How Compute works.

A launchpad with compute behind it.

Compute is live on Solana mainnet. Launch a Pump.fun coin with one wallet signature, trade it, and put its creator fees to work on GPUs, starting with AI images. You sign every payment in your own wallet.

Each project has its own coin, access rule and GPU budget. Compute credits are separate from any coin. There is no $COMPUTE token.

Compute has not had an independent audit. See what’s live today for exactly what works now and what is coming live soon.

Launch a coin in one signature.

Create a project, configure its application and access, and set a daily job budget. Saving a project is free and does not submit a token transaction.

Pump.fun. One wallet signature creates the coin on Pump.fun and locks its creator-fee split: 50% to the creator, 20% to Compute’s burn wallet, which buys back this coin and burns it (Compute does that by hand; it is not automated), and 30% to the coin’s GPU budget. The split is written into the launch transaction and lives in Pump.fun’s fee-sharing config with its admin revoked, so neither the creator nor Compute can change it. Pump.fun’s fee program, which Pump controls, enforces it, and Compute re-reads the config on chain after the launch to check that it holds exactly these shares. The launch also adds a GPU prepay worth 2 hours of the GPU the creator picks ($1.60 with the default GPU, about 53 images) to the coin’s GPU budget. That is a prepaid budget credit, not a rented machine. Anyone can then buy and sell on Pump’s bonding curve, and on PumpSwap once the coin graduates.

Meteora. Coming live soon. Meteora launches are built and tested on devnet but are not approved for mainnet, so you can’t launch there today.

Existing tokens. Link an existing mint after creator-authority verification. Its launch origin and liquidity stay where they are. Any launch, trade, or transfer requires a separate wallet signature.

Follow the token lifecycle. A linked project’s Trade tab shows saved chain observations, the current venue and the age of an observed migration. An unavailable observer keeps the last result clearly out of date. Migration age starts when Compute first sees that state; every trade still checks the chain again before a wallet review.

Use earned fees to fund useful work.

Every coin launched from Compute splits its Pump creator fees on-chain: 50% to the creator, 20% to Compute’s burn wallet and 30% to the coin’s GPU funding, which pays for the coin’s GPU work. 20% of creator fees go to Compute’s burn wallet, which buys back this coin and burns it. Compute runs that by hand; it is not automated, and the chain does not enforce it. These fees come from token trading. They are not the same as trading volume, and future fees are never counted as money available to spend.

Your 50% is yours. You can also choose to put part of it toward a project’s budget:

There are three separate wallet reviews: collect eligible SOL fees into your wallet, convert an explicitly selected allocation to USDC, then pay the project’s exact USDC invoice. Collection and conversion are each off by default, and in-app conversion also needs a separately reviewed mainnet pool; a saved plan never switches them on. Conversion creates no credits, and USDC beyond the chosen deposit remains in your wallet.

The creator-fee split is locked at launch. Admin rights over the split are revoked, so neither the creator nor Compute can change it. It is enforced by Pump.fun’s fee program. An editable Compute plan has no effect on this on-chain configuration.

The saved monthly cap limits reservations and verified deposits made through this explicit creator-funding flow. It covers the frozen UTC funding month across plan revisions; it does not limit ordinary top-ups, creator-wallet collections or Pump’s native fee split. An unresolved conversion or an issued invoice keeps its reservation until the service can establish a safe outcome. The daily job-spending ceiling remains a separate limit.

Ordinary Pump creator vaults are keyed by creator wallet and may combine earnings from several coins. Shared configurations have their own recipient rules. The fee inspector labels that scope. You explicitly select finalized wallet receipts for a project, and those sources cannot be allocated twice across projects. No automatic funding or token purchase follows from collection, conversion or a saved plan.

Pump fee-sharing specification · Creator fee collection

A funded balance, with limits.

Compute credits are nontransferable prepaid usage accounting, denominated in USD. They are not a token, bank account, yield product, or noncustodial escrow. You prepay them in canonical USDC or native SOL and sign every payment in your wallet. SOL is priced at a live quote locked for 10 minutes, and the treasury holds the SOL. A deposit only becomes spendable after the correct finalized transfer is verified. Project budgets take USDC.

Submitting a job reserves the quoted amount. A successful, validated and durably stored result settles the charge. A failed job releases the reservation. Unclear provider outcomes remain in reconciliation to avoid duplicate delivery and charging.

Each project’s budget stays separate. Credits in a coin’s GPU budget, including what Compute credits to it from creator fees and the launch prepay, can only pay for that coin’s GPU work and can’t be withdrawn. Account balances, access rules, daily caps, and concurrency checks are enforced when work is submitted. An insufficient balance pauses new jobs. Images are charged per finished image, so idle managed GPUs cost you nothing. A rented GPU or a coin’s dedicated GPU is billed by the hour, busy or idle.

Unspent deposits remain service obligations. They are not earned compute revenue and cannot also be counted as money for provider rewards or another project. Credits are prepaid and non-refundable except where the credit policy says so. Whether withdrawals are open is shown in the Withdraw tab of your wallet.

Your GPUs, work, and earnings.

Community GPUs can’t earn yet. Today you can register a Linux/NVIDIA machine and run a connection test. Paid jobs, rentals and payouts for community machines are not live; the steps below describe how they are built to work.

The provider workspace has three views. Overview shows your machines, current jobs, job history, earnings and reported GPU utilization, memory, temperature and power. Machines handles registration, connection credentials and availability. Earnings & payouts shows approvals, accepted compensation and USDC settlement requests.

Not sure what your hardware suits? Check your rig first: paste the diagnose output and see which work fits, without registering.

Register your Linux/NVIDIA node and use the diagnostic worker to report its hardware. This connection process does not run jobs or earn compensation. A heartbeat shows connectivity; it does not verify hardware or prove useful work.

