Embeddings
Large batches with clear inputs, outputs and deterministic accounting.
Product · In development
T2 PLANNED
Rent a GPU by the hour. Or fund a batch job.
Abakos is designed as a self-serve hourly rental marketplace, like Akash or Vast.ai, plus funded batch jobs for programmatic demand — both settle through the same escrow and can unlock the useful-work block reward.
Three ways to buy compute
Self-serve rental, programmatic jobs, or a volume deal.
ON-DEMAND
Hourly rental
Browse provider listings by GPU model, region and price. Launch an instance instantly, pay per minute from a stablecoin balance — the same model as Vast.ai or Akash's reverse-auction market.
FUNDED JOB
Batch jobs
Post a workload with a deadline and input commitment, lock stablecoin in escrow, and let an eligible provider claim it. Used by the Developer API and Abakos Chat under the hood.
OFFTAKE
Enterprise volume deal
Committed monthly volume at negotiated rates and a dedicated integration, instead of clicking through listings — for buyers who already run recurring inference or training load.
Enterprise & offtake partners
Buy in bulk without the listing UI.
Self-serve rental fits ad-hoc demand. Recurring, high-volume buyers need reserved capacity and a price that reflects commitment — the same pattern Render Network uses for its Compute Client API partners, and how Together AI buys Pearl-powered inference: a direct integration, not a shopping cart.
PLANNED
Two integration paths
Reserved capacity — a longer-term lease against the same escrow rails, at a negotiated rate, matched to specific providers or partners.
Direct API offtake — your systems call a dedicated Abakos endpoint automatically; volume settles on the same chain without a human clicking through the marketplace.
Target listings
What renting will look like.
| Provider | GPU | Region | Price / hr | Reward-eligible |
|---|---|---|---|---|
| gpu_provider_07 | RTX 4090 24GB | EU-West | target: low | Yes — supported workload |
| gpu_provider_12 | H100 80GB | US-East | target: mid | Yes — supported workload |
| gpu_provider_21 | A100 40GB | AP-South | target: mid | Pending — arbitrary workload |
Illustrative target design only. No listings, prices or providers are live. Real prices will be set by providers, not by Abakos.
Batch job lifecycle
Fund. Verify. Settle.
01
Create
Choose a registered open model, workload, input commitment and deadline.
02
Fund
Lock stablecoin in escrow. Unfunded jobs never count as useful work.
03
Assign
An eligible provider claims the job and commits hardware capacity.
04
Prove
The provider submits output, work units and the PoUW certificate.
05
Settle
Payment, protocol fee, burn and block reward become publicly traceable.
Supported workloads
Verifiable work unlocks the full reward.
Renting a raw GPU for arbitrary code is easy; proving that the cycles were "useful" is not. Abakos ships verification workload-by-workload, starting with batch inference.
Classification
Structured outputs suitable for verification and retries.
Real-time inference / arbitrary rental
Streaming and unverified rental sessions earn the 25% floor only, until proof coverage expands.
Escrow
Payment is the usefulness signal.
Abakos does not guess whether matrices look like AI. A session or job must be funded, valid and bound to proven work.
IN DEVELOPMENT
What is not live
On-chain escrow, per-minute metering, the listing matcher, model registry and settlement are specified but not implemented in production.
Have GPUs to sell instead of buy?
This page is the buy side. If you're a data center or GPU cloud operator who wants to sell capacity into the network, go to the supply-side partner track instead.
Stay close to delivery
Get the right update.
Choose your path through the page you are on. We send product and testnet updates—not token-price promotion.
By joining, you agree to the privacy notice. Unsubscribe any time.