DeepSeek makes the floor permanent

DeepSeek confirmed on Monday that the 75% discount it had been running on V4-Pro is now the list price, not a promotion. Output tokens are $0.87 per million; cached input tokens are $0.003625 per million. Claude Opus 4.7 sits at roughly $75 per million output tokens. That is an 86-fold gap on the output side between two models that are, on paper, in the same frontier-class weight tier.

The interesting question is not whether DeepSeek can hold this price. It is what the existence of the price does to the revenue model of the labs that cannot match it.

What was actually announced. DeepSeek's pricing page now lists V4-Pro, its 1.6 trillion parameter Mixture-of-Experts model (MoE: a model architecture where only a fraction of parameters fire on each token, cutting inference cost), at the post-discount rates with no expiry. The lab attributes the durability of the pricing to its deployment on Huawei Ascend 950 supernodes, with further capacity scheduled for the second half of 2026. The relevant primary disclosure is the pricing page itself; everything else is commentary.

Two things matter about this disclosure. The first is that DeepSeek did not announce a new model, a new benchmark, or a new round. It announced that a promotional price is now the price. The second is that it tied the announcement to a hardware story rather than a software efficiency story. That is a deliberate framing choice.

The frame: inference economics. The structural lens here is straightforward. Frontier-class inference has two cost components — the amortised cost of the GPUs running the model, and the cost of electricity and cooling. Western labs run on Nvidia H100 and H200 silicon, leased at rates set by a supply-constrained market. Their per-token cost floor is, roughly, what Nvidia and the hyperscalers decide it is.

If Ascend 950 throughput on a 1.6T MoE is within striking distance of H100 throughput at substantially lower hardware cost, and that is the claim, not yet independently verified, then DeepSeek's cost floor is structurally lower than Anthropic's or OpenAI's. The pricing is then not a discount. It is the new floor for one supply curve, with the Western labs on a different, higher curve.

That is the frame's prediction. The evidence partially fits.

Where the frame holds. DeepSeek's willingness to commit to the price permanently, rather than running another six-month promotion, is informative. A lab running below cost as a land grab does not typically remove the option to revert. Making the price permanent is a signal, possibly an honest one, that the unit economics work at this level. The Ascend 950 deployment timing aligns. The capital flow from High-Flyer (DeepSeek's quant-fund parent) into Huawei silicon is the kind of capex commitment that the AI performativity frame predicts: spending at a scale that reshapes the market regardless of whether the product ships well.

The export-control reading also fits. The US bet was that restricting Nvidia's highest-end chips to Chinese buyers would slow Chinese frontier capability. The observable result, eighteen months in, is that Chinese capital redirected to Huawei, Huawei shipped Ascend 910B and then 950, and DeepSeek is now using that hardware to set a price point that Western labs cannot match on Nvidia leases. The sanctions did not block the capability. They funded the alternative supply chain.

Where the frame breaks, or at least bends. Three caveats sit on the other side.

First, DeepSeek has not disclosed gross margin at these prices. The claim that Ascend 950 closes the cost gap is asserted by the lab and amplified by sympathetic press. It is not demonstrated. High-Flyer has the balance sheet to subsidise V4-Pro indefinitely if it chooses to; permanence in pricing is not the same as profitability in pricing. The frame predicts a structural cost advantage. The evidence shows a price. Those are not the same thing.

Second, the API rack rate is not the enterprise rate. Anthropic and OpenAI do not compete with DeepSeek on the public price card for their largest contracts. Claude Enterprise bundles compliance, audit logs, SLA guarantees (service-level agreements: contractual uptime and response guarantees), data-residency commitments, and integration support. GPT-5 enterprise pricing is opaque and tier-negotiated. The public token price is a marketing artefact for a meaningful slice of the labs' actual revenue. That slice is not on the table.

Third, data residency. EU enterprise, US federal procurement, regulated financial services, and healthcare are largely closed to Chinese-hosted inference regardless of price. The TAM (total addressable market: the maximum revenue a product could reach if it captured every potential buyer) for residency-unconstrained workloads is real but not the whole market.

What this is a case of. This is the third move in a sequence. First, Chinese labs demonstrated they could train frontier-class models on restricted hardware (DeepSeek V3, late 2024). Second, they demonstrated they could match Western capability on reasoning benchmarks (V4, early 2026). Now they are demonstrating they can sustain frontier inference at a price point that breaks the Western margin assumption.

Each step is consistent with the intelligence-explosion-signals frame, which watches for compounding algorithmic and hardware progress that compresses the capability gap faster than incumbents can defend their pricing. The signal here is not that DeepSeek is better. It is that the cost of being roughly as good has dropped by an order of magnitude, in public, with permanence.

What I'd watch. Three things.

One: whether Anthropic or OpenAI move on their public API pricing in the next 90 days. They probably will not match $0.87. The question is whether they cut Opus and GPT-5 output rates by 30 to 50% to narrow the visible gap, which would itself signal pressure.

Two: whether any tier-one enterprise, a bank, a retailer, a SaaS platform, discloses meaningful production workloads routed to DeepSeek for non-residency-constrained tasks. Batch document processing, code-review pipelines, retrieval pre-processing. Those are the workloads where the price differential is large enough to override procurement inertia. None has been announced. One announcement changes the conversation.

Three: independent Ascend 950 benchmarks at MoE inference scale. The cost story rests on a hardware claim that has not been third-party verified at enterprise load patterns. If the throughput holds, the frame strengthens. If it degrades, DeepSeek's permanence is a subsidy, not a structure.

The reading I'd hold lightly: token-API revenue as the load-bearing revenue line for frontier labs is on borrowed time, and the pivot to seats, platforms, and outcome-based contracts is now urgent rather than optional. The labs that have not started that pivot will find the next twelve months harder than the last twelve.

Footnotes

¶