The platformto rent
Rent GPU compute by the minute, with reputation-scored operators and SSH access in under sixty seconds. Pay what the work costs.
Bring your tools.
Run them on real silicon.
Every node is NVIDIA hardware. Every CUDA, PyTorch, TensorFlow, HuggingFace, vLLM, or Ollama workload runs the way it would on a server you owned.
What you actually need.
Nothing you don't.
Sub-minute allocation
Pay, then SSH. The auto-allocator watches confirmed payments every ten seconds, picks an idle GPU matching your tier and region, and hands back credentials in under sixty seconds.
Reputation that is earned, not bought
Every operator carries a transparent score: 60 percent uptime, 25 percent buyer ratings, 15 percent completed jobs. The formula is public. The math is public. There is no pay-to-rank.
Per-minute billing, refunds for what you did not use
Stop an on-demand or spot rental early and the minutes you did not use are refunded the way you paid: to your balance, to your card, or as USDC on Solana to the wallet on file. Usage is billed up to the next whole minute. Reserved terms are non-refundable on early termination.
Ephemeral SSH, no shared keys
Every rental gets a fresh session credential that expires when the rental does. No long-lived passwords, no operator can replay your access after you leave.
Six audiences.
One rate sheet.
Whether you're fine-tuning on a Saturday or running production inference behind an enterprise SLA, the same per-minute meter and the same SSH access path serve you both.
Research teams
Burst training without an AWS bill
Spin up 8x H100 for a weekend of fine-tunes, kill the job, pay for the minutes you used. No quota requests, no reservations, no calendar invite with a sales rep.
Solo developers
A real GPU at indie prices
RTX 4090 or L40S for less than the cost of a model API call. Ship the side project, learn the stack, keep your weekend.
Generative AI studios
Image, video, audio at minute-level cost
Stable Diffusion XL, Flux, Suno-style audio, video diffusion — all run on the same SSH-accessible boxes you'd rent from a hyperscaler, except you get billed per minute and refunded on early stop.
Enterprise ML teams
Committed terms without the contract
Commit 7, 30, or 90 days on RESERVED tier, priced at checkout, and skip the procurement process. Solana settlement and on-chain receipts replace the PO chase.
AI startups
From experiment to production on one bill
Spot tier for batch training, reserved for serving, on-demand for the random Tuesday spike. One rate sheet, one cost meter, one settlement chain — no four-cloud invoice reconciliation at month-end.
Web3 + DePIN builders
Crypto-native infrastructure with the receipts to prove it
Wallet sign-to-pay, USDC settlement, every operator carrying a public reputation score. Verifiable from the chain, no centralized invoice flow, fits naturally inside a token-driven product stack.
Three steps.
Pay, allocate, run.
1// You pay2POST /v1/buyer/compute/requests3{4 "gpuTier": "H100",5 "gpuCount": 1,6 "tier": "ON_DEMAND",7 "durationDays": 18}
Real machines,
real operators.
Inventory is contributed by independent node operators. Each one carries a reputation score derived from uptime, ratings, and completed jobs. You can read the formula. You can read the source.
Targets we hold
ourselves to.
Six workers,
one rate sheet.
The moving parts that keep allocations fast, prices honest, and operators accountable. Every one of them is in production today.
Auto allocator
Watches confirmed payments every ten seconds, picks an idle node by tier and reputation, mints an SSH credential before the buyer can refresh the page.
Per-minute meter
Heartbeats roll into a buyer-facing cost ticker every sixty seconds. Stop early and the unused minutes return to the wallet that paid.
Reputation engine
Daily recompute on a transparent formula. The score, tier, and the math behind both are public. There is no pay-to-rank tier.
Spot preemption
When ON_DEMAND demand spikes, SPOT rentals get a 90 second grace window and a partial refund. RESERVED tier is exempt by contract.
