Designed, not live. Nothing trades until two audits are clean, the pilot blockers found in our own risk review are fixed and checked again, counsel has answered, the B300A grade is ratified and default protection is in place and attested. The pilot will be open to certified businesses only.
Akash Compute Forwards
A design for a cleared forward market in GPU compute, delivered on attested nodes. It is not open, and nothing on this site trades.
The price of every hour of a future month is fixed today, and the month's hours are delivered one by one on attested nodes.
The problem
Superintelligence has a liquidity problem.
By that we mean an open market for a future hour of compute. You can buy one, you can sell one, and anyone can see the price.
For most of the compute being built, that market doesn't exist yet. You can reserve a few weeks ahead from a big cloud, at that cloud's price. One team runs an order book where you can buy GPU time and sell it again. That's how early this is.
Four companies alone spend hundreds of billions of dollars a year on capex, and they say most of it goes to data centers and servers.
Amazon, Microsoft, Alphabet, Meta. Company filings and results calls, calendar years, each company's own capex measure. 2026 is company guidance as of 29 Oct 2026. Total capex, not AI only. Checked 11 Nov 2026.
Stijn Van Nieuwerburgh, "Financing the AI Buildout", Brookings Papers on Economic Activity, conference draft, Sep 2026, Tables 1 and 5. A scenario, not a forecast. United States only.
Those data centers sell what they make by the hour. An hour that isn't used is gone. You can't store it, and you can't deliver it late.
Grain, metals, gas, power and freight all learned to sell future output today, in the open.
The Chicago Board of Trade opened in 1848. In March 1851 came the earliest forward contract on record: 3,000 bushels of corn, sold for delivery in June.
The London Metal Exchange opened in 1877. Its daily prompt dates still run three months ahead, the time copper took to sail from Chile.
Gas: Patrick Heather, OIES Paper NG 198, June 2025, Table 10. Power: ACER, 2025 Monitoring Report, 5 Nov 2025, Figure 2. Freight: SGX Group, 9 Jan 2026. Gas counts options as well as futures. Power counts exchanges and broker platforms. History: CFTC, US Futures Trading and Regulation Before the Creation of the CFTC. Joseph Santos, A History of Futures Trading in the United States, EH.Net Encyclopedia, 16 Mar 2008. HKEX, The London Metal Exchange, 8 Jul 2020, Table 1.
Gas, power and freight trade forward in open markets. Most GPU compute still trades on private deals.
TodayThe next twelve months
A buyer can see today's prices. For later months, most GPU compute has no market price at all.
Start month, from this month to twelve months ahead
Today a price exists, and it depends on who you ask.
You know how many GPUs you need in March. Getting a price for March means a round of sales calls. So you sign a long contract today for capacity you might not need, or you wait and pay whatever March turns out to be.
The provider.
You have machines arriving in the spring. You'd like to sell April today, at today's price, to whoever needs it. There is almost nowhere to do it, so you build and hope the buyers are there when the machines are.
The agent.
An agent can pay by the hour. An agent can't get on a sales call.
The market maker.
You quote forwards in power, gas or freight. The lab and the provider above both want to fix a price, but for most GPU compute there is no book to quote into.
The design
A forward market for the GPU-hour.
Akash Compute Forwards is a design for a cleared forward market in GPU compute. You fix the price of a future month of B300 compute today. Any attested provider delivers it. A program on Solana stands between buyer and seller.
One contract
1 node · 8 × B300 · 1 calendar month · any attested provider · priced in USD per GPU-hour
A node-month is every hour of one calendar month on one node. Its value is the price per GPU-hour × 8 GPUs × the hours in that month (744 in December).
Step 1 of 4: One daily auction
Step 1 of 4: One daily auctionStep 2 of 4: The program in the middleStep 3 of 4: Assigned before the monthStep 4 of 4: Checked every hour
One daily auction
Once a day, an open call auction sets one price per contract. Every fill trades at that price.
A program on Solana stands between every buyer and seller. Each day it values every position and moves the price change between accounts in US dollars, settled today in USDC.
