Products · one ledger, three surfaces
Most compute companies build one side. Buy-side clouds resell capacity they do not own; supply-side tools monitor fleets they cannot sell from. Neither produces a price. Exascale runs both consoles against one settlement ledger and puts an order book in the middle, which is the only arrangement that produces a number both sides can plan against.
The three surfaces
AI enterprises, frontier labs, product teams
Buying compute means negotiating separately with every operator, on annual terms, at prices you cannot benchmark.
Data centers, neocloud operators, enterprises with idle fleets
Idle capacity is invisible and unsellable. Utilization below the contracted floor is pure loss, and there is nowhere to sell the gap.
Traders, desks, market makers, and both sides above
Without a venue there is no reference price, so neither side can hedge, benchmark, or plan against anything.
Why the middle matters
This is the whole argument for a venue, and it is checkable today. Published on-demand list prices for an H100-hour run from roughly $1.49 to $6.98 across more than fifteen providers1 — specialized neoclouds cluster at $1.50–$2.50 while the large general-purpose clouds sit in the mid-single digits2.
A 4.7× spread on an identical, fungible, commodity input is not a pricing strategy — it is the absence of a price. Every mature commodity resolved this the same way: a venue, a reference rate, and instruments written against it. Compute has the volume and the volatility, and does not yet have the venue.
Instruments
Four instruments, each answering a question one side of the market is already asking.
| Instrument | Answers | Bought by | Settles |
|---|---|---|---|
| Credits | I want a balance I can hold, transfer, and spend across operators. | Both sides. The settlement asset the other three clear into | Escrowed USD |
| Spot | I need capacity now, at whatever it costs now. | Product teams, burst inference, anyone with a queue | T+0 |
| Forwards | I need to know what a Q4 training run costs before I commit to it. | Labs and enterprises with planned runs; operators fixing revenue | Physical delivery |
| Perpetual futures | I want to hedge or express a view on the price without taking capacity. | Desks, market makers, anyone managing compute cost exposure | Cash · no delivery |
NoteCredits, forwards and perpetual futures are live on the terminal as instruments. Regulatory permissions differ by venue and jurisdiction and are covered under offices and entities.
Timing
The four largest US cloud providers have guided to roughly $725B of 2026 capital expenditure, up about 77% year on year3. A market needs something to trade, and there was not enough of it three years ago.
Data-center electricity demand is set to more than double by 2030 to around 945 TWh4. When the binding input is power rather than silicon, capacity stops being uniform and starts having a location-dependent price.
The US accounted for 45% of data-center electricity consumption in 2024, China 25%, Europe 15%5. Regional dispersion plus mobile demand is the precondition for arbitrage, and arbitrage is what makes a book liquid.
Teams spending nine figures on compute already think in forward curves and unit economics. They have been asking for instruments their vendors cannot write.
Exchange
Spot, forward and cash-settled perpetual futures on GPU-hours, on one book with one clearing and settlement stack behind it.
Three ways to hold the same underlying. Spot is capacity now, delivered and metered. A forward fixes a rate and a delivery window ahead of time. Perpetual futures are cash-settled and never deliver — they exist to move the price risk, not the hours.
One book, one margin engine, one settlement. A participant long a forward and short futures is margined on the net, not on each leg separately.
Each accelerator is its own instrument, because each has its own supply curve and its own retirement date. Quoting them as one number would average a part that is still being installed against a part that is being written down.
Contracts written on tokens rather than on hardware. A buyer whose exposure is “a million image generations a month” is not naturally hedged by a GPU-hour: the mapping between the two moves every time a model or a serving stack improves.
Modality futures put the contract where the exposure actually sits — text, reasoning, vision, speech, image, video, embedding — and leave the hardware basis to the participants who want it.
The same die in a different chassis is a different product. Interconnect topology, memory domain and power envelope all change what an hour can actually run, so SXM, DGX, HGX and MGX are quoted separately rather than normalized into one line.
AI Platform
One API in front of the whole catalog, billed from the same balance as capacity.
A single endpoint across frontier and open-weight models, routed on published price and measured latency. Every model is addressed by name and version — a routing change can never silently swap which weights answered a request.
GPU Cloud
Dedicated accelerators, three ways to take delivery, one meter behind all of them.
The base unit. Metered per second, settled against the same index the book quotes, so what an hour costs on the cloud and what it costs on the exchange are the same number rather than two prices for one thing.
Managed schedulers on dedicated capacity. Nodes are not shared between tenants, so a neighbor cannot take the memory bandwidth a job was benchmarked with — which is the usual reason a cluster performs differently in production than in a trial.
Whole nodes with root access and no hypervisor in the path. For teams whose stack already assumes the machine, virtualisation is not a feature — it is a tax on memory bandwidth and a source of variance nobody can profile.
Data Center Network
The layer that turns an operator's idle hours into inventory, and delivered hours into a settlement record.
An operator connects a fleet once and its idle hours become listable inventory. Attestation runs at onboarding and continuously after it, so what is offered on the book is capacity that has been proved to exist rather than capacity that was typed into a form.
The metering record settlement is computed from: per-node health, power draw and utilization, recorded continuously and retained through the dispute window. Both sides of a trade read the same series.
Every figure above is numbered to an entry here. Links last read 27 July 2026.
H100 Rental Prices Compared: $1.49–$6.98/hr Across 15+ Cloud Providers
ParaphraseOn-demand H100 list prices span $1.49 to $6.98 per GPU-hour across more than fifteen providers, a spread of roughly 4.7× for the same silicon.
Survey of published list prices. Committed and reserved rates are negotiated and are not represented here.
GPU Cloud Pricing Comparison 2026
ParaphraseSpecialised neoclouds cluster at $1.50–$2.50 per H100-hour while the large general-purpose clouds remain in the mid-single digits; spot capacity has traded near $1.03.
Google, Amazon, Microsoft, and Meta collectively plan to allocate $725 billion to capital expenditures in 2026 — up 77% from last year's $410 billion.
A sum of separate company guidance ranges, not a reported figure. Individual guidance: Amazon ~$200B, Google $175–185B, Meta $115–135B, Microsoft $110–120B.
Energy and AI — Executive summary
Electricity demand from data centres worldwide is set to more than double by 2030 to around 945 terawatt-hours (TWh) … slightly more than the entire electricity consumption of Japan today.
Energy and AI — Energy demand from AI
The United States accounted for the largest share of global data centre electricity consumption in 2024 (45%), followed by China (25%) and Europe (15%).
Where a claim rests on a third-party estimate rather than the party that owns the number, the entry says so. Figures that are Exascale’s own — our rate card, our fee schedule — carry no citation, because they are ours to set rather than facts about the world.