Every scarce asset follows a familiar path: first it is produced, then it is priced, then it is financed. Oil found its benchmark barrel. Electricity found its forward market. Produce, price, finance.
Compute, the scarcest commodity of this century, is on that path now. Anyone can rent a GPU in minutes. Price benchmarks print daily on financial terminals. The world's largest exchanges have announced compute futures. What compute still lacks is a financial layer of its own: a native way to hold it as an asset, earn it as yield, own a claim on tomorrow's inference, or run a business or an agent that pays its own compute bill from its own revenue.
Compute Finance is the name for that missing layer. It is the financial system in which compute itself is the underlying asset. DeFi financialized capital; Compute Finance financializes compute. Equity introduced dividends. Proof of work introduced a new way to reward participation. DeFi expanded what people could do with digital assets. In Compute Finance, you hold an asset and it earns compute you can actually use.
Nobody owns decentralized finance, and nobody will own Compute Finance either. But someone has to name it.
Mining has already run this playbook
Hashpower began as something you run: machines in a warehouse, electricity in, blocks out. Then marketplaces like NiceHash turned it into something you rent, with an order book where miners sold raw hashpower and buyers set the price through open competition.
For years, that was the whole market. The venue had prices, but the industry had no benchmark. The turning point came when Luxor published hashprice, a standardized public rate for a unit of hashpower. Forward contracts followed, settling against that benchmark. Eventually a hashrate future listed on a regulated exchange. Today a miner can hedge future revenue the way a farmer hedges a harvest.
Spot first. Benchmark second. Instruments last, because an instrument cannot exist before the reference it settles against. Mining proved the sequence on a small stage. Compute is following the same script on a far larger one.
Where compute stands today
Hyperscalers sell compute reserved and on demand. Clouds run spot instances. Independent marketplaces like Vast.ai and RunPod rent GPUs by the hour, and decentralized networks like Akash, Render, and io.net do the same onchain, with a long tail of smaller markets behind them. The spot market for compute is real. It is no longer the layer that needs to be invented. What has not been built is everything a spot market is supposed to sit on top of.
The price layer is arriving now. Rental benchmarks for the most traded GPUs publish daily and carry tickers on financial terminals. The first forward curves have been published. The largest exchanges have announced cash-settled compute futures, still pending regulatory approval and not yet trading. Large buyers and sellers already sign bilateral forward deals over the counter. Piece by piece, traditional finance is beginning to treat compute as a tradable commodity.
Now look at the bottom of the stack, where compute is actually consumed, and the picture inverts. The working unit of account there is the prepaid credit, whether it sits with a hyperscaler or an inference router like OpenRouter. It is contractually non-transferable, confers no property right, and often expires quietly in an account. Renting does not solve the problem either: a rented GPU goes back at the end of the hour and never hands you the underlying unit of compute. A credit is a voucher, and a voucher is not an instrument. A voucher is spent. An instrument is held, valued, and redeemed.
The price layer has arrived. The capital layer has not.
The missing layer has a shape
Start with what Compute Finance is not. It is not chip equity, not a data center REIT, not a GPU marketplace, and not a prepaid credit. All of those exist today, and none of them lets you hold the asset itself: compute, the metered, delivered, consumed resource. That distinction matters beyond trading. A holder gains exposure to compute demand without running hardware, and a supplier gains a demand side deeper than hourly rentals.
The system rests on a small set of primitives, each with early, scattered ancestors.
Compute yield. In ordinary staking, you lock an asset and a network pays you interest. Compute yield changes the payout: you stake capital, and the return is compute. The difference sounds small and is not, because it converts compute from an expense into something a balance sheet can earn. Early fragments already exist. Venice pays stakers in inference rather than interest, and onchain credit desks turn GPU financing into yield. Nothing in the design requires the staker to be a person. An agent can hold a position that pays out in the very resource it runs on. Once compute is a cash flow, it can be valued. Once it can be valued, it can be traded.
Compute claims. The second primitive turns access into an asset: a scarce, transferable right to draw compute, priced by a market rather than a rate card. A claim like that behaves like productive property. It pays whoever holds it. It can be bought when compute is cheap and sold when demand spikes, posted as collateral, borrowed against, or passed on. This is the moment a market begins pricing compute itself rather than the companies that sell it.
Self-funding compute economies. Nearly every AI service alive today pays for inference out of someone's patience: a venture subsidy or a corporate budget. A self-funding economy reverses the flow. Part of the fees a service earns routes directly into the compute account that keeps it running. The service earns its compute instead of asking a treasury for it. For autonomous agents this is not a convenience. An agent that cannot pay for its own inference is a demo. An agent that can is an economy.