On 13 November we set three tests for any fix. Here is how the design answers them.
It delivers the hour, or pays for the gap.
A buyer that holds its position into the month is assigned attested nodes for the month. Every hour is checked. A missed hour costs the buyer nothing, and the provider pays its cash value plus a penalty. From the next hour the program bids for replacement compute, or the buyer takes the rest of the month in cash.
You can see the book.
The book shows the total size at each price. The program publishes the indicative clearing price and the imbalance before the call.
It says what isn't ready.
Nothing trades yet. The status page lists what is designed, what is decided and what is still proposed.
The venue has no token of its own: dollars go in, and compute comes out.
Accounts are in US dollars and settle today in USDC. AKT is never margin. Once ACT, Akash's dollar-pegged compute token, exists on Solana with the transfer terms the design requires, a certified account may post it for part of its margin.
See it run
A simulated day in the market.
This is a simulation of the design's rules, with illustrative prices and a compressed clock. Each contract clears once a day at the call. Watch the curve move as orders arrive, and the board snap at each clear.
SIMULATION. Illustrative prices. Nothing trades.
B300A EU
JAN28Last 7.10 USD / GPU-hour SIM0.00 from the clear before
The selected contract is B300A EU JAN28. Its last clear was 7.10 USD per GPU-hour, 0.00 from the clear before. Prices are simulated.
Use the left and right arrow keys to read the price at each month.
Simulated history SIMForward curve, indicative SIMSimulated. Set one year after 17 Nov 2026.
The clearing program sits in the middle. Choose a party to see who it pays, who pays it, what it stands behind and what it delivers.
Compute
Dollars
Protection and checks
Show the links that carry
The clearing program and the ten parties around it, with what each one pays, is paid, stands behind and delivers.
The clearing program and the ten parties around it, with what each one pays, is paid, stands behind and delivers.
ProvidersSell node-months and deliver attested hours.
Pays
Its buyers, through the program, the cash value of any hour it misses, plus a penalty. Fees to the program.
Is paid by
Its buyers, through the program, for every hour it serves, less the delivery fee, every day.
Stands behind
Its own margin, which covers its side of the price change.
Delivers
Attested B300 nodes, hour by hour, for every month it sold.
Buyers, including AI agentsBuy node-months.
Pays
Its provider, through the program, for every hour served. The daily price change when the price falls. Fees to the program.
Is paid by
The program: the daily price change when the price rises, and the cash value of any hour its provider misses.
Stands behind
Its own margin.
Delivers
Nothing. It receives the compute, or the gap in cash if its provider fails.
Market makersQuote both sides in the daily auction without needing hardware.
Pays
Fees to the program, at the lower maker rate for an order that rested through an earlier close. The daily price change when it moves against them.
Is paid by
The program: the daily price change when it moves their way.
Stands behind
Their own margin.
Delivers
Nothing. A maker still short at assignment with no attested capacity is in default.
The clearing programThe counterparty to every trade. It runs the auction, the daily mark and margin, and assigns buyers to nodes.
Pays
Each day's price change to the side it favours. Providers for served hours. A share of fees to the validators who check every lease.
Is paid by
Each day's price change from the side it goes against. Buyers for served hours. Every participant's fees.
Stands behind
Every trade, as its counterparty, within limits. A certified defaulter's loss passes through the order of default protection up to a price published every day. An anonymous defaulter's shortfall falls on that settlement's winners.
Delivers
The one-month lease record that names the provider, the buyer, the nodes and the month.
Certification agentsCertify businesses for the pilot, set their allowances, and absorb a defaulting account's loss after its own margin and its guarantors.
Pays
The loss on a defaulting account it certified, from its own dollars held by the program.
Is paid by
The accounts it certifies, a daily charge for the allowance it extends, and a share of what is recovered.
Stands behind
The accounts it certifies, up to its own dollars held by the program.
Delivers
Checks on each business: legal existence, signing authority, domicile and a recoverability score.
GuarantorsStand behind a certified account, next in line after its own margin.
Pays
What it pledged, if that account defaults.
Is paid by
The account it stands behind, on terms they agree privately.