In Machines, reopen connection instructions whenever needed and set the minimum USD pay you accept per image. Paid work stays disabled until you explicitly accept those terms. The server rejects new assignments below your minimum; existing jobs retain their agreed price when you change it. The separate opt-in Docker runner is built to serve one fixed image workload. An authorized operator must approve the exact GPU device, reported inventory, model revision and container digest after reviewing benchmark and isolation evidence. The runtime and model must already be provisioned locally. The runner accepts one job at a time, with a bounded lease and an explicit USD compensation quote.

Use Request workload review in Earnings & payouts to submit the benchmark evidence reference and SHA-256 for a connected node. Your request shows pending review, any information the operator needs, and the review history. You can revise evidence after feedback or withdraw a request. The operator separately records an approval for the exact device and workload; a submitted or closed request never enables jobs by itself.

The GPU image includes a benchmark kit with 100 fixed prompts, separate cold and warm runs, repeated seeds, image review and actual cost records. Run it on your intended hardware before submitting evidence. Its report does not cover the complete container startup or application delivery path, so the operator also reviews those measurements and isolation checks.

Dispatch requires enabled services, a current approval, matching connection reports, user access and a funded reservation. Results pass request/model binding, full image decoding, moderation and private-storage checks. Customer charging and the provider obligation commit together. Failed or expired work releases its reservation and creates no payable.

Accepted work is recorded as an unpaid USD obligation. USDC settlement is reviewed and signed by an operator, and compensation is marked paid only after the exact transfer is finalized. There are no token payouts.

Keep unresolved runner journals so interrupted uploads can reconcile the same result. Pause blocks new assignments; revoked credentials or qualifications block further access or acceptance. Saved power preferences remain advisory. Operator approval and valid output are not cryptographic proof that a particular GPU performed the task; stronger verification and an open network remain future work.

GPU rentals alongside the launchpad.

GPU rentals are an additional product. Project launches, coin GPU budgets and the image studio stay separate.

Compute GPUs. Live. Rent a GPU that Compute rents from its GPU provider, by the hour, paid from your credits; you can pay in SOL or USDC to top them up. Each hour is charged when it starts, and stopping early does not refund the rest of that hour. The GPU runs on a machine hosted by a third party, so do not put secrets or personal data on it. It is deleted with all its files when you stop or your credits run out. A dedicated GPU per coin, billed hourly from the coin’s budget, is live. The bigger models (A100, H100, H200 and B200) unlock for a coin’s dedicated GPU once the coin graduates from Pump’s bonding curve to PumpSwap.

Community GPUs. Coming live soon. The marketplace is built for independently supplied GPUs too, with owners setting their own price per GPU-hour and renters paying it plus a disclosed marketplace fee. That side is not open: community machines can register and run a connection test, but they can’t be rented or earn yet.

When independent rentals open, you will choose a GPU and a compatible workload, review its exact price, and accept the renter policy before you start. Rentals reserve the whole listed machine; lowering a minimum memory requirement does not reduce its price. Your GPU rentals keeps quotes and receipts available for recovery. If a response is lost, recover the original request instead of starting another payment. Stop requests remain pending until cleanup is confirmed; uncertainty does not release held funds or hardware.

There is no Compute token.

No Compute token exists. Compute has not issued a token and is not selling one. Anything sold as $COMPUTE or under a similar name is not from Compute.

Project coins identify individual communities. They can gate access to a project when configured, but holding one does not create credits or an entitlement to free GPU time.

What’s live today.

Model APIs may power agents, but an agent application and its underlying compute are separate layers. Video generation, model training and unrestricted containers are not offered.

Compute has not had an independent audit.

Start with the image studio.

Image generation runs FLUX.1 [schnell] on NVIDIA GPUs in a managed Runpod pool, at $0.03 an image, paid from a coin’s GPU budget or your own credits. The studio accepts a prompt, an output format, and an optional seed, and shows the request price, model identity and execution provider before submission. Results are private to your account unless they were made in a coin’s studio, where they are shared to its gallery by default.

On a project’s Overview, check the listed image formats, price per image, access requirement and community budget before opening the studio. The observation time shows when those figures were checked; a stale observation does not confirm current availability. Holder-only studios show the exact mint and minimum holding in token base units.

In the studio, choose community credits or your own workspace explicitly. On a launched coin, community credits are a daily number of free images for its holders, by the share of the supply they hold (the coin’s Studio tab lists the tiers, and the day resets at 00:00 UTC); the coin’s creator is not limited. Adding credits cannot resolve a missing token holding, exhausted project daily allowance or unavailable GPU service. Refresh to check again, or use the free image editor. If the project’s rules change during a check, review the refreshed settings before continuing.

The free image editor works without external credentials. Add exact text to artwork, choose a square, portrait, or banner canvas, and export a PNG at the displayed dimensions. Local uploads stay in your browser.

Terms, privacy and support.

Terms and credit policy. Credits are prepaid and non-refundable except where the Terms say so. Credits reserved for a job that fails are returned automatically. Read the Terms & credit policy.

Privacy. Images you make in your personal studio are private to your account unless you share them. Images made in a coin’s studio are shared to that coin’s public gallery by default, and you can remove each one from Your generations. Prompts and wallets are never shown. Prompts and generated images are processed by Runpod (image generation) and checked with OpenAI’s moderation service. Rented GPUs run on machines hosted by third parties, and a container does not hide inputs from a machine’s operator. Read the Privacy notice.

Signatures. Saving a project, a machine registration or a funding plan does not submit an on-chain transaction. Every launch, trade and credit top-up needs your wallet signature.

Support and reports. Get help with a payment or a job, or report content such as an image or a coin.