GPU Cluster Data: What an AI Agent Can Verify

By Commertize Desk · Published 19 November 2026

Take one asset: a cluster of 1,024 GPUs in a leased colocation hall, owned by a single sponsor. It has 1.6 megawatts of contracted power. One customer has signed a three-year contract for 60 percent of the capacity, and the rest is sold by the hour. For October the sponsor reports 90 percent utilization.

An AI agent screening compute assets for an allocator, or buying hours for its own workload, cannot take that paragraph on trust. It needs GPU cluster data it can check. The numbers below are illustrative; the method is the point.

The arithmetic the agent runs first

Before any cryptography, the agent tests whether the claims fit together.

| Item | Calculation | Result |
|---|---|---|
| Capacity for a 30-day month | 1,024 GPUs × 720 hours | 737,280 GPU-hours |
| Contracted share | 60% of capacity | 442,368 GPU-hours |
| Reported billed hours | 90% of capacity | 663,552 GPU-hours |
| Hourly sales implied | Billed minus contracted | 221,184 GPU-hours |
| Energy at an average 1.3 MW draw | 1.3 MW × 720 hours | 936 MWh |

Each row gives the agent something to compare. Billed hours cannot exceed capacity. The contracted share should match the contract. The energy figure should sit close to the utility or colocation meter. If the meter shows 500 MWh for the month, a 90 percent utilization claim does not survive, whatever the dashboard says.

What the agent can verify by machine

That the hardware exists and is what the sponsor says. Modern servers can produce signed statements about their own components and firmware. The Internet Engineering Task Force's RFC 9334, Remote ATtestation procedureS (RATS) Architecture (January 2023), names the roles. An Attester (the machine) produces Evidence. A Verifier appraises that Evidence against reference values and the manufacturer's endorsements. A Relying Party (here, the agent) receives the Attestation Result. The agent does not inspect 1,024 devices. It checks that a verifier it accepts has appraised them, and when.

That the usage numbers came from the sponsor and were not altered. A metering endpoint can sign each response. RFC 9421, HTTP Message Signatures (February 2024), defines how a server signs chosen parts of an HTTP message so the receiver can confirm the origin and detect changes. A signed, dated reading of 663,552 billed hours is something the agent can store and show its principal. An unsigned figure on a web page is not.

That a third party stands behind a claim. Some facts are not the sponsor's to assert. The colocation operator can state the contracted power. An auditor can state that billed hours tie to invoices. The World Wide Web Consortium's Verifiable Credentials Data Model v2.0 (W3C Recommendation, 15 May 2025) describes the format: an issuer signs a claim about a subject, a holder presents it, and a verifier checks the signature and the issuer's identity. The agent's question changes from "is this true?" to "who signed it, and do I accept that signer?"

That the readings agree with each other. With signed billed hours from the sponsor and a signed energy reading from the facility, the agent repeats the arithmetic above on evidence, not on claims. Independent sources that reconcile are worth more than one source stated twice. The same reasoning sits behind oracles and proof of reserve for AI agents.

What the agent cannot verify

Machine checks stop at the edge of the machine. Four items still rest on documents and people:

A careful agent reports these as open items to its principal. It does not fill them with assumptions.

What this means for the sponsor's asset

For the sponsor, the list above is a publishing guide. Three steps make a cluster readable by machines:

  1. Publish capacity, billed hours and uptime as signed, dated readings, not as figures in a PDF.
  2. Ask the facility operator and the auditor to issue their statements in a form software can check.
  3. State plainly which facts are documents (title, liens, lease, customer contract) and where an authorized reader can obtain them.

An asset that answers an agent's first ten questions with evidence gets read by more buyers of compute and more allocators, at any hour, with no call scheduled. For how compute and power assets are structured for a digital register, see tokenization for energy and digital infrastructure.

What to do next

If you are building an agent that reads offerings for an asset owner, start with Commertize's machine-readable index at commertize.com/llms.txt.

Prefer email? Send a short asset summary to deals@commertize.com. We reply within one business day, with no obligation to engage us.

Asset class: GPU Compute

Tokenize an Asset