Generating the Answer Got Cheap. Testing It Did Not.

Yesterday we argued that AI vendors have quietly stopped selling capability and started selling speed and price (the tokens-per-second pitch). Here is the other half of that trade. While text models fight a price war, the money that still wants a margin is moving to places a price war cannot reach.

Look at how CuspAI packaged its round. The company raised a $450 million Series B for its MIRA materials discovery platform, and the announcement did not lead with the model. It led with the AI Materials Foundry, a network of more than 45 members that includes Nvidia, Meta and a set of major industrial labs.

That ordering is the story. Generating a candidate material is now cheap. Making it and measuring it is not.

## The economics quietly inverted

In text and code, generation and verification cost roughly the same, which is to say almost nothing. You can retry a hundred times before lunch. That symmetry is why language models improved so fast and why they are now competing on price.

In materials, biology and hardware, generation collapsed to near-free while verification did not move at all. A candidate alloy still has to be synthesised. A device still has to be put on a patient. A structure still has to survive a launch. When one side of a loop gets a thousand times cheaper and the other side does not, the bottleneck relocates, and so does the profit.

Which is why the interesting AI deals of the past month were not model deals.

Japan committed up to one trillion yen over five years to its Noetra programme and a national AI factory built on Nvidia hardware. The word attached to it is "physical AI", not chat. Siemens is putting its Xcelerator stack inside the design loop for reusable spacecraft. In India, a small seed round went to Bioscan Research for AI-enabled brain injury detection devices. That last one is a rounding error in dollar terms, and it is instructive anyway: the product is an instrument, and it will be judged on clinical measurement rather than on a benchmark.

## The hedge

This is not a claim that language models are finished. The scale is not close. CuspAI's $450 million sits against tens of billions moving through compute and infrastructure deals in the same window. AI-for-science also has a long history of impressive demos that never survived contact with a lab, and one well-structured round does not reverse that record.

The physical world pushes back on the compute side too. Texas has ordered its grid operator to audit every new data centre before it can connect, with an interconnection queue of roughly 474 gigawatts, about 90 percent of it data centres. Announcing capacity and energising it are different activities. The same gap between the claim and the physical result runs through this entire thesis.

So treat what follows as a lens, not a forecast.

## What to do with this

**Check who owns the test loop.** Next time you read an AI-for-science announcement, look past the model for the verification asset: a foundry, an instrument, a wet lab, a clinical channel, a manufacturing partner. If the round funds only the model and leaves testing to someone else, the company has bought the cheap half of the problem and left the expensive half on the table. CuspAI buying a consortium on day one is the tell worth copying.

**Watch for the second and third one.** One round shipping with a factory network attached is an outlier. Three in a quarter would mean the market has repriced verification as the scarce input, and that would change what an AI company is expected to own. We track these as they land in the deal tracker; it is a cheap pattern to watch and an expensive one to miss.