Stratechery: Kimi K3's Rise Shows Marginal Costs, Not Just R&D, Now Define AI Competition
- Kimi K3, an open-weights Chinese model, approached state-of-the-art capability and sparked debate on X over whether Chinese models threaten US frontier labs.
- Kimi K3 charges $3 per million input tokens and $15 per million output tokens, undercutting Sol's $5 input/$30 output pricing, though the article argues raw token price is misleading.
- Author Ben Thompson distinguishes R&D (a fixed cost unrelated to revenue) from COGS (inference cost, which scales directly with revenue), arguing open-weights models are not actually free to serve.
- Reasoning models like Kimi reportedly need far more chain-of-thought tokens than Sol to reach the same answer, which can erase any apparent price advantage since tokens are not fungible but the resulting intelligence is.
- Model COGS depends on footprint (memory/accelerators per replica) and inference efficiency (architectural choices like Mixture-of-Experts that cut computation per generated token).
Hacker News 의견들
Why exactly is distillation bad? LLMs are basically distillation of the whole open internet anyway, scraped by frontier labs and then distilled again. The US should just pass a law that says training data collection is fair use and bars ToS clauses banning distillation.
I'm immediately sold on that. Sorry, OpenAI and Anthropic.
Wait, if it's that easy, why do the Chinese labs bother distilling from another model instead of just distilling the raw internet from scratch themselves?
If LLMs can hoover up copyrighted data for training, seems fair they can also use LLM outputs since those aren't copyrightable. Barring ToS restrictions on distillation legally though is a much tougher sell, a company should get to pick its customers.
This whole distillation panic is just American exceptionalism dressed up as securities defensiveness. Anthropic's own reports show Chinese labs doing minimal API calls, nowhere near enough to actually train a frontier model that way, the claim is nonsense.
Government can't literally 'bar' a ToS, ToS isn't law. All OpenAI or Anthropic can do is refuse to enforce it themselves by banning accounts, like a coffee shop enforcing a dress code.
Banning distillation is like banning someone from using a compiler to build a better compiler.
There's a real distinction though: building an LLM from raw data is value-add, distillation is value-extraction. Maybe the internet itself should be taxed as a public good, but that's a separate question from distillation.
Rules for thee but not for me, frontier labs love applying that standard to everyone but themselves.
Distillation being 'bad' is just felony contempt of business model.
This feels like time-traveling into the SpaceX IPO for later-round OpenAI and Anthropic investors.
Me, I'm just terrified of finally getting decently priced inference.
The article's claim that agent harnesses are a sticky moat doesn't match my experience building my own harness, the model does almost all the work and a barebones harness like mini-swe-agent still performs fine.
Harnesses will commoditize eventually, but right now harness quality still matters a lot, especially outside terminal-style use cases.
For me the stickiness is really just that providers gate their subsidized plans behind their own harness, not that the harness itself is technically hard to replace.
I think by 'harness' he means the whole product experience, not just code, I stick with Codex mainly because the Mac app is better and I've built habits around it, that kind of lock-in grows as they add more workflow tools.
Yeah it really does feel like harnesses are minimal value-add sitting on top of the raw token pipe.
I built my own self-improving harness in a weekend and it covers 90% of what I'd use Claude Code or Codex for, the remaining 10% just needs a few more prompts, so long-term I don't buy that harnesses are sticky at all.
Nobody's talking about the DC angle: there are surely lobbyists already pushing to restrict Chinese models on national security grounds, the same way TikTok got forced into divestiture and Huawei got locked out of the US market. Chinese models tied to CCP-linked entities are probably headed for a similar fight in Washington or the courts.