Who’s Afraid of Chinese Models?

This is Ben Thompson’s July 20, 2026 Stratechery analysis of Chinese open-weight AI models. It covers inference economics, frontier-lab strategy, China’s industrial policy, distillation, and cybersecurity.

Software drove the marginal cost of copying and distributing a product toward zero. Thompson’s Aggregation Theory explains how that favored scaled companies that controlled demand instead of physical distribution.

AI requires inference for every answer and agent run. Kimi K3 may be an open-weight model near the frontier, but serving it still consumes expensive compute. Marginal costs are back.

Open weights let a company avoid the fixed research and development cost of building a model. They do not eliminate cost of goods sold. Inference spending rises with usage.

Kimi K3 cost $3 per million uncached input tokens and $15 per million output tokens, compared with $5 and $30 for GPT-5.6 Sol. Kimi looks cheaper, but Thompson says token price is the wrong unit of comparison.

NVIDIA describes AI infrastructure as a token factory, measured through tokens per second, time to first token, tokens per watt, and token cost.

Buyers want correct work, not tokens. Models can use different numbers of tokens to reach the same result, so a cheaper token does not guarantee cheaper intelligence.

Intelligence COGS has five parts:

• Model footprint: the memory and accelerators needed for each replica.
• Inference efficiency: the computation used per token.
• Memory efficiency: how many requests can run concurrently.
• Serving efficiency: batching, scheduling, caching, and utilization.
• Token efficiency: how much generation the correct result requires.

When several models can complete a task equally well, the result becomes fungible. Providers then compete through cost structure.

Thompson uses three suppliers with ten units of capacity each and marginal costs of $10, $15, and $20. Demand for 25 units at $20 lets the first supplier earn $10 per unit, the second $5, and the third nothing on its final five units.

He calls the example simplified because lower-cost suppliers can undercut the marginal supplier and demand changes with price. Its mechanism is that the highest-cost capacity needed to meet demand helps set the clearing price. Lower-cost suppliers keep the spread.

The marginal supplier may still carry R&D costs or infrastructure debt. If it fails, supply shrinks and prices rise until another supplier enters or incumbents expand.

Frontier intelligence is not yet a commodity. Compute remains scarce, and buyers still seek Anthropic and OpenAI rather than an interchangeable capability tier.

NVIDIA sells accelerators. Compute owners resell capacity to model companies. Anthropic, for example, secured SpaceXAI’s Colossus 1 capacity to expand Claude usage. Frontier providers sell inference at another markup.

Anthropic and OpenAI also serve new capability levels months before competitors. That gives them time and scale to optimize model behavior, token efficiency, and serving costs.

Compute scarcity creates a price umbrella. Lower Chinese API prices do not prove lower marginal serving costs. As capacity expands, frontier providers can lower prices and apply their optimized infrastructure to older capability tiers.

Thompson gives four reasons.

First, frontier-lab finances assumed training would consume more compute than inference. High inference prices funded the next training run. Agentic demand now offers enough volume to support lower prices.

Second, inference produces data about how people use a model. Providers can use that data to improve the next version, giving them a reason to lower prices and grow usage.

Third, labs can own more of the customer experience. Claude Code and Codex create sticky workflows. That threatens software companies, while competitive open models help those companies defend their customer relationships.

Finally, Thompson attributes part of Anthropic’s reaction to its belief that only tightly controlled organizations can be trusted with advanced AI. Open-weight alternatives challenge that belief.

Alibaba announced Qwen3.8 Max as a 2.4-trillion-parameter model and said its weights would be released. Xi Jinping’s World AI Conference speech called for open source, openness, collaboration, and sharing as AI moves into the physical world.

Thompson describes the strategy as commoditizing complements. China’s robotics, manufacturing, and other physical industries benefit when intelligence becomes widely available. Open models also attract outside work into an ecosystem China can use.

They prevent the United States from holding an asymmetric AI advantage. Wider access weakens the pricing and strategic position of U.S. frontier labs while giving China and other countries a shared model base.

Distillation does not explain China’s whole open-model lead. Chinese labs also have strong researchers, compute, pretrained models, software-hardware codesign, and increasingly capable post-training. Distillation still compresses the final gap between a strong base and a near-frontier system.

Chinese labs can query frontier models for reasoning traces, tool use, grading, and other training signals. U.S. open-model builders must follow provider terms that prohibit this, so they often wait for a Chinese release and learn from a model that has already learned from the U.S. frontier.

Thompson wants U.S. law to treat training-data collection as fair use and bar model terms that prohibit distillation by U.S. companies. His goal is to reduce dependence on China while preserving open-weight competition.

An autonomous agent breached Hugging Face and left more than 17,000 logged events. Hosted frontier APIs blocked attack commands, exploit payloads, and command-and-control artifacts because their guardrails could not distinguish incident response from attack.

Hugging Face ran China’s GLM 5.2 on its own infrastructure. It could analyze the evidence without guardrail lockout and keep attacker data and credentials inside the company.

Thompson says attack-capable models are already widely available, so defenders also need access to the best models. He wants broader defensive access to systems such as Fable and Sol and equal footing for U.S. open-weight builders. Frontier labs should compete on model quality, not define an access regime that leaves American defenders dependent on Chinese models.