AI Around the Globe: China — a parallel AI stack

1 · The Models

A crowded frontier, not a single champion.

Ask a Western researcher to name a Chinese model eighteen months ago and you'd hear one word: DeepSeek. That single-lab narrative is outdated. China's frontier is now a genuine multi-polar contest among at least five serious labs.

The Price Leader

DeepSeek

Hangzhou-based, backed by the High-Flyer hedge fund. V4 is a 1.6-trillion-parameter mixture-of-experts model with a million-token context window — the first major frontier model explicitly engineered around Huawei's Ascend accelerators rather than Nvidia silicon.

The Developer Ecosystem

Alibaba Qwen

Crossed one billion cumulative Hugging Face downloads faster than any open-weight family in history — in one month, more downloads than the next eight competitors combined, including Meta's Llama. Nearly all released under Apache 2.0.

The Sovereign Proof Point

Zhipu AI GLM

GLM-5 was the first frontier-class model trained entirely on Huawei Ascend chips, zero Nvidia hardware in the loop. Zhipu was China's first publicly listed AI company, with enterprise reasoning and tool-use that has edged out Claude Opus on specific coding tasks.

The Agentic Specialist

Moonshot Kimi

Tsinghua-founded startup built around agentic capability — coordinating hundreds of sub-agents and thousands of sequential steps for long-horizon tasks, and shipping a terminal coding agent comparable to Anthropic's Claude Code.

The Sovereignty Proof

Huawei openPangu

Exists to prove a fully sovereign model — trained, hosted, and served on non-U.S. hardware top to bottom — is viable at production quality. A proof of concept for any government or enterprise unwilling to depend on Washington-controlled infrastructure.

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.

2 · The Regulatory Architecture

No single law. Many sharp tools.

Beijing has deliberately avoided a single, comprehensive AI statute. Instead: dozens of narrow, fast-moving rules issued by the Cyberspace Administration of China, each targeting a specific technology as it becomes visible in the market.

Aug 2023

Interim Measures for Generative AI Services

China's first dedicated generative AI regulation, arriving within months of ChatGPT's debut. Requires lawfully sourced training data, mandatory security assessments, and content upholding "socialist core values."

Ongoing

Deep Synthesis Provisions

Targets deepfakes specifically — real-name verification of anyone whose face or voice is synthesized, plus mandatory consent from the person being replicated.

Sep 2025

Mandatory AI content labeling

Visible labels on AI-generated text, images, audio, and video, plus invisible watermarks in file metadata. National standard GB 45438 spells out the byte-level implementation.

May 2026

National digital ID for humanoid robots

A mandatory national digital-ID system assigning every commercial robot a 29-digit traceable code — a regulatory reflex ahead of embodied AI moving through public spaces.

Draft, 2026

Anthropomorphic AI rules

Would require conspicuous alerts that users are talking to a machine, automatic reminders after two continuous hours of use, and annual audits of how minors' data is handled.

Jan 2026

Cybersecurity Law amendments

Wrote AI into national cybersecurity legislation for the first time — provisions on algorithm R&D, training-data infrastructure, and AI ethics rulemaking.

The Carnegie Endowment calls this "regulation through technical control" — distinct from both the American and European models. Beijing embeds governance directly into system architecture rather than policing outputs after the fact. It is operationally demanding to track, precisely because there is no single master text to read. But it is fast: China can write a rule, watch the market react, and issue the next one within months.

3 · The Chip War

Huawei's rise and the limits of self-sufficiency.

No honest account of AI in China can avoid the chip question — it is the single variable that has reshaped every other part of the ecosystem.

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.

4 · Data Centers and the Power Grid

A $740 billion bet.

China is building a captive market at a scale no purely private buildout could ever generate on its own — guaranteeing Huawei, Cambricon, and the rest of the domestic chip industry a customer base large enough to fund years of iteration, even while their products trail Western silicon on a spec sheet.

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.

5 · Where the Capital Is Actually Going

Not foundation models. Robotics.

If you want to see China's highest-conviction bets, don't look at foundation-model labs — look at robotics.

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.

6 · The Cost Story

A deliberate race to the floor.

Arguably the single most consequential export China has produced this cycle — more consequential, in the near term, than any individual 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.

7 · The Fears

Three distinct anxieties, rarely discussed together.

Social & psychological

Inside China

Regulators moved to address AI companion apps after documented cases of compulsive dependency, including among minors. A well-documented labor anxiety also runs through public discourse as a state-subsidized robotics industry works to automate factory and service labor.

Strategic & dual-use

Inside Washington

The export-control debate is fundamentally about whether restricting China's compute access is working — or whether it simply forced a more resilient, self-sufficient domestic stack the U.S. now has less leverage over.

Dependency & content control

Inside the rest of the world

Every Chinese model carries hard-coded content restrictions on politically sensitive topics. Functionally irrelevant for most commercial workloads — a real, deliberate constraint for anything touching political speech or governance.

8 · Breakthroughs Worth Taking Seriously

Strip away the geopolitics and the engineering is real.

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.

9 · Talent and Data

The quiet inputs nobody argues about.

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.

10 · What This Means for the Rest of Us

Three conclusions for anyone placing infrastructure bets.

01

"China is behind" should be retired

The best closed U.S. models still hold an edge, typically months rather than years — but on cost-efficiency and open-weight availability, China now sets the terms other ecosystems respond to.

02

Compute and model quality have decoupled

China's chips still trail American chips on raw performance, probably durably so. But models trained on those inferior chips are, on many practical benchmarks, no longer inferior.

03

The price of intelligence is falling toward zero

China is the primary force driving that curve down. Stress-test any proprietary AI advantage against a world where a "good enough" open-weight Chinese model is one API call away from your customers.

China's AI story in 2026 is not a story of catching up. It's a story of a state, its chipmakers, its labs, and its capital markets moving in genuine coordination toward a single, coherent goal — and getting uncomfortably close to it.

Next in this series: Europe — where the calculus, the regulation, and the fears all look very different.

References

Bibliography

The complete numbered bibliography (37 sources) accompanies the article at the original publication. To preserve citation-link integrity and avoid drift, please consult it directly: the full bibliography at Artnet Inc..

Mishak's Take

Matthew A. Mishak, J.D. · Founder, LegalTek.ai

Alex's field report is the clearest single map of the Chinese AI stack I've read this year, and it lands with direct operational consequences for U.S. attorneys. Three things I want practitioners to sit with:

1. Vendor due diligence now includes model provenance. If a legal-tech vendor fine-tunes on Qwen, DeepSeek, or GLM, that fact is material to Rule 1.6 confidentiality, Rule 5.3 supervision of nonlawyer assistance, and ABA Formal Opinion 512's competence duty. Ask the question in writing.
2. Assume an import-restriction scenario. Firms quietly built on Chinese frontier APIs should have a model-agnostic portability plan — exportable memory, documented U.S. fallback — before an executive order, not after.
3. Content restrictions are a candor issue. Irrelevant for a slip-and-fall memo; not irrelevant for international arbitration, sanctions work, human-rights litigation, or First Amendment matters. Clients deserve to know what the model will and will not answer.

Read Alex's original at Artnet Inc. for the full argument and the 37-source bibliography that supports it — the reporting is his, and it's worth your time.