Stands behind
One certified account, up to a stated limit.
Delivers
Nothing to the market directly.
Default protectionCovers a certified participant's default after its own margin and the layers above it.
Pays
A certified defaulter's counterparties, up to a price the program publishes every day.
Is paid by
Its sources are set out in the white paper.
Stands behind
Certified accounts only. It never pays for an anonymous account.
Delivers
Payment up to a limit published in advance.
The AssociationPursues claims against a defaulter after the fact.
Pays
Nothing to the market directly.
Is paid by
Nothing. What it recovers goes to the parties owed, in the order the rules set.
Stands behind
The rulebook every certified participant signs.
Delivers
The claim against a defaulter, which the defaulter's certification agent brings to arbitration. Arbitration takes twelve to eighteen months.
Akash validatorsCheck each lease every hour in random panels, and sign the lease prices the reference price reads.
Pays
Nothing until the chain cutover. From then, each service node posts a bond of its own AKT, burned only for a provable fault.
Is paid by
The program: a share of fees, in dollars.
Stands behind
Their signatures on every check.
Delivers
An hourly check of every lease, and signed lease prices.
GovernanceA proposer suggests changes, and Akash governance can veto them during a timelock of seven days at its initial value. AKT token governance on Solana takes the proposer's role later.
Pays
Nothing.
Is paid by
Nothing.
Stands behind
Nothing. It cannot take away a credit already released.
Delivers
Rule changes, each one behind a timelock and a veto.
The Akash lease marketSupplies the lease leg of the reference price as data. It is not a counterparty.
Pays
Nothing.
Is paid by
Nothing.
Stands behind
Nothing.
Delivers
Prices of open, paying Akash leases, read and signed by validators.
AKT
What the market would do to AKT supply.
AKT comes into play in three places: surplus default protection above a buffer is spent on AKT that is bought and burned, each service node locks a bond of its own AKT from the chain cutover, and ACT posted as margin keeps AKT locked on-chain. None of the three is switched on yet. The market prices and margins in US dollars and settles today in USDC, and AKT is never margin. Pick a size of delivered compute to see the estimate.
Delivered compute a year$7M · $70M · $700M
Delivered compute a year
These are examples of size, not a forecast of this market.
Each panel has its own scale, from zero to its tallest bar, so the height of a bar in one panel cannot be compared with the height of a bar in the other. The figure under each bar is its value.
Bought and burned in a year, at most
$7M
about $0.11M Estimate. Could change heavily.
$70M
about $1.1M Estimate. Could change heavily.
$700M
about $11M Estimate. Could change heavily.
Locked in service-node bonds, at least
$7M
about $0.26M Estimate. Could change heavily.
$70M
about $2.6M Estimate. Could change heavily.
$700M
about $26M Estimate. Could change heavily.
The most the burn could spend in a year at that size: about $1.1M (about 1.8M AKT) Estimate. Could change heavily.
The least locked in service-node bonds at that size, once bonds apply: about $2.6M of AKT (about 4.3M AKT) Estimate. Could change heavily.
The most AKT bought and burned in a year, and the least AKT locked in service-node bonds, at three sizes of delivered compute a year. Every figure is an estimate that could change heavily.
In the early years the burn is likely to spend about $0.1M a year or less. Estimate. Could change heavily.
A GPU-hour is perishable: an hour that is not served in its hour is gone. A provider that buys GPU servers and a buyer that plans months of work both need to fix the price of hours that do not yet exist. The paper proposes a forward market in which a contract is one node of one GPU grade, in one region, for one calendar month. A program on Solana stands between every buyer and seller. Each day it clears an open auction, values every position at a price built from the market's own trades, and moves the day's price change between accounts in US dollars. Before each month it pairs buyers with providers' attested nodes, and validators check every lease each hour. Every protection has a limit published in advance.
Fix the price of a future month of B300 compute today, delivered by any attested provider. If plans change, sell the month in a later auction before it starts.
Quote both sides of a B300 node-month, where the natural buyers and sellers are both hedging, at one price a day. You can sell without hardware and buy back at any close before the month.