Delivery, and how each hour is checked
A contract is delivered as compute: one attested node of its grade, in its region, for every hour of its month.
Assignment before the month
Trading in a contract ends with the last close before its delivery month. Right after that close, the program pairs every buyer that holds its position with providers’ attested nodes, in a fixed order. The lead time before the month is a placeholder still to be set.
- Netting. A short counts as attested up to the free attested capacity in its own registry entry, and beyond that it is a financial short. Financial shorts are netted at the contract’s mark against buyers that asked for cash, with no default, shared pro rata and never by draw. The exact form of that sharing is a default that can change.
- Stay. A buyer that held nodes last month may ask to stay. Capacity its provider marked for continuation goes to those buyers before anyone else. The program cannot promise a stay, since no provider has to sell the next month.
- The rest. Remaining buyers are paired with remaining attested capacity by a draw from public randomness, keeping each buyer with one provider where possible.
- Imbalance. A financial short left after the netting is in default, as a provider that did not deliver. The shorts in default and the buyers they strand are closed together at the reference price as one group, pro rata to the stranded hours, in a form that is a default and can change. The buyers are settled in cash, and where the reference price is below the contract’s mark they give the difference back.
The reference price is the market’s price for replacement compute, and every close at it in delivery and at assignment uses its average over the trailing close-out window.
The lease
At the opening hour of the month, the program records a one-month lease naming the provider’s registry entry, the buyer’s key, the nodes and the month. The program runs its own leases and uses nothing from the Akash chain’s lease modules. By default, a provider may replace a failed node with another of the same grade from its own fleet. The downtime counts as missed hours, and the buyer’s state on the failed node is lost with no cash for it.
At assignment the program publishes each pairing, and each buyer sends its new provider a short message carrying the key it will deploy with. The buyer then deploys to the public endpoint in the provider’s registry entry, in whatever form the provider’s software sets.
How each hour is checked
The check reaches the node, not the buyer’s workload. A validator sends a fresh challenge to a small agent on the node, which signs it with a key bound to the node’s attestation report and to the lease. The provider cannot fake the answer without the hardware, and the buyer cannot sabotage it.
Every hour, each lease is checked by a random panel of selected Akash validators. The hour is served if enough of the panel got a valid answer and missed if enough got none. Anything in between goes to a wider panel within the same hour, whose result stands. The panel sizes and thresholds are placeholders still to be set. Once a day a keeper posts each lease’s signed record of served and missed hours for settlement.
Paid for every hour served
For each served hour, the buyer’s account pays the provider the frozen mark times the GPUs per node, less the delivery fee of 3% at its initial value. When a contract enters delivery, its mark freezes at the trimmed average of the daily marks over its settlement window. How that average is trimmed is a default that can change, and the window’s length is a placeholder still to be set. Daily margin has already moved the difference between that mark and each side’s contract price, so both sides net their own contract price. Providers are paid daily, whether or not the buyer uses the hour.
What a missed hour costs
For a missed hour that nothing replaces, the buyer pays nothing. For an hour that replacement compute serves, the buyer pays no more than the frozen mark. On each node of the lease, the provider pays two amounts for every missed hour.
- A penalty, set from data and never less than the stress rate times the hour’s value at the frozen mark. The penalty’s exact form is a default that can change.
- The hour’s cash value. Where a replacement hour filled, this is what it cost above the frozen mark. Where nothing replaced it, it is the reference price’s excess over the frozen mark. Every missed hour carries it, the earliest included.
A short outage is therefore a few missed hours charged at their cash price, not a default, and the lease continues. Both amounts are drawn from the provider’s posting, which is decided for the cash value and a default that can change for the penalty. For a certified provider, what the posting cannot pay of the cash value is fronted in the order of loss, no further than the gap at the reference price and the covered price, and the provider must repay it within its grace period or be closed out. For an anonymous provider nothing is fronted, and the buyer bears what the provider’s margin cannot pay.
Replacement compute for the buyer
From the next hour after a miss, the program bids for replacement compute for the buyer. By default this cover is an hourly bridge, and the buyer restores from its own checkpoint on whatever node serves the hour. When cover starts is open, and the next hour is its placeholder. A cover bid never pays more than the covered price, and an hour with no fill inside that price is settled in cash at the reference price. By default the bid takes any price inside that limit and never sets a clearing price. Cover fills never move the mark or the reference price. What cover costs above the reference price is drawn from the provider’s posting before anything else, and nobody fronts it. What the posting cannot pay of it is the buyer’s own cost, and where the provider is certified the buyer holds it as a claim against the provider.
When an outage becomes a walk
A walk is an outage that the program treats as a default. When a run of consecutive missed hours reaches a set length, a placeholder still to be set, the program opens the close-out and bids, for each buyer, a strip for the rest of its month. A provider whose checks resume before the strip fills, and within one epoch (the interval between daily closes), keeps its lease but still pays what the run cost. Otherwise the walk is fixed when the strip fills or the epoch passes. It is fixed at once, with no time to resume, when a second run on the same lease reaches the set length in the same month, or when a new run begins after the lease has reached its monthly limit of missed hours. That limit is a high quantile of leases’ monthly missed hours, never below a floor tied to the run length. It stays at the floor until the market has its own record or the Akash provider uptime history it is to start from has been obtained. For a certified provider, each block of missed hours in a month as long as the set run length is reported against its certification agent.
By default, at the walk each buyer chooses replacement, which applies unless it chooses otherwise, or cash at the reference price for the rest of the month, and it may switch to cash at any hour. When the strip fills, the buyer moves to one node for the rest of the month, so the lasting move happens at most once. The defaulter’s later months are taken over through the auction, as the next chapter explains.
An example from the white paper
The white paper’s running example is an H100A position, and the paper labels its prices an illustration, not a forecast (WP §10 and Appendix A.3). In it, a certified provider is short fifty eight-GPU nodes with 600 hours left, 240,000 GPU-hours, at a frozen mark of 2.00 USD per GPU-hour. At the paper’s illustrative stress rate of 30%, its in-delivery minimum is 144,000 USD, and the covered price is 2.60. A squeeze lifts the reference price to 2.90, and the provider fails to top up and is closed out. Its posting pays its buyers 144,000 USD as cover runs, and the remaining 72,000 USD is their claim against it. The arithmetic is the same for B300A. None of these figures is a B300A price.
Changing provider
When a buyer’s next assignment names a different provider, or a replacement moves it during the month, a handover moves its state. A buyer whose workload keeps no state can skip the snapshot. Before the cutover, the outgoing provider streams a disk snapshot, encrypted to the key the buyer deploys with, to the incoming provider. At the cutover it checkpoints GPU memory and process state, so that state such as a model’s cache moves with the buyer.
The opening hours of a handed-over lease are handover hours. A missed hour there is not billed, carries no penalty, and counts toward no run and no cover. By default it also carries no cash value and counts toward no monthly limit. An outgoing provider late with its checkpoint pays the buyer a handover penalty. The window’s length and the deadline are to be set once providers have tested a full-size restore. The program does not check that a snapshot restores, so a buyer that cares keeps its own checkpoints off the node.