Source · Published with the author's permission
This commentary summarizes and reacts to "AI Around the Globe: China — a parallel AI stack" by Alex Progrebinsky, published at Artnet Inc. The underlying reporting, framing, and figures are Alex's. Read the original for the full argument and bibliography — it's worth your time.
Most U.S. lawyers still hear "Chinese AI" and think of one lab from eighteen months ago. Alex Progrebinsky's field report for Artnet Inc. is a useful corrective — and it lands at a moment when firms are making multi-year vendor bets on legal-AI platforms whose model provenance they cannot always name.
Alex's thesis, in one line:
"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."
Below is a lawyer-focused summary of his five main moves, followed by what I think it means for the U.S. legal profession right now.
1. Five labs, not one
The single-lab narrative is dead. Alex identifies five serious Chinese frontier labs, each with a distinct strategic role:
The Price Leader
DeepSeek
Hangzhou-based, backed by the High-Flyer hedge fund. V4 is reported as a 1.6-trillion-parameter mixture-of-experts model with a million-token context window — 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. Nearly all releases are Apache 2.0 — making Qwen the default starting point for cost-sensitive developers worldwide.
The Sovereign Proof Point
Zhipu AI (GLM)
GLM-5 was reportedly the first frontier-class model trained entirely on Huawei Ascend chips — zero Nvidia hardware in the training loop. Enterprise reasoning and tool-use benchmarks are competitive with top Western closed models on specific coding tasks.
The Agentic Specialist
Moonshot Kimi
A Tsinghua-founded startup built around long-horizon agentic capability — coordinating many sub-agents across thousands of sequential steps, with a terminal coding agent aimed at the same use cases as Claude Code.
The Sovereignty Proof
Huawei openPangu
Exists to prove a fully sovereign stack — trained, hosted, and served on non-U.S. hardware top to bottom — is viable at production quality. A reference architecture for any government or enterprise unwilling to depend on Washington-controlled infrastructure.
Alex's takeaway is that the Chinese open-weight ecosystem now holds four of the top five positions on independent open-weight leaderboards, closing the gap with the best closed Western models to a narrowing single-digit margin. His causal claim is the interesting one: the gap didn't close because China caught up on compute — it closed because export controls forced an entire industry to obsess over efficiency.
2. Regulation through technical control
Beijing has deliberately avoided a single comprehensive AI statute. Instead, Alex documents dozens of narrow, fast-moving rules from the Cyberspace Administration of China — generative AI service licensing, deep-synthesis (deepfake) real-name verification, mandatory AI content labeling (GB 45438), national digital IDs for humanoid robots, and draft anthropomorphic-AI rules requiring conspicuous "you are talking to a machine" alerts.
He borrows a phrase from the Carnegie Endowment worth remembering: regulation through technical control. Beijing embeds governance directly into system architecture rather than policing outputs after the fact. It is operationally demanding to track — there is no single master text — but it is fast. That's the opposite of the EU's single-statute approach and the U.S.'s sector-by-sector approach, and it's a useful third data point for anyone building a global AI-governance program.
3. The chip war and the $740B buildout
$0
Nvidia data-center chip shipments to China in the quarter ending April 2026 (down from $4.6B a year earlier)
~60%
Share of China's domestic AI chip market Huawei is projected to capture by end of 2026
~$740B
Estimated total China buildout for compute + power grid (≈5 trillion yuan)
Alex is careful to note the counter-evidence. Citing the Council on Foreign Relations, he flags that Huawei's own roadmap shows next-generation chips arriving 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. His summary line captures it well:
"Quality gap widening, deployment gap narrowing."
For lawyers, the interesting fact isn't that Huawei will or won't catch Nvidia. It's that Beijing is guaranteeing a captive domestic market large enough to fund years of iteration regardless — which means the U.S. leverage that export controls were supposed to create is decaying on a clock.
4. A deliberate race to the price floor
Alex argues that the single most consequential Chinese export this cycle isn't any one model — it's the price curve. He cites DeepSeek's cheapest tier at roughly $0.14–$0.30 per million input tokens, StepFun's Step 3.5 Flash at $0.10 per million, and academic teams training competitive derivative models on top of Qwen for as little as $30 to $50. The strategic point:
"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."
This is the sentence I'd tape to the wall of every legal-tech buyer's office.
5. What this means for U.S. legal practice
Alex's piece is a geopolitics field report, not a legal-ethics brief. Here is where I think his facts land for practicing attorneys:
Vendor due diligence just got harder
If a U.S. legal-tech vendor quietly fine-tunes on a Chinese open-weight base model — Qwen, DeepSeek, GLM — that fact belongs in your vendor questionnaire. Model provenance, training-data jurisdiction, and hard-coded content restrictions are now material to Rule 1.6 (confidentiality), Rule 5.3 (nonlawyer assistance), and ABA Formal Opinion 512's competence duty.
An import ban is a real policy scenario
Progrebinsky bluntly predicts the U.S. government will move to restrict Chinese frontier models. Firms that quietly built workflows on Qwen or DeepSeek APIs need a portability plan — model-agnostic prompts, exportable memory, and a documented fallback to a compliant U.S. model — before the executive order, not after.
The floor price of intelligence is Chinese
When a competent open-weight model runs at pennies per million tokens, every U.S. legal-AI pricing model built on 'GPT-4-class inference is expensive' collapses. Expect renewed pressure on billable-hour economics for research, summarization, and drafting — and on legal-tech SaaS margins.
Content restrictions are a professional-responsibility issue
Every Chinese frontier model ships with hard-coded restrictions on politically sensitive topics. Irrelevant for a slip-and-fall memo; not irrelevant for international arbitration, sanctions work, human-rights litigation, or First Amendment matters. Attorneys owe clients candor about what a model will and will not answer.
Three conclusions worth carrying forward
Alex closes with three infrastructure-bet conclusions. Paraphrased, with attribution:
"China is behind" is retired
The best closed U.S. models still hold an edge — measured in months, not years. On cost-efficiency and open-weight availability, China now sets the terms the rest of the ecosystem responds to.
Compute and model quality have decoupled
Chinese chips still trail American chips on raw performance, probably durably so. But models trained on those chips are, on many practical benchmarks, no longer inferior.
The price of intelligence is falling toward zero
China is the primary force driving that curve down. Any legal-AI advantage that assumes expensive inference should be stress-tested against a world where a 'good enough' open-weight Chinese model is one API call away from your client.
Attribution: This post summarizes and comments on "AI Around the Globe: China — a parallel AI stack" by Alex Progrebinsky, published at Artnet Inc. (source). Quoted passages are reproduced in short form under fair use with attribution and with the author's permission to cross-post commentary. All reporting, figures, and characterizations of the Chinese AI ecosystem are Alex's; legal-practice commentary is mine.
Not legal advice: This article is for informational and educational purposes only and does not constitute legal advice. No attorney-client relationship is created by reading it. Consult a licensed attorney in your jurisdiction for advice on any specific matter.
LegalTek.ai is a technology company, not a law firm.