Rubin's real claim is a 10x cut in the cost of a token

NVIDIA's Vera Rubin platform is in full production: seven chips, five rack-scale systems and one supercomputer, with partner availability in the second half of 2026 and AWS, Google Cloud, Microsoft and OCI among the first to deploy instances. The headline figures are up to a 10x reduction in inference token cost and a 4x reduction in the number of GPUs needed to train mixture-of-experts models, both measured against Blackwell.

The training number will get the attention. The inference number is the one that changes what you can build.

Training cost is a capex conversation. Inference cost is a product constraint.

Training efficiency determines what a well-funded lab can afford to produce. It affects the model supply, and almost nobody reading this is training frontier models.

Inference cost determines what the rest of us can afford to run in a loop. That distinction is doing enormous work right now, because the architectures currently being designed — agentic systems that plan, call tools, check their own output and retry — consume tokens at a multiple of what a single-shot request ever did. An agent that reasons over five steps and verifies each one is not five times more expensive than a chatbot turn. With retries, tool-call context and the verification passes, it is frequently an order of magnitude more.

The reason most verification is missing from production agent systems is not that teams think it is unnecessary. It is that it doubles the bill.

What a 10x cut actually unlocks

Every safety and reliability technique that works today is one people decline to run at scale because of unit economics. Self-consistency sampling. Independent verifier passes. Adversarial self-critique. Running the same task through two different models and comparing. All well understood, all effective, all routinely cut in the first cost review.

A 10x reduction moves those from a research luxury to a line item. That is a larger change to what a deployable autonomous system looks like than any capability improvement announced this year — because the blocker was never that we lacked techniques for making agents more reliable. It was that reliability was priced as a premium nobody would pay.

The concentration worth naming

Reporting puts major hyperscalers each deploying close to 1,000 NVL72 racks — roughly 72,000 Blackwell GPUs — every week. Whatever else that describes, it is a supply chain with one dominant input, and the pricing power sits upstream of every company building on top.

For anyone planning a multi-year autonomous system, that is a strategic dependency worth stating in the risk register rather than leaving implicit in a cloud bill. The 10x is real and it is not yours — it is a vendor's improvement, passed through at a rate the vendor chooses.

What we would do with it

If your inference costs fall materially over the next year, the question is what you spend the saving on. The tempting answer is more scale: more users, more calls, more autonomy. The better one, for anything operating where failure is expensive, is to buy back the assurance that was cut for budget reasons — put the verifier pass back in, run the second opinion, keep the self-consistency samples.

Cheaper tokens are an opportunity to make systems more defensible, not merely larger. Very few roadmaps will spend them that way.

Written by the Devence Lab research team.