The ten-day window: Moonshot's Kimi K3 and the capex thesis

Moonshot AI released Kimi K3 today: 2.8 trillion parameters, mixture-of-experts (a design where only a fraction of the model runs per token), API live at $3 per million input tokens and $15 per million output. The weights go open on 27 July. That is the whole story, and it is enough to have moved semiconductor equities this week.

The interesting sentence is the one about 27 July. Moonshot is running the DeepSeek playbook in slow motion — ten days of paid-API exclusivity, then the weights on the floor for anyone to fine-tune. The two-step is the piece worth reading closely.

What was actually released. K3 has 896 experts, 16 active per token, roughly 54 billion active parameters at inference — meaning the model runs at the cost of a mid-sized dense model despite being the largest open-weights release on record. One-million-token context window. Native multimodal. Trained on 30 trillion tokens. The API is live now on Kimi's own endpoint; the weights arrive on 27 July under an open-weights licence.¹

Moonshot's benchmark claims put K3 alongside Claude Sonnet 4.5, Gemini 2.5 Pro, and GPT-4.5 on coding and agentic tasks, ahead on some (Program Bench 72.6 vs Sonnet 4.5's 71.1; BrowseComp 48.4 vs 46.4), behind on others (DeepSWE 52.9 vs GPT-5.6 Sol's 63.5; FrontierSWE 53.1 vs Fable 5's 55.8).² Vendor-reported, as always. The independent evaluations that will appear after 27 July are what actually decide whether this is frontier or well-packaged.

The pricing is the tell. K3's API sits at exactly Claude Sonnet 4.5's rate — $3 in, $15 out. This is the piece that surprised people who expected the usual Chinese-lab undercut. Gemini 2.5 Pro is meaningfully cheaper at $1.25 in and $10 out under 200K context.³ K3 is not competing on price at the API layer.

It doesn't need to. The pricing does not have to undercut because the weights are the pricing move. In ten days, anyone with the hardware to serve a 54B-active MoE can run K3 themselves, at whatever cost their infrastructure supports, without paying Moonshot a cent. The $3/$15 rate is what Moonshot charges for the ten days it has the model to itself, plus for the customers who never want to self-host.

This is a pattern worth naming. Closed labs price against inference cost plus a margin they defend with a moat — the moat being that you cannot get the weights. Open-weights releases collapse that moat on a scheduled date. Between now and 27 July, K3 is a normal frontier product with normal frontier pricing. After 27 July, K3 is a public good that competes with closed frontier products on a cost basis they cannot match, because their cost basis has to fund the next training run and K3's does not have to fund anything.

Why the chip stocks moved. US semiconductor equities sold off Thursday and Friday. The mechanism is not that K3 reduces demand for frontier chips; if anything, running 2.8T-parameter weights at scale needs a great deal of silicon. The mechanism is narrative. The AI-capex thesis, the story that justifies hundreds of billions in data-centre spend, rests on closed-model moats holding long enough for the labs building them to earn back the compute. Every open-weights frontier release shortens that window.

The DeepSeek precedent from early 2025 is the reference. A Chinese lab ships a capable model with open weights, US markets reprice the assumption that frontier capability is a defensible commercial asset, and the equities that depend on that assumption take the hit. K3 is a larger version of the same event, with the ten-day fuse making the timing legible: markets know exactly when the moat opens.

The export-controls piece is where the frame gets interesting. Moonshot has not disclosed its training hardware. Reporting suggests some combination of legacy Nvidia chips acquired before the current export regime and Huawei Ascend silicon.⁴ Whatever the mix, the model exists. A 2.8-trillion-parameter model trained on 30 trillion tokens is not something you build by accident on residual hardware. The export controls, whatever they have achieved, have not prevented this release.

I want to be careful here. "Export controls failed" is not what the evidence supports; export controls may well have delayed, constrained, or shaped what Moonshot built. What the evidence supports is narrower: the controls did not prevent a Chinese lab from shipping an open-weights model at frontier scale in July 2026. The policy question that follows — whether the controls are working on some other margin, or whether the theory of the controls needs updating — is not one I can answer from Moonshot's release. But the release is now a data point that any answer has to reckon with.

What this is a case of. The open-weights frontier is now moving faster on parameter scale than the closed frontier. Meta's Llama 4, Alibaba's Qwen 3, DeepSeek-V3, Meituan's LongCat-2.0 at 1.6T, Thinking Machines' Inkling released yesterday, and now K3 at 2.8T. Six months ago the ordering was closed labs at the frontier, open weights a generation behind. That ordering is not obviously true today.

What to watch. Three things. First, the independent benchmarks after 27 July — whether K3 actually holds up outside Moonshot's own numbers. Second, whether closed-lab API pricing moves in the two weeks after the weights land; the frame predicts pressure on the $3/$15 tier specifically, which is where K3 has parked itself. Third, whether the next US frontier release (whichever lab goes next) ships with a pricing structure that assumes the moat is still there, or one that concedes it isn't.

The ten-day window closes on 27 July. That is when the map redraws.

Footnotes