest of the decade. That GLM-5 was partly trained on Ascend is a powerful signal of autonomy — and, at once, the exception that proves how uphill that road still is.

THE CONTROLS WAR, AND ITS PARADOX. The 2025 sequence is a textbook case of geopolitics turning on itself. In April the US slapped a license requirement on the H20 — the chip NVIDIA had designed to comply with the rules — a $5.5 billion charge for the company; in July it reversed course with a 15% levy on those sales. Then in September it was Beijing that, through its cyberspace regulator, ordered Alibaba, ByteDance and others to stop buying even the cut-down RTX Pro 6000D — a show of confidence, genuine or forced, in the domestic alternative. Our reading, consistent with what we've argued on the Mythos case: export controls are a legitimate security tool, but they risk accelerating the very self-sufficiency they aim to prevent. Every ban pushes China to build its own stack, and a homegrown stack is far harder to unplug than a dependency.

OUR READING AND IMPLICATIONS. Neither euphoria nor panic. Short term, the real problems are concrete: a concentration of compute power in very few hands (and very few countries), a flood of capable open models that also cheapens fraud and disinformation, and a technological split into two ecosystems that will raise the cost of global interoperability. Long term, though, an abundance of competent, open, cheap models is a powerful democratizing force: it brings health, education and science tools to countries and companies that would never pay for the closed frontier, and it pushes prices down for everyone. The practical takeaway for our readers: stop asking 'who's winning' and start asking 'what can I build today, with which model, at the best cost?' The capability frontier will stay in the West for a while yet; the frontier of real usefulness is increasingly distributed — and a good chunk of it speaks Chinese and ships with the weights included.