Pinout Compute is live on Hedera testnet. See the machines and prices

Compute your agent buys for itself.

Your agent picks the machine, signs one x402 payment, holds it for as long as the job needs, and hands back what it did not use. Settled on Hedera, so the bill it pays is one anybody can recheck.

Install on your agent
OpenAI CodexClaude CodeOpenClawHermesGemini CLIGrokAntigravityOpenCode
orTry Now (Hosted Agent)

It pays for 41 seconds

Not the hour most hosts bill for them. The meter counts seconds and stops the moment the job does.

Runs dry, keeps the work

When credits hit zero the machine is held, not destroyed. The agent tops up and the job continues from where it stopped.

It gets the change

Unused seconds return to the wallet the moment a session closes. No expiry, no leftover balance to write off.

Built on the rails that make this possible.

Hederax402Hedera Consensus ServiceHIP-991Mirror node verified

Why compute needs its own kind of payment

x402 settles one request at a price you already know. Compute does not work like that. Nobody can say how many seconds a job needs until it has run, which leaves you guessing up front or signing thousands of tiny payments as you go.

Pinout sits in the middle. You buy a pool of credits once, a meter draws them down while the work happens, and whatever survives is refunded when the session closes.

One job, start to finish

Watch an agent rent a graphics card

You hand over a job and a spending limit. It works out what hardware that needs, pays for it over x402, and gives you back whatever it did not use. This is a real session from the deployed service, replayed.

you

Cut me out of this clip and put a gradient behind me.

clip.mp400:10 · 1080p · 14.2 MB
budget ceiling set to 0.06 ℏ
  1. You

    You ask

    A job and a spending ceiling. Nothing about which machine to use.

  2. Agent

    It picks the machine

    Reads the catalogue, matches the job to a lane, and signs one x402 payment.

  3. Agent

    The card runs

    Seconds are checkpointed to a Hedera topic as the work happens.

  4. Hedera

    It runs dry

    Credits reach zero. The machine is held and the meter stops.

  5. Agent

    It tops up

    A second x402 payment, and the same process carries on from where it stopped.

  6. Agent

    Work finishes

    The job completes and the machine is handed straight back.

  7. Hedera

    Hedera refunds the rest

    Seconds bought but never used return to the agent's wallet on chain.

  8. Hedera

    The bill is on chain

    Settlement anchors to a HIP-991 topic that anyone can recompute from the mirror node.

Machine pricing

Machines powerful enough for any task

Your agent reads this catalogue and picks what the job needs. Prices are all-in: accelerator, vCPU and memory, billed per second held. No minimums, no egress fees, no reservation.

See the machinesRead the source
>80GB VRAM
B300 SXM6
287 GB VRAM192 GiB RAM16 vCPUs
6,209,000 tinybar/s
→
B200
191 GB VRAM128 GiB RAM16 vCPUs
5,149,000 tinybar/s
→
H200
150 GB VRAM96 GiB RAM12 vCPUs
3,796,000 tinybar/s
→
A100 SXM4
85 GB VRAM64 GiB RAM8 vCPUs
2,296,000 tinybar/s
→
24 - 48GB VRAM
L40S
48 GB VRAM32 GiB RAM8 vCPUs
1,712,000 tinybar/s
→
A10
24 GB VRAM16 GiB RAM4 vCPUs
912,000 tinybar/s
→
L4
24 GB VRAM16 GiB RAM4 vCPUs
778,000 tinybar/s
→
T4
16 GB VRAM8 GiB RAM2 vCPUs
474,000 tinybar/s
→
CPU ONLY
cpu-4
8 GiB RAM4 vCPUs
147,000 tinybar/s
→
cpu-2
4 GiB RAM2 vCPUs
74,000 tinybar/s
→
cpu-1
1 GiB RAM1 vCPUs
30,000 tinybar/s
→
Use cases

What agents run here

Inference

Serve a model for the seconds a request needs, with no idle hour attached.

Training

Long runs that top up mid job rather than dying when a budget lapses.

Autonomy

An agent with a wallet buying its own hardware, inside a ceiling you set.

Media

Diffusion, matting and vision work, billed for the seconds it actually took.

Why Hedera

Why this runs on Hedera

Four platform properties make per-second settlement practical. No other network offers all four.

Your agent never holds gas

Hedera's facilitator signs as fee payer and submits on the agent's behalf. The agent's balance moves by the price and nothing else, so it carries no separate gas float to monitor or refill.

The network keeps the ledger

Consumption is checkpointed to a Consensus Service topic with a running hash maintained by consensus nodes. The party issuing the bill is not the party building the audit trail.

Settlement that costs to publish

Every settlement is anchored to a HIP-991 fee-charging topic at real, unrecoverable cost. Publishing final numbers is expensive by design, which makes inflating them expensive too.

Auditable by anyone, free

Every record sits on Hedera's public mirror node. No key, no signup, no rate-limit agreement. And no smart contract anywhere in the system, both tiers are Consensus Service.

Billing

See exactly what you paid for

Every session is itemised down to the second, with the charge and the refund recorded as they happen. Below is live activity, straight from Hedera.

Settled sessions, read from Hedera just now
reading the topic…

Reconcile any invoice yourself with the open-source verifier, or hand a session id to your finance tooling and let it check the arithmetic.

# rebuild the bill from public data, not from our word for it
$ npm run verify -- 3c145a57-d8d5-4c7e-a974-9f92426e075e

  PASS  contiguity      contiguous [0, 141) across 5 checkpoints
  PASS  tier-binding    anchor commits to seq 142 + running hash
  PASS  arithmetic      141 x 30000 = 4,230,000 tinybar
  PASS  burn-count      ledger count 141 matches events received
  PASS  commitments     all 5 checkpoint commitments match

  VERIFICATION PASSED  bill recomputed from the mirror node
Questions

Answers.

What if an agent runs out mid-job?

The machine is held and the meter stops, so the wait costs nothing. The agent tops up and the job resumes from where it stopped. If no top-up arrives within the grace window the machine is released and the balance refunded.

Do I need HBAR for gas as well as for the work?

No. On Hedera the facilitator pays the network fee and submits for you, so your agent holds exactly what it plans to spend and nothing else.

Is there a contract to audit?

None. Metering and settlement both run on Hedera Consensus Service, so there is nothing to deploy, upgrade, or audit for reentrancy.

Is this production ready?

It runs on Hedera testnet today, against real CPU and GPU hardware. Mainnet is the next step. Current limits and economics are documented in the repository.