Solutions · two units, one balance
An AI credit buys a unit of finished work — a million tokens, a video second, a scanned page. A GPU credit buys an hour of a named accelerator. Both settle against the same balance, both are transferable, and both clear T+0. Most customers hold some of each, because the right answer changes by workload and by quarter.
AI credits
The mistake most compute contracts make is pricing everything in GPU-hours when the buyer thinks in pages, minutes or agent-hours. Each modality below bills in the unit its buyer already plans in, and settles in dollars underneath.
Chat, extraction, summarisation, classification, long-horizon reasoning.
Completion, review, migration, agentic repair against a repository.
Generation, editing, upscaling, background and product composition.
Text-to-video, image-to-video, interpolation, restoration.
Transcription, diarisation, translation, synthesis, voice conversion.
Embeddings, vector and hybrid retrieval, reranking, grounded answers.
Layout parsing, table extraction, handwriting, signature and stamp detection.
Scoring, imputation, anomaly detection, demand and risk forecasting.
Continuous inference on live camera and sensor feeds at the edge or in region.
Long-running tool use, planning, and multi-step execution against real systems.
Mesh and scene generation, physics simulation, synthetic data for robotics.
Learned environment models for planning, policy training and evaluation.
NoteRates for each modality live on the pricing page. For calibration on the text line, published frontier-model list prices currently run from $1 / $5 per million input / output tokens at the small end4 to $3 / $15 for mid-tier frontier models5. Exascale prices against the market, not against a single vendor.
GPU credits
A GPU credit is an hour of a specific accelerator in a specific region, not a vague claim on “capacity”. That specificity is what makes it tradable: two H100-hours in the same region are fungible, and an H100-hour and a GB300-hour are not.
| Class | Memory | Bandwidth | What it is for | Notes |
|---|---|---|---|---|
| H100 SXM | 80 GB HBM3 | 3.35 TB/s | The market's reference unit. Fine-tuning, mid-size training, dense inference. | The most liquid instrument on the venue — and the one whose list price varies most between providers, from roughly $1.49 to $6.98 per GPU-hour for identical silicon6. |
| H200 SXM | 141 GB HBM3e | 4.8 TB/s | Long-context inference and memory-bound serving where H100 spills. | NVIDIA describes it as the first GPU to offer 141 GB of HBM3e at 4.8 TB/s1. The memory step, not the FLOPs, is what moves serving economics. |
| B200 | 192 GB HBM3e | 8 TB/s | Blackwell-generation training and high-throughput inference. | The generational step from Hopper. 192 GB per GPU3 changes what fits on a single device and therefore how many devices a run needs. |
| B300 (Blackwell Ultra) | 288 GB HBM3e | 8 TB/s | Reasoning-model inference where FP4 throughput and memory both bