Frontier commodities
The massive price disparity between Chinese inference tokens and US frontier tokens was justified by a simple observation: the tokens were not the same thing. The capability gap between something like DeepSeek V3 Flash and Anthropic’s leading models was so large that the 10-to-100x price difference was easy to defend. The expensive tokens unlocked use cases the cheap ones simply couldn’t touch. Both were commodities, sure, but it was like comparing petrol to coal, or gold to copper. And many assumed the leading edge held by US frontier labs would hold, or even widen, over time, justifying valuations approaching a trillion dollars for OpenAI and Anthropic.
Kimi’s K3 launch has led a lot of people to question the infallibility of American dominance on the AI frontier. The marginal output token, which had started to feel like an arbitrary accounting unit designed to rip us off, given that per-token prices could differ 20x while every consumer subscription costs roughly the same twenty dollars a month, now looks a lot closer to a commodity than we thought. K3 has some rough edges, but the consensus is that it has the same shape as a frontier model and belongs in the same class: it landed at roughly half of Opus 4.8’s price, beat it on most software benchmarks, and took frontend and design outright.
So the critical question is this: are inference tokens differentiated tech IP that can command software-like premiums, or are they commodities?
Two things complicate a clean answer.
First, model performance is multi-dimensional. You’re not comparing points on a single axis, you’re computing the distance between vectors. If two models are sufficiently differentiated, then perceived reliability or brand can still command a premium.
Second, there is real segmentation: your tokens are not our tokens, in both difficulty and modality. It’s absurd to spend frontier tokens on image classification, and you can’t meaningfully compare a video-generation model to a coding model in the first place.
But the last few weeks have been like watching the playbook flip-flop several times over. The market has been re-pricing the commodity question in real time, with the loudest voices reversing their opinions in a matter of days, sometimes hours. Fable 4 launches and blows everyone out of the water, widely considered to be as much as eighteen months ahead of the field. That leads Anthropic to decide its tokens are differentiated enough to pull Fable 4 from its subscription plans. Then Grok 3, Sol 1.5, even Mistral Large Spark launch and impress, and Anthropic is forced to keep extending Fable access, because Fable’s edge was being commoditized in real time. I suspect Anthropic needed the “we are not a commodity” narrative going into a trillion-dollar IPO. And I suspect that now, in the wake of Kimi and Thinking Machines, it won’t be able to pull Fable or any future frontier model from its subscription again. It would decimate the user base.
So it certainly feels, at this moment, like AI is commoditizing across a handful of complexity levels and modalities. The open-weights and sovereign-data movements are likely to accelerate that. And the open, collaborative nature of global development is dismantling the objections to commoditization in real time. Take the three biggest objections in turn.
1. American frontier labs will always lead, because of their head start and their data moat. Maybe. But open models collaborate freely and build on each other, and the new open labs have empirically shown a faster rate of progress. And the distillation question cuts the wrong way for the moat: if there was distillation, that is a stronger argument for commoditization, not a weaker one, because it means any future edge won’t persist either.
2. Open models are a red herring, because the weights aren’t legible and self-hosting is too expensive anyway. But legibility was never the point. You don’t need to read the weights to build on them; you can use one model to extract intelligence and make the next one better. And self-hosting was never about individuals. The point is that when many providers host the same open model, competition drives the marginal token down to the cost of electricity.
3. Geopolitics leaves Anthropic and OpenAI a captive market. But commoditization isn’t constrained by borders or regulation. Even if US enterprises never buy from Moonshot directly, the work Moonshot does, along with every other open-weight lab in the world, helps open labs like Thinking Machines, whose first open model shipped this week, outcompete Anthropic. And there is no reason US firms can’t buy from allies: Sakana in Japan, or SEA-LION in Singapore, both aligned with the open-weight movement.
The point of this post isn’t to doom-and-gloom Anthropic or OpenAI. I think the coincidence that so much of the open-model innovation has come out of China has distracted us, or maybe just made things more interesting, from the more significant fact underneath it: humanity has discovered a commodity whose output keeps increasing at breakneck speed. It is precisely because inference is commoditizing that intelligence can be priced at the true cost of generating and serving it, and that a far greater slice of humanity gets to benefit from it.