Deploy a model
Tell us the model, the shape of the work, and how fast it has to be. We show you where it should run and why, before anything is provisioned.
Model source
Where the weights live. We price against the architecture, so a fine-tune is priced as its base.
Deployment details
What this deployment is called, and what kind of work it does. The workload class sets the shape and the default bound.
Used in the endpoint, in logs, and on the receipt.
What the service does, not how many tokens it moves.
Constraints
Anything you declare here is treated as a bound and never optimised away. Everything you leave blank, we choose and justify in the plan.
The only policy input. Declaring it once authorises every later move that holds it.
Optional. A placement above this is excluded even if it is the cheapest that meets your bound.
Only if you are contractually stuck with one. Pinning a region narrows the search and usually costs money.
Spot is offered only where the workload class declares it tolerates interruption.
Preview
Exactly what we will send. The console posts this to the placement API, and you can post the same body yourself.
Plan
Every placement we considered
Ranked by cost per served token. A placement that misses your bound has no cost per served token, so it says undefined rather than a number.
| Placement | $/MSVT | p99 first token | Basis | Why we believe it |
|---|
Calling it
The endpoint is OpenAI-compatible, so anything that already speaks to a chat completions API speaks to this without changes.
Overview
What the corpus knows, and what your placements cost against it. Everything here is denominated in served tokens, because GPU-hours are what you rent and served tokens are what you get.
What choosing wrong costs
One model, every placement that meets the class bound, cheapest against dearest. Computed from the corpus, no traffic required.
Your first command
Nothing here needs an account. The plan is real, and the credential stays on your machine.
Spend
Served tokens delivered inside your bound, and what they cost.
pilot
When it last ran, what it checked, and what it decided.
Recent activity
Every event across the account, newest first.
Corpus freshness
When the physics last moved, and whether the published Index has caught up.
Analytics
Four questions, in the order they get asked. Three of them need traffic, and say so until there is some.
Cost per served token, over time
Not cost per GPU-hour and not cost per request. A token that missed its bound is not in the denominator.
Compliance
The share of output delivered inside the bound. This is the difference between tokens generated and tokens you can sell, and it is the number no other platform reports.
What eviction cost
Spot is cheaper and work in flight is redone. We charge you for the redone work rather than absorbing it, so it has to be visible.
Corpus coverage for your workloads
How much of what we quoted you rests on a measurement rather than a specification sheet. Available now, because it does not need traffic.
Batch jobs
Work with no user waiting. We price the whole job before you submit it, and run it on capacity nobody else can safely use.
New batch job
An input file, a model, and a deadline. Everything else we choose and justify before you commit.
JSONL, one request per line, in the OpenAI batch format. We count the lines and the tokens to price the job.
Priced against the architecture. A fine-tune costs what its base costs.
The only constraint that matters for batch. A longer deadline buys cheaper capacity, and we show you how much.
Spot is roughly a third the price. Work in flight is redone, which costs time rather than correctness.
Files
Input for batch jobs and output when they finish. JSONL in, JSONL out.
Settings
Keys, providers, and how we bill.
API keys
For the placement API and for submitting jobs.
There are no keys, because there are no accounts. The placement API is open and holds no customer data: it answers where a workload should run and creates nothing. Keys arrive with provisioning from the browser, which is the first thing worth protecting.
Capacity providers
Where we rent. Your credential stays with you: we assume a role, we never store a key.
How we bill
Two lines, never one. You see what the capacity cost and what we charged, separately.
A token that missed your bound is not billed. That is the unit doing its job: we are paid for delivered work rather than for elapsed time, so a placement that underperforms costs us before it costs you.
Deployments
Every workload we are holding, where it runs now, and when pilot last checked whether that is still the right answer.
Receipts
What moved, what it saved, and the evidence. Realized, available and foregone are three different things and are never added together.