The fee
WARP trades against tokenised NVDA on a Pons curve. Every buy and every sell pays 3% of its value, charged in NVDA. That 3% splits two ways: a 2% creator tax, which is the only thing that funds holder compute, and a 1% Pons fee, which funds the launchpad and operations.
Per side, not per round trip
A round trip is two events and the curve charges each. Someone who buys and later sells has paid twice. The headline number is the one people quote; the effective number is the one people feel, so both are printed.
The creator tax is the only inflow to compute. No tokens are issued to fund it, there is no treasury of WARP to sell, and there is no emission schedule. If nobody trades, nothing is collected and the hourly settlement divides zero.
Buying capacity
The 2% creator tax arrives in NVDA. It is priced in dollars using the Chainlink NVDA feed, so the figure the purchase is made against is the same figure anyone else can read, and it is then spent on RTX 5090 capacity at the current rate.
The 5090 price is updated weekly. Hours are fixed at the price in effect when they are credited — a later price change does not revalue an hour already in your ledger, in either direction.
Why compute and not cash
Paying holders cash in proportion to their holdings is the shape of a security nearly everywhere anyone is reading this. Compute is a good: you receive access to a machine, priced at what the machine costs. That distinction is not a wrapper around the same thing — it is the load-bearing wall of the entire design, and everything else is built to keep it true.
The hourly settlement
Once per hour, every wallet holding at least 1,000 WARP is recorded, and the capacity purchased in that window is allocated by each wallet's share of eligible WARP — the sum of balances at or above the threshold, not total supply.
Wallets under the threshold are not in the denominator. They neither receive nor dilute.
Why a snapshot rather than a stream
A balance that flickers inside the hour cannot be paid twice, and the arithmetic can be checked afterwards against a block. An hour is short enough that buying in does not mean waiting a week, and long enough that the distributor is not writing constantly.
Nothing is staked. No deposit, no lock-up, no escrow, no approval to sign. Balances are read at a block and left alone. Sell in the same minute if you want — you are simply not in the next snapshot, and what is already in your ledger stays there.
Eligibility
The threshold is 1,000 WARP, measured at the snapshot block. It exists so the ledger is not filled with wallets owed a millionth of a second, which costs more to record than it is worth.
There is no upper cap and no tier ladder. A wallet holding a hundred times more receives a hundred times more, because the allocation is a proportion and dressing it up as anything else would be a lie about what proportion means.
The ledger
Credits accrue to your address, denominated in RTX 5090 hours. They do not expire. They are not pushed anywhere. They are not transferable between addresses. They stay in the ledger until you claim them onto a key.
Keys
Claiming attaches credits to an API key. Each key holds up to 40 RTX 5090 hours. You can hold several, name them, and revoke any of them — revoking detaches a balance, it never burns one.
Anyone holding a key can spend its balance. There is no second factor on a bearer token. Treat it the way you would treat any other API key with money behind it.
The gateway
The gateway speaks the OpenAI chat protocol. If your application already talks to that shape of API, changing the base URL and the key is the whole integration.
curl https://gateway.warp.example/v1/chat/completions \ -H "Authorization: Bearer $WARP_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"qwen2.5-7b","messages":[{"role":"user","content":"hello"}]}'
The seconds a request occupies a GPU are deducted from the key's balance. There is no card on file, no minimum and no monthly commitment, because there is nothing to bill — you are spending an hour a trade already paid for.
Pods
A pod is a whole 5090 with a root shell. Choose an image, it boots, you SSH in, and it bills by the second it is running.
warp pod create --gpu 5090 --image pytorch:2.6-cu126 ssh pod-9f2a1c@ssh.warp.example
A stopped pod bills nothing, and a persistent volume survives a stop so you are not re-downloading weights each session. Pods are the right surface when you want the card rather than an endpoint — training, a custom kernel, anything the gateway's model list does not cover.
What credits are not
They are not cash and cannot be redeemed for it. They are not transferable. They are not a yield, an APY, a dividend, or a claim on revenue. Nothing about them is minted, and none of them is an NFT.
They are a prepaid balance on a machine, funded by fees somebody already paid. The amount depends on trading volume, on eligible supply, and on the price of a graphics card — and all three can move to a level where the answer is close to nothing.
No APY appears anywhere on this site. Not in the docs, not in the estimator, not in a post. A rate implies a promise about the future and there is no mechanism here that could keep one.
Risks, stated plainly
Volume is the only inflow
A quiet month collects little and purchases little. That is the design working exactly as written, not a failure.
GPU prices move against you
The same tax buys fewer hours when 5090 rental gets more expensive. You are exposed to an input cost nobody here controls.
The venue holds a lever
A launchpad can generally redirect its own share of trading fees by an owner action on its factory. Where that is true it belongs here, with the notice period and who can execute it, rather than in a contract nobody reads.
The compute has to actually exist
An hour is only worth something if the gateway and the pods are up. That is an operational dependency on real machines, and it is the part of this system that is not on a chain and cannot be.
Glossary
| Term | Means |
|---|---|
| Creator tax | The 2% of each trade that funds holder compute. Two thirds of the 3% fee. |
| Pons fee | The remaining 1%. Funds the launchpad and operations. It never enters the compute pool. |
| Eligible WARP | The sum of all balances at or above 1,000 WARP at the snapshot block. |
| Settlement | The hourly allocation of purchased capacity to eligible wallets. |
| Credit | One unit of prepaid compute, denominated in RTX 5090 hours, fixed at the price of its hour. |
| Claim | Attaching credits from your ledger to an API key. Your action. |
| Gateway | The OpenAI-compatible endpoint credits are spent through. |
| Pod | A dedicated RTX 5090 with SSH, billed by the second while running. |
