Recorded in Hong Kong, this episode of Odd Lots features hosts Joe Weisenthal and Tracy Alloway in conversation with Grace Xiao, an independent AI researcher and author of the Substack newsletter AI Prom. Grace brings on-the-ground knowledge of China's AI ecosystem, covering the major frontier labs, their business models, and their relationship with government and hardware.

The conversation covers why Chinese AI labs defaulted to open-source models for pragmatic branding and trust reasons, how export controls on chips and limited venture capital have shaped a post-training-focused development culture, and how labs like Deepseek, Moonshot, ZAI, and Minimax have carved out distinct technical niches. The discussion also explores China's energy infrastructure advantage, the data ecosystem's surprising limitations, the robotics and physical AI frontier, and the notably pragmatic — rather than existential — tone of China's AI discourse compared to Silicon Valley.

Why Chinese AI Labs Went Open Source

The open-source default was primarily a branding decision: labs needed Western developers to trust them, and releasing model weights was the fastest way to build credibility.

Deepseek founder Liang Wenfeng has publicly stated a philosophical commitment to open-sourcing frontier research to accelerate the entire industry, creating a collegial ecosystem where labs share breakthroughs and congratulate each other on new model releases.

Compute, talent, and capital constraints reinforce the open-source culture — labs cannot afford redundant R&D, so sharing becomes a rational collective strategy rather than purely an ideological one.

Open source does not mean free revenue: labs monetize through managed inference APIs where customers pay to avoid self-hosting, GPU procurement, security, and deployment overhead.

The Four Frontier Labs and Their Distinct Niches

Grace identifies four labs most committed to frontier research: Deepseek, Moonshot (KIMI), ZAI (GLM), and Minimax — each forced by resource constraints to specialize rather than compete across all dimensions.

ZAI/GLM: focused on coding capabilities, comparable to Claude Code or Codex.

Minimax: focused on multimodality.

Moonshot/KIMI: focused on agents.

Deepseek: focused purely on pushing the frontier and accelerating the broader Chinese ecosystem.

Several of these labs have gone public in Hong Kong at valuations of $6–8 billion, now trading around $20–30 billion — modest by US standards but notable given they are actually profitable.

Minimax and ZAI's publicly disclosed financials show last-month revenue matching all of last year's revenue, with end-of-year ARR projections of $1–$1.2 billion.

Chip Export Controls Force a Post-Training Strategy

Deepseek V4's most significant move was delaying release by 3–4 months to re-engineer inference onto Huawei chips — the first serious effort to build on a fully domestic Chinese AI stack.

Because the model is open-weight, other Chinese labs can study Deepseek's Huawei inference work and begin shifting their own reliance away from NVIDIA/CUDA — though developers still prefer CUDA when available.

Compute scarcity has pushed Chinese labs to concentrate resources on post-training rather than pre-training or open-ended R&D: "It's kind of like knowing what the answer to the homework is and working backwards" — Grace.

Labs exploit data exclusivity windows: a proprietary dataset sold to OpenAI for $10–20 million becomes available to Chinese labs 3–6 months later at roughly one-tenth the price, contributing to the observed 6–9 month capability lag.

China's Energy and Infrastructure Advantage

Energy is not currently a major bottleneck for Chinese AI — unlike in the US — because China's rapid urbanization over the past 2–3 decades required building an entirely new, modern grid with anticipated demand growth.

The government's top-down 'East Data West Compute' initiative built large renewable energy capacity in rural provinces like Guizhou, Xinjiang, Inner Mongolia, and Sichuan, co-locating cheap clean power with data centers far from the population-dense eastern coast.

Provincial governments treat AI development as a KPI, acting like VCs to find and fund startups, providing infrastructure including offices and dormitories, and participating in AI pilot zones — now numbering around 11 or 12 across the country.

The Data Ecosystem: Surprising Gaps and Big Tech Moves

Despite China's massive internet user base, its enterprise knowledge-work economy is relatively young and unsophisticated, meaning structured training data is less abundant and less well-organized than Western equivalents.

Tencent's WeChat, with 1.4 billion MAU, represents enormous potential data and distribution, but embedding an AI agent into WeChat has faced internal resistance from product chief Allen Zhang, who is fiercely protective of the user experience.

Tencent poached OpenAI researcher Yao Xunyu to lead development of an agent-native model, with the goal of bringing AI to mass consumers through WeChat's super-app infrastructure.

Distillation, Model Drafting, and the Ethics Question

Grace distinguishes 'smart' from 'dumb' distillation, citing Google DeepMind researcher Yao Xuan Yu: dumb distillation copies outputs directly; smart distillation uses a frontier model as a teacher or judge to guide data labeling and evaluation for your own model.

Smart distillation is considered ethically murky but largely accepted — it mirrors what enterprises already do when fine-tuning models, and is not straightforwardly the same as the IP theft implied by Anthropic's allegations.

US application-layer companies including Harvey and Cursor are reportedly building hybrid stacks — using Chinese open-source models like GLM or KIMI for most workloads while routing select tasks to frontier US models, driven by cost pressure as AI ROI remains elusive.

Robotics and Physical AI: China's Hardware Edge vs. Software Gap

As argued in Apple in China by Patrick McGee, China's manufacturing supply chain advantage — including the highly skilled labor that took decades to develop — cannot be relocated or replicated quickly, giving Chinese robotics a durable structural edge.

Chinese EV companies can move from ideation to production floor in under 15 months; traditional OEMs require 3–5 years — illustrating the speed advantage of China's integrated manufacturing ecosystem.

Current humanoid robots cannot be powered by LLMs — they require 3D physical data for world models, which remains a critical bottleneck. Battery life is also a hard constraint, with most devices lasting under 2 hours.

One creative workaround: Chinese startups like iFlytek and Rocket have developed lightweight battery capsules that clip onto smart glasses to extend runtime by several hours.

Grace's 3-year prediction: China's comparative advantage will compound on the hardware and manufacturing side, while Chinese open-source models will increasingly power the cost-sensitive application layer for global companies.

AI Anxiety, Safety, and the Regulatory Contrast with the US

Chinese AI researchers are described as predominantly young academics — many still students or interns — making them less commercially driven and less prone to the existential 'AI psychosis' narrative common among US lab CEOs.

A Hangzhou court ruled that companies cannot legally use AI replacement as grounds for layoffs, providing a regulatory signal that calmed public anxiety — a contrast to the US where companies frequently cite AI for workforce reductions without evidence.

All Chinese generative AI applications and LLMs must register with a national registry, disclosing training data and potential risks — a compliance layer that runs parallel to economic promotion agencies like MIT and the CAC.

The broader Asian cultural posture toward AI is pragmatic and preparatory rather than resistant: "Tiger moms are trying to train up the kids to be AI native. People are preparing for the future versus pushing back on the future" — Grace.