A market for AI tokens
Buy tokens below list.
Sell the ones you're sitting on.
Every AI company buys inference by the million and has no market to sell it into. The desk is a counterparty on both sides — it buys committed capacity from holders below list, and resells it to buyers under list, with the spread published rather than buried in an enterprise contract.
Indicative list rates today. Desk bid and ask are set when the market opens.
- Spread published, not buried
- Sell below list, buy under list
- Card, bank or stablecoin
The board
Published list rate per million tokens, across the ten models the market quotes. One million tokens is roughly three quarters of a million words.
| # | Model | Vendor | Input / 1M | Output / 1M | Blended* | Context | Tier |
|---|---|---|---|---|---|---|---|
| Loading prices… | |||||||
*Blended assumes a 3:1 input-to-output split, which is roughly what a typical production workload runs. Figures are each provider's published list rate, not a desk quote. The market opens soon and this board becomes live bid and ask.
Size it
What does your volume actually cost?
Almost nobody is billed a single clean rate. Input and output are priced separately, and output costs several times more. Move the split and watch the bill move with it — this is the number that matters when you size a commitment.
- Pick any model from the board
- Set monthly tokens and your input/output mix
- See cost per million, per million words, and per day
When the market opens, this same calculator shows the desk bid and ask alongside list, so the saving is a number rather than a promise.
Output share — output is the expensive half
Indicative list pricing. No desk quote is implied.
Two sides, one counterparty
The desk takes the other side of the trade, which is what makes it a market rather than a price list.
If you buy
You name the model, the volume and the input/output mix. The desk quotes per million against a live reference and settles through whichever rail you prefer. No annual lock-in, no seat negotiation, no procurement call.
If you hold
Committed spend, an unused enterprise agreement, spare inference capacity, or a model you no longer use — all of it is an asset here. The desk buys it below list and pays out in your currency.
How the spread works
The desk buys at one price and sells at another. That margin is how the market pays for itself — and it is published on the board, so you can always see what the desk is making on your trade.
What you get
A counterparty instead of a vendor relationship: a quote you can compare, a price you can walk away from, and settlement in the instrument you already use for business.
Settlement
Settle the way your finance team already does
Three rails, quoted per desk when the market opens. Nothing exotic, nothing that needs a new banking relationship to start trading.
Status
The market opens soon
Right now the board shows published list rates, so you can size the opportunity. When the market opens, the same board carries live desk bid and ask, the calculator prices your volume against them, and trades settle.
Nothing on this page is a quote, an offer, or a commitment to transact.
Questions
What is a market for AI tokens?
AI models are metered in tokens, usually priced per million. Most companies buy that capacity on one vendor's own terms, and companies with committed volume or spare capacity have nowhere to sell it. The desk is a counterparty on both sides.
What does a million tokens actually buy?
About three quarters of a million English words, or roughly 150 to 200 pages. On a flagship model that ranges from a few dollars for input to well over a hundred for output, which is why the split matters more than the headline rate.
How do I sell capacity I already have?
Committed spend, an unused enterprise agreement, spare inference capacity, or a model you have outgrown. All of it is an asset on this market, and the desk buys it below list.
Is the market live yet?
Not yet — it opens soon. The prices on this page are each provider's published list rate, shown so you can size the opportunity. No desk bid or ask is quoted until the market opens.
How do I settle?
Card for smaller desks, bank transfer for invoiced volume and net terms, and stablecoin where the counterparty prefers it. All quoted per desk at open.
Why does the input and output split matter so much?
Output tokens cost several times input on nearly every model, because generating them takes more compute than reading them. A workload that is 75% input runs dramatically cheaper than one that is 75% output on the same headline rate.
The market opens soon.
Get the desk's opening terms before the first print.