Executive summary

Meta priced an open model at $0.10 per million tokens on its contributor tier, with cache reads at $0.002 per million tokens, according to Nathan Lambert's Interconnects.ai analysis. The infrastructure underneath that model is pricing in the opposite direction.

The model layer and the infrastructure layer look like one price. They are decoupling in opposite directions.

Over the same stretch, NVIDIA's gross margin rose from 72.4% to 75.0%, and Data Center revenue grew 117% year-over-year to $89.0 billion, according to NVIDIA's Q2 FY27 earnings release filed with the SEC.

Three implications for capital allocators:

Model-layer savings are real and can be negotiated. What the next budget meeting has to establish is whether those savings are being reinvested into compute lock-in or measured separately from it.

For practitioners

Run the math on a self-hosted Llama or Kimi K3 deployment and the checkpoint's price disappears into rounding error. Moonshot's Kimi K3 ships as a 2.8-trillion-parameter mixture-of-experts model with 16 of 896 experts active per token and a 1-million-token context window, according to Kingy AI's analysis of AI capex trends. Serving a model at that scale requires GPU infrastructure sized to the active parameter count and context window, not to the download size of the checkpoint.

This is where "just self-host it" stops being a cost argument and starts being a capacity-planning problem. Jensen Huang said as much without saying it directly: on the Q4 FY26 call he credited "Grace Blackwell with NVLink" for "an order-of-magnitude lower cost per token," attributing the gain to interconnect hardware, not to any shift in model pricing, according to NVIDIA's Q4 FY26 earnings release. The efficiency lives in the fabric between GPUs. It does not live in the weights file.

Practically, this changes what "evaluate open weights" should mean as a project. The comparison that matters is not API price per million tokens against zero. It is API price per million tokens against a fully loaded self-hosting cost: GPU-hours at whatever NVIDIA's channel is currently pricing them, plus the engineering time to build a serving stack.

The UK's Department for Business and Trade self-hosted an LLM specifically to keep sensitive trade data from leaving its perimeter. That was a security and sovereignty rationale, not a cost one, according to the UK government's digital trade blog. That is the honest use case for open weights today: control over data residency and model behavior, purchased at whatever NVIDIA charges to run the hardware underneath it. A team evaluating a self-host migration should price the GPU allocation first and treat the checkpoint as free, because it already is.

Deep dive

Nathan Lambert's own framing, buried in his mid-2026 analysis of the open-closed gap, undercuts the tidiness of the capex-reallocation story. "Supply is fully dictated by economics," he wrote, according to Nathan Lambert's Interconnects.ai analysis. That collapses a distinction the capex-reallocation thesis leans on. If model supply and infrastructure economics are the same variable, then a checkpoint's price and NVIDIA's margin are not two decoupled series worth tracking separately. They are one series, observed from two ends.

The counter-evidence to that view is Anthropic. A lab that could not afford to compete on price would produce exactly this pattern: high list prices, weak uptake reported by the Financial Times on its top tier, and public signals of capacity constraint rather than pricing discipline. If Anthropic's restraint is a compute problem and not a strategy, the price war among its competitors says less about model economics than about who currently has enough GPU allocation to give tokens away.

That reframes the CIO's renewal leverage. A price match from OpenAI or Meta on a procurement desk this quarter is not evidence that model access is now abundant. It may be evidence that OpenAI, Meta, and a handful of Chinese labs currently have enough spare capacity to subsidize the model layer while they fight for developer share, in a way Anthropic currently cannot. Lambert's own comparison of open and closed models describes the capability gap between them as "intertwined" with the funding dynamics behind each lab, according to the same analysis. Today's discount could be a temporary function of which lab has the compute headroom to burn, rather than a permanent feature of the market.

There is a version of the next twelve months in which this resolves in the CIO's favor anyway. If Vera Rubin-class hardware delivers the kind of interconnect gain Huang described for Grace Blackwell, and if that gain reaches cloud GPU rental pricing rather than staying captured in NVIDIA's margin, the self-hosting math in a Kimi K3-scale deployment improves independent of what any model vendor charges per token. An efficiency gain and a margin expansion happening in the same hardware generation, reported by the same company, is consistent with a story where NVIDIA passes some of the cost-per-token improvement to customers and retains the rest as margin. The filings do not break out which share goes where, and no source here does either.

The Hugging Face deal adds a second channel through which that retained share could grow. The Register's framing of the deal, that NVIDIA is "tightening its grip on the ecosystem surrounding open models" while pledging not to "squeeze too hard," names the mechanism without resolving whether it will be used, according to The Register's coverage of the acquisition.

The acquisition is not scheduled to close until the first half of 2027, subject to regulatory review, which means the year in which Hugging Face's product roadmap would visibly bend toward NVIDIA's stack, if it does, has not started. The comparison that would settle the question, a self-hosted Kimi K3 or Llama deployment's dollar-per-million-token cost, measured the way OpenAI's and Meta's API prices are quoted, does not exist in any public source cited here.