AI agent GPU sourcing: discovery to quote request
AI-narrated product walkthrough. 30 September 2026.

Discovery is the first step
An AI agent can help a team discover GPU capacity and prepare a supplier quote request. It should also be precise about what has happened. Finding an offer is not reserving it. Submitting a request is not provisioning a machine. In this Machine Age AI walkthrough, we follow the supported path from public discovery to an authenticated quote request. We use the actual guide and developer interface. The example request contains placeholders rather than credentials, and we do not send a purchase or payment. The service description is dated the thirtieth of September, twenty twenty six.

Read the current capability response
Begin with the capability response at slash A P I, slash agent manifest. Read it before deciding what actions to enable. Then read slash A P I, slash live inventory. Both are public, read only requests and do not require an account credential. The developer page provides a button that runs the inventory request and displays a sample response. Current capabilities include discovery and quote requests. Automatic provider provisioning, tokenisation and secondary trading are unavailable, and native checkout is on hold. An integration must not treat a documented future action as something it can perform today.

Inspect freshness and price basis
Inspect the observation time and expiry time before using an offer. If the response has expired, refresh it. Check the source status rather than assuming a successful HTTP response means every provider is available. For an individual offer, record the GPU model, memory, quantity, region, storage and full instance price. Do not multiply a complete instance price by the number of GPUs again. Provider pools need different treatment: a pool can report aggregate capacity across several models and require a bid. Its total does not prove that your exact GPU model or requested configuration is available.

Build an explicit shortlist
Turn the owner's workload into a small set of explicit filters. Collect the model, framework, memory requirement, GPU count, region, start window, duration and budget. Ask whether several GPUs need to share a machine and what communication between them the job requires. Keep those assumptions visible in the shortlist. If performance has not been measured, say so. Do not claim that two machines are equivalent just because the GPU label matches. Include acceptable alternatives, such as a different start time or region, only where the owner has allowed them. That makes the later supplier conversation more useful.

Use an owner-issued permission
A public inventory read is different from a request made for an account. The account owner signs in to the agent workspace and issues an expiring permission with the scopes the integration needs. For this flow, those are quote reading and quote requesting. Keep the key in a secret manager and send it only in the authorization header to the intended service. Never put it in a URL, source code, public logs or the quote description. The key is an account permission. It is not a wallet private key, a capacity reservation, or authorization to exceed the owner's stated task.

Submit the request once
The request body uses the request quote action, a fresh UUID, a current offer identifier, the workload, a future UTC start time, and the duration in whole hours. Replace every placeholder in the guide before sending. For an Akash pool, also provide a listed GPU model and a positive GPU count within the reported total; the supplier must still confirm the selected model. If the account has no verified email, include a valid contact email. After a timeout, reuse the same UUID with the identical body. Creating another identifier for the same uncertain request can create duplicates.

Report what is still unconfirmed
The result is a quote reference with awaiting provider status. Present that reference to the owner, together with the requested scope and everything still unconfirmed. Retrieve the status through the authenticated trading endpoint. Machine Age AI reviews availability, pricing and delivery with suppliers; the request itself does not return a provisioned GPU endpoint. Keep supplier confirmation, commercial agreement and tested access as separate events in your own workflow. The full walkthrough and machine readable guide are linked from machine age dot A I slash developers. Start with a read only integration, then add the authorized request step once its boundaries are clear.