Kimi K3 charges $3 per million tokens; Thompson says that's not the risk

Kimi K3 undercuts Sol on price, and Thompson says that's not the scary part

Ben Thompson's essay was prompted by two releases: Moonshot AI's Kimi K3, a 2.8-trillion-parameter model, and Alibaba's decision to open-source its 2.4-trillion-parameter Qwen3.8 Max. His argument runs against the panic: intelligence is becoming a plain commodity, and frontier labs' current pricing power comes from scarcity, not some durable technical edge.

Thompson separates two kinds of cost that get conflated in the "China is winning" narrative: the fixed cost of training a model (R&D) and the variable cost of running it (COGS). Running inference costs money in direct proportion to revenue; training a model is a sunk cost that doesn't scale with usage. K3 prices tokens at $3 per million input and $15 per million output; a frontier rival he calls Sol charges $5 and $30. He lists five things that actually determine that COGS gap: model footprint, inference efficiency (mixture-of-experts architecture), memory efficiency (the KV cache), serving efficiency (batching and caching), and token efficiency, meaning how many tokens a model burns to reach a correct answer.

Crucially, Thompson argues tokens themselves aren't fungible: a token from one model isn't interchangeable with a token from another. But intelligence, meaning a correct answer at a given task, is becoming exactly the kind of commodity where the lowest-cost producer sets the market price and everyone else either matches it or goes under. On distillation specifically, meaning Chinese labs training on frontier labs' own outputs, he doesn't dismiss the advantage, but reframes it as a policy failure: US open-weight developers are stuck following terms of service that ban distillation, while Chinese labs simply aren't. His proposed fix is legislative, not protectionist: make training-data collection explicit fair use, and ban terms of service that forbid distillation.

The one moat he still credits is the agent harness. "Claude Code and Codex are proving quite sticky," he writes: once a developer picks a coding harness, they tend to stay, and that stickiness matters more for non-technical users who'll never touch a second tool.

But the essay's sharpest material isn't economic at all. It's a cybersecurity story he uses to reframe the whole question. When Hugging Face investigated a breach earlier this year, its own incident responders couldn't get American frontier models to help: the models' safety guardrails read forensic analysis of the attack as an attack itself. Hugging Face ended up running the investigation through a Chinese model, Z.ai's GLM-5.2, instead. Thompson's real complaint is that the Trump administration's restrictions on using frontier models like Fable and Sol for defensive cybersecurity work created this dependency: the US locked its own defenders out of using US models to defend US infrastructure.

The thread tested 50 models and got the opposite answer

Thompson's "intelligence is a commodity" framing drew the most pushback, and the strongest came with data attached.

"Interestingly, frontier models did worse on the tasks than open weight models. However, when it came to costs, the picture was reversed: frontier models were much, much more token-efficient. In fact, almost no open-weight model was able to meet the initial token budget, while almost all frontier models did." — an engineer who benchmarked roughly 50 models across every major family for a token-budgeted production system

That's close to the inverse of Thompson's framing: not that intelligence has flattened into a commodity, but that frontier models are winning specifically on the efficiency metric he says should be converging.

Was the distillation advantage ever uniquely Chinese?

Someone questioned whether Thompson's central asymmetry, that Chinese labs distill freely while American ones can't, actually holds up.

"Is this an assertion that is backed by evidence? … On the stand in a California federal court on Thursday, Elon Musk was asked if xAI has used distillation techniques on OpenAI models to train Grok, and he asserted it was a general practice among AI companies." — the former founder of an ML company, later CTO at an AI-policy startup

If a US frontier lab's own founder is testifying under oath that distillation is industry-wide, Thompson's China-specific policy failure starts to look like an everybody-does-it problem with a legal fig leaf attached.

Is the harness moat real, or does it just feel that way until you switch?

The other leg of Thompson's argument, that Claude Code and Codex stickiness is a durable moat, ran into direct experience.

"I have been building my own agent harness for a while, and I can tell with confidence that the harness almost does not matter, the entirety of the AI magic is the model itself." — an engineer who has been building his own agent harness

A separate commenter described switching from Claude Code to Codex, and before that from Cursor, each time with no friction at all: the opposite of the lock-in Thompson describes.

If nobody's much afraid of Chinese models, who actually is?

The thread answered the essay's own headline question more literally than Thompson did.

"The people who are most afraid of Chinese models are the VCs who poured into Anthropic and OpenAI at astronomically high valuations. Anthropic is valued at $1.2T and OpenAI is targeting $850B." — a commenter with over a decade on HN and 6,700-plus karma

Those valuations were priced for labs commanding premium margins on API access. If Kimi K3 and Qwen3.8 Max force a race to the bottom on token prices, the fear Thompson's title asks about isn't really about national security. It's about whether the two most-valued private companies in tech are worth what investors just paid for them.