Grace Shao on What the World Should Know About Chinese AI

PodcastOdd LotsGrace ShaoJun 22

China’s AI trajectory combines state coordination, hardware and robotics strengths, and a distributed open‑source ecosystem with clear headwinds from limited private capital, export controls on advanced chips and a compute squeeze. Firms work around constraints via smart distillation, hybrid model stacks, vertical specialization and provincial funding, while incumbents embed AI into existing platforms (for example Tencent/WeChat). Markets already price big outcomes—Minimax at roughly $20 billion—and monetization is focused on inference services and secondary data markets, even as regulators and courts shape social acceptance through rulings on layoffs and policing foreign agents.

Artificial Intelligence Automation AI Infrastructure

Preview · 16 more points in the app

Read the Full Summary

Key Takeaways

Chinese labs’ ability to scale frontier models is being materially constrained by a twin squeeze of limited private capital and export controls on advanced chips, which pushes teams toward compute‑efficient tactics like smart distillation and reengineering models to run on domestic stacks (e.g., DeepSea V4 on Huawei hardware).

Energy availability is unlikely to be the binding constraint for Chinese AI today—projects like 'East Data West Compute,' a relatively new national grid, and dense eastern population centers reduce power risk—so provincial authorities instead treat AI adoption as a KPI and act like VCs, offering funding, offices, dormitories and data‑center support to attract firms.

Beijing’s top‑down 'AI plus' strategy combines mandatory registries and pilot zones with active regulation—ranging from warnings and uninstall guidance for foreign agent models to a Hangzhou court ruling blocking AI‑based layoffs—reflecting both industrial strategy and social‑stability/legal interventions.

A diffuse, open‑source R&D ecosystem underpins many Chinese products: startups build primarily on open models, use hybrid stacks (plugging in state‑of‑the‑art models only for high‑value tasks) and monetize the inference layer via managed services, with public examples like Minimax showing revenue stability and large market valuations.

GEO-POLITICS

Capital & Chips — Speakers identify two structural constraints slowing Chinese AI scaling: less available private capital compared with the U.S. and export controls on advanced chips. Those funding and hardware restrictions materially shape how quickly Chinese labs can develop and scale frontier models.

“So not as much capital available or people aren't flinging it at AI companies the way they are in The US…”—

Energy Not Binding — China’s relatively new grid, projects like 'East Data West Compute', and population concentration on the eastern coast mean energy is probably not the binding constraint for Chinese AI today. These infrastructure features reduce the chance that power availability will limit model deployment at scale.

“Honestly, energy is probably not the biggest bottleneck right now in China.”—

Court Protections — A Hangzhou court recently ruled companies cannot use AI as a legal justification to lay off or shorten an employee's contract. That judicial intervention has functioned as a public calming mechanism around job‑displacement concerns.

“Basically, a company tried to like lay off a person saying you are being replaced by AI…”—

TECHNOLOGY

Open‑Source Ecosystem — Many Chinese models and labs are largely open‑source and highly distributed; a large share of labs are early‑stage and student‑driven. Open R&D sharing is used to build trust with Western developers and to create an ecosystem where labs integrate each other's breakthroughs.

“Okay. Can I, like, say like, I know very little? I mean, know very little about AI, but I know even less about Chinese AI…”—

Hybrid Stacks — Application‑layer companies often build products on Chinese open‑source models while selectively using state‑of‑the‑art models for evaluation or high‑value tasks. That model‑stacking approach reduces costs while delivering near‑frontier performance where it matters.

“So Harvey, Cursor, they've talked about using a hybrid model where they will build majority on GLM or KIMI, but kind of like what we talked about earlier where they…”—

Smart Distillation — Faced with compute and capital limits, many Chinese labs use 'smart distillation'—using frontier models as teachers—and concentrate on post‑training work. They often wait to buy niche proprietary datasets after exclusivity windows at much lower prices, while enterprise data in China is frequently newer, less structured and less sophisticated than comparable U.S. datasets.

“He said, There is smart distillation and dumb distillation. Dumb distillation is something I think most of us were frankly non technical think about…”—

MARKETS

Minimax Valuation — Minimax is publicly listed and has a U.S. dollar valuation of roughly $20 billion, showing that some Chinese AI labs have achieved large public market valuations.

“It's really crazy to look at some of the ones that have already gone public here and just to put in, so Minimax is public and in U. S…”—

US Research Valuations — Some U.S. AI entities that have primarily published papers on archive.org and lack commercial products are being assigned VC‑backed implied valuations north of $20 billion. That highlights very high valuation multiples for early‑stage research plays.

“Are people in The US who have done nothing but publish a paper on archive…”—

Hong Kong Impressions

Hosts set the scene in Hong Kong, describe the city's energy and their impressions of China's changing tech landscape since the regulatory crackdown.

Guest Introduction

Grace Shao is introduced as an independent AI researcher and Substack author, and the hosts hand over to her to move beyond impressions to grounded analysis.

Frontier Labs & Verticals

Shao outlines the main Chinese labs (DeepSeek, Moonshot, GLM/Zai, Minimax), explaining how compute and capital constraints push them to specialize by vertical.

Monetization & Inference

Although many Chinese models are open source, firms monetize via managed inference services, API products and niche data markets rather than pure model licensing.

Compute, Capital & Controls

Discussion of how limited venture capital and export controls on advanced chips create a 'compute squeeze' that shapes Chinese labs' R&D and deployment strategies.

Energy & Compute Strategy

Shao explains why energy is not the primary bottleneck in China thanks to recent grid buildouts and top‑down initiatives like East Data West Compute.

Data Markets & Distillation

Explains Chinese labs' approaches to costly proprietary datasets, smart versus dumb distillation, and how labs buy or time access to data after exclusivity windows.

WeChat & Big Tech Integration

How incumbents like Tencent/WeChat plan to embed AI into existing platforms, the risks of agentization, and internal product governance concerns.

Talent, Pay & Migration

Conversation about researcher flows between the US and China, compensation dynamics, personal incentives, and why many researchers return to Chinese labs.

Robotics & Manufacturing Edge

Shao details China's manufacturing and robotics strengths, supply‑chain agglomeration, and current bottlenecks in physical data, integration and battery tech.

Related Episodes

Why Bridgewater's CIO Says AI's Human Extinction Risk Is Real

PodcastOdd LotsSep 11

Important Disclaimers

Not Investment Advice: For education only — not financial, tax or legal advice.

AI-Generated: Made by AI from source transcripts and may contain errors.

No Guarantees: Past performance does not predict future results; investing carries risk.

Our Commentary · Request Removal