The State of Compute & Hardware APIs

About this paper

Every other report in this series covers one industry. This one covers a stack, and the finding is the distance between its floors. The companies that rent access to chips score 36.5. The companies that design and fabricate those same chips score 10.7 — and on agent-readiness the gap is nearly ten to one. Eighty-one percent of the silicon layer sits in the Minimal band: an index entry, a description, a link, and nothing a machine can consume.

One organization settles the question of whether this is about capability. NVIDIA scores 31.1. NVIDIA NIM — the inference software NVIDIA sells on top of its own silicon — scores 75.7, the highest score in the market. Same company, same engineers, 44.6 points apart. What separates them is which product got treated as a product. The rest of the report is what that costs, what the leaders publish that nobody else does, and why the most-published resource in the compute and hardware market turns out to be models.

What's inside

1. Executive summary — a twenty-six point gradient down one stack
2. How this cohort was built, and why it matters
3. The state of the market
4. The resource taxonomy — whose vocabulary is this?
5. The composite Kin Score, and why it's the foundation
6. Agent-readiness — the market an agent cannot reach
7. The blueprint
8. The standards layer — the stack has standards, and they stop at the rack
9. The regulatory layer — the most-regulated market in the series
10. Standards and artifacts for the agentic turn
11. Security, auth, scopes and compliance
12. Provider-by-provider intelligence
13. The investable thesis
14. Where this is going

What you get for $500.00