Revenue loops. Every compute entitlement is somebody's obligation. A claim on tomorrow's inference only works if someone is bound to deliver it, and delivery costs money every day the claim exists. The loop that routes real revenue back to whatever funds delivery is the unglamorous part, and any serious design has to answer the same questions. Who owes the compute? What finances the serving? What happens if that funding falls short? A system with good answers produces claims that hold their value. A system without them produces IOUs with better branding.
The frontier. Beyond these sit a settled unit of compute, cleared and liquid derivatives, and hedging as routine as it is in energy markets. That chapter is already being drafted from above by the exchanges. It will be finished from below, by the yield, the claims, and the cash flows underneath, because derivatives need something real to settle against. Where the two directions meet, compute becomes a mature asset class.
The barrel problem
The strongest objection to all of this is fungibility. A barrel of oil is a barrel of oil, but one hour of compute is not another. The physical unit of compute is the flop, a unit of mathematical work, and even the flop refuses to stand still. An H100 is not a B200. The same chip counts differently at different precisions. Real workloads lean as hard on memory and networking as on raw math. And unlike oil, silicon ages on a schedule: the flagship accelerator of one generation becomes the mid-tier workhorse of the next. If the unit will not stand still, the objection goes, nothing can settle against it.
Mining answered this objection once already. The machines securing Bitcoin were never uniform: competing ASICs across generations and efficiency curves, running on cheap and expensive power alike. Hashprice did not pretend the machines were identical. It abstracted them into the one thing buyers actually purchase, expected output per unit of hashpower per day, and let every machine price itself against that rate. The heterogeneity moved into the basis, where finance has always kept it, from crude grades to regional power markets.
Compute is starting down the same path. Early benchmarks already quote the rental rate of a reference chip rather than the virtues of any particular rack. A financial unit of compute does not need every GPU to be equal. It needs a settleable reference that heterogeneous hardware can trade around. Defining those references, and the discounts and premiums that orbit them, is some of the most valuable unclaimed work in the category. Until then, the working grammar is simple: flops are what the hardware produces, tokens are what you spend, compute is what you own.
Why now
Compute demand keeps outrunning supply and prices swing hard, and commodities with that profile have historically grown financial layers. It would be surprising if compute stayed the exception. And the layer still missing is exactly the layer crypto was built to provide: programmable claims, transparent collateral, and markets anyone can enter.
There is also a new customer at the door. Agents are the first economic actors made entirely of compute. They rent no office and hire no staff. Inference is their payroll, their rent, and their raw material at once. An economy of agents cannot run on corporate credit cards and monthly invoices. It needs compute that can be held, earned, and spent by software. For the agent economy, a financial system for compute is the precondition.
Compute is already near the midpoint of the road, roughly where hashpower stood between its first benchmark and its first exchange-listed future. The top of the stack is being built in public by the exchanges. The bottom, the part that touches the people and agents who actually consume compute, is still wide open.
What we are building
The 0G Ecosystem is building at the bottom of that stack, and today the ecosystem is shipping two key implementations.
Ascend, live today, is the 0G Ecosystem's liquid staking product and the entry point into Compute Finance. Users stake 0G and receive a0G, keeping DeFi composability while their 0G is staked. a0G is also the asset used to mint the next piece.
Infinite AI (iAI), scheduled to launch September 29 via Ascend, is a compute-focused digital asset. Users mint iAI with a0G and stake eligible iAI to receive compute credits for use across supported 0G AI products, including 0G Private Computer, with access to more than 130 AI models currently, and the 0G App, where users can chat with AI and build and launch apps. Under initial parameters, eligible staked iAI is designed to receive compute credits with a stated usage value of more than $1 per day across supported services, subject to applicable product terms.
The full loop is short: stake 0G, receive a0G, mint iAI, stake iAI, earn compute, use AI. It is a first, concrete answer to the primitives above. Ascend is the staking and liquidity layer. iAI is the compute claim. Compute credits spent across 0G's products are the revenue loop that ties the claim to real usage on a real network.
Naming a new primitive
Pieces of this have been named before: tokenized hardware in one corner, GPU-backed credit in another. None of the names held the whole. The whole deserves a name that says exactly what it is, the way decentralized finance did, and the name should belong to everyone who builds in it.
Compute Finance. The short name is not ours to pick. Decentralized finance became DeFi and traditional finance became TradFi only after enough people were saying it every day. The crowd did the shortening, and it will do it here too. Maybe ComFi. Maybe something better. The category comes first.
An asset class, not a product. The financial system of the machine economy, still early enough to name, still open enough to build.
The 0G Ecosystem is building here.
0G Ascend, Infinite AI (iAI) and related compute credits are not available in all jurisdictions and are subject to eligibility criteria and applicable product terms. Compute credits are intended solely to access supported 0G AI services, are not redeemable for cash, and do not represent interest, dividends, or any type of guaranteed return. Product features, pricing, timing and availability are subject to change. Digital assets involve significant risk. This is not investment, financial, or legal advice.
