The open-weights tier is now Chinese, and DeepSeek just set the floor

Two things happened over the weekend that I think are the same story.

On Sunday, OpenRouter — the router layer that most developers use to address multiple LLMs through one API — crossed a threshold most people had been expecting and then stopped expecting: four Chinese open-weights models (Kimi K2.6 from Moonshot, DeepSeek V4, Zhipu's GLM-5.1, and Alibaba's Qwen 3) together account for roughly 60% of traffic through the platform. The day before, DeepSeek announced that its 75% price cut on V4-Pro, the headline of its previous pricing move, was now permanent rather than promotional.

These are not independent events. The first describes where developer demand sits. The second describes the cost basis that put it there. Taken together they tell you that the open-weights tier of the market, the tier where weights are downloadable or accessible without vendor lock-in, has tipped, and the tipping is structural rather than a moment of enthusiasm.

What the primary signal actually says. OpenRouter's own 2025 usage study, which tracked roughly 100 trillion tokens through the platform, already showed Chinese open-weights families dominating the open-weights segment by share of traffic, with Qwen the single largest contributor and DeepSeek, MiniMax, and StepFun close behind. The 60% figure being reported this week is the accumulated share across four Chinese labs, not a single product, and the methodology underneath it (token volume rather than request count or dollar value) is worth keeping in mind. This is developer-tier API traffic. It is not enterprise contract value. GPT-5, Claude 4, and Gemini 2.5 Pro still dominate the closed-source enterprise tier, and that tier is where most of the dollars sit.

But the developer tier is where the next generation of products gets built, and the routing layer is where price discovery happens.

The DeepSeek cut is a cost signal, not a discount. Promotional pricing is something a vendor can withdraw. Permanent pricing is a statement about cost basis. When DeepSeek says the 75% reduction on V4-Pro is permanent, what it is telling the market is that its inference economics support that price as a steady state. Either the underlying GPU-hour cost is genuinely that low — which would speak to efficiency gains the export-control regime was supposed to prevent — or DeepSeek is willing to run the model at low or negative margin to hold the developer tier, which would speak to a strategic decision about where the market is being contested.

Both readings produce the same downstream consequence. US closed-source labs that compete on the developer-API tier now have to price against a permanent reference point that sits 75% below where V4-Pro previously cleared. Anthropic and OpenAI can decline to follow, which means ceding the price-sensitive developer tier and concentrating on enterprise. Or they can follow, which compresses the margin story that justifies the current round of valuations. There is no third option that preserves both share and margin at the open-weights/API frontier.

Model weight lineage, with the nationality flipped. The model-weight-lineage frame, until recently, mostly described US enterprise procurement worrying about IP encumbrances in acquired or licensed weights from US labs — the carve-out question in M&A, the talent-weight dependency question in compensation. At 60% OpenRouter share, the question reverses. Regulated UK and EU buyers, defence-adjacent buyers, and anyone with a procurement framework that takes provenance seriously now face a different question: can they take a production dependency on weights whose lineage runs to Hangzhou, Beijing, or Hangzhou-by-way-of-Singapore?

In most regulated procurement frameworks I have read, the answer is currently no, but the no is doing work it was not designed to do. Existing frameworks were written to govern data flows and vendor relationships, not to govern the lineage of mathematical objects that have been distilled, fine-tuned, and re-hosted through multiple jurisdictions. The market is bifurcating between developers, who adopt on capability and price, and regulated enterprise, which cannot adopt without policy cover that does not yet exist. This is a slightly strange arrangement and I think it is going to produce slightly strange outcomes — including, plausibly, an entire intermediary layer of Western-hosted, audited deployments of Chinese open weights, sold as "sovereign" by virtue of where the GPUs sit rather than where the weights came from.¹

Export controls measured the wrong variable. The chip export-control regime, tightened progressively through 2024 and 2025, was built around a specific theory: training compute is the binding constraint on Chinese frontier capability, and restricting access to leading-edge accelerators will slow the frontier. That theory may still be correct about training the next generation of frontier closed models. What it has not done is prevent capability diffusion at the open-weights tier, where the diffusion vector is not "PRC labs training new frontier models on H100s" but a combination of efficient training on permitted hardware, distillation from existing weights, and architectural improvements that move the cost curve faster than the controls move.

The policy instrument and the diffusion vector were misaligned. I would not call this a policy failure exactly, controls on training compute may be doing exactly what they were designed to do, but I would note that the variable that mattered for developer adoption was inference cost and weight availability, and the controls did not touch either.

Meta's Avocado silence. The last credible US open-weights frontier candidate, Meta's "Avocado" line, has gone quiet. No announced date, no signal. This could be quality-gating, and Meta has form for delaying releases until they meet an internal bar. It could also be capability stall. Either way, the absence is a market-structure event. If Avocado does not ship in a form that competes with V4 or Kimi K2.6 on the developer-API tier, the open-weights tier structurally lacks a US-origin alternative at the frontier, and the procurement bifurcation I described above hardens into something more permanent.

What to watch.

• Whether Anthropic or OpenAI cuts API pricing on a flagship model in the next 60 days. If they do, the floor has been accepted. If they do not, watch for explicit repositioning toward enterprise and away from developer-tier API.
• Disclosure from OpenRouter on methodology and dollar-weighted share, not just token share. The token-volume framing is load-bearing for the 60% figure and currently unspecified.
• Any UK or EU enterprise procurement guidance on PRC-origin weights. The absence of a framework is itself a market signal; the appearance of one would move billions of dollars of decisions.
• An Avocado release date, or a credible report that there is not going to be one.

The map this week is straightforward. The developer tier of the open-weights market is Chinese, the price floor has been set by a lab that says it can hold it permanently, and the US response at the same tier has gone silent. Capital allocation decisions that assume a symmetric open-weights market are working off an old map.

Footnotes

¶