Solana settlement
Buyer USDC payments confirm on Solana via Helius webhooks, with an on-chain poll as fallback. Service fees for a completed rental are recorded as an earning for the operator's node on the UTC day it completed; they become withdrawable after that day ends plus a cooldown, and a withdrawal is reviewed by an admin before funds move on the rail chosen in the operator portal.
Carbon estimator
Per-rental CO2 from GPU TDP times region grid intensity. Buyer dashboard surfaces grams per active rental and monthly totals.
Trust is
non-negotiable.
No checkbox compliance copy. Just the controls we actually run, named honestly. If a claim is missing, it is because we have not earned the right to make it.
Ephemeral SSH credentials
Every rental gets a fresh session token. It dies when the rental ends. No shared passwords, no operator can replay your access.
Encryption at rest and in transit
Postgres rows encrypted at rest by the database provider. All API and dashboard traffic over TLS 1.3.
JWT auth, no shared admin tokens
Per-user signed tokens with role separation. The old HMAC admin token from Phase 1 was replaced in M1.
Published audit records
Third-party security audit reports, when we publish one, are recorded at /audits with a link and a SHA-256 hash so you can verify the artifact yourself. We vouch for the record, not the audit.
Three calls
to a running GPU.
No SDK to install. No proprietary auth scheme. Just REST over HTTPS with bearer tokens and an OpenAPI spec you can read.
Plain REST
No SDK lock-in. curl, fetch, requests, all work the same.
Open spec
OpenAPI 3 schema for every public endpoint.
Per-minute meter
Cost ticks update on a 60s cadence via WebSocket.
Ephemeral creds
Session tokens expire when the rental ends.
curl -X POST https://a2e-api.onrender.com/v1/buyer/compute/request \-H "Authorization: Bearer $TOKENOS_TOKEN" \-H "Content-Type: application/json" \-d '{"gpuTier": "H100","gpuCount": 1,"tier": "ON_DEMAND","durationDays": 1,"sshPubKey": "ssh-ed25519 AAAA... you@laptop"}'
“Pay only for the minutes you used. Refund the rest.”
Operating principle
TokenOS DeAI Network, Per-minute meter
/min billing precision
Three tiers,
one rate sheet
Per-minute metering on every tier: on-demand and spot settle to the minute when you stop early; reserved bills its full commitment up front. The published rate is the on-demand lane, shown for an H100; SPOT and RESERVED are priced at checkout, and the marketplace surfaces live per-operator pricing for every GPU class.
On-Demand
Pay only for the minutes you use. Cancel anytime.
~$5.5375/hr at retail
- Per-minute settlement; 1-day minimum (1-hour on hourly rentals)
- Refund unused minutes on early stop
- Reputation-scored operators
- Ephemeral SSH credentials per session
- Live cost meter in your dashboard
Spot
Preemptible with 90 seconds notice. Priced at checkout.
not part of the published schedule
- Discounted below on-demand; exact rate shown at checkout
- 90 second eviction warning before preemption
- Refund prorated for unused minutes
- Optional checkpoint snapshot for restart
- Best fit for batch and training workloads
Reserved
Commit 7, 30, or 90 days. Priced at checkout.
not part of the published schedule
- Commitment pricing shown at checkout
- Choose 7, 30, or 90 day terms
- Billed up front; non-refundable on early termination
- A paid claim on operator uptime; no capacity guarantee is sold
- Best fit for long-running production
USDC payments confirm on Solana. See live rates for every GPU tier
What you would
pay elsewhere.
On-demand list prices from AWS, Azure, and GCP for the same hardware (us-east-1, East US and us-central1), recorded in May and June 2026 and not refreshed since. Reservation and committed-use discounts can lower their numbers 30 to 60 percent; TokenOS does not require any commitment to get the figure shown.
1 × NVIDIA H100 for 720 hours per month. TokenOS rate of $5.5375 / GPU-hr is the published on-demand price; no reservations or commitments required.
Pick a GPU,
pay by the minute.
Sixty seconds from payment to an SSH prompt. Refunds for unused minutes. No tickets.
Pay with USDC on Solana or card. No subscription; rentals start at one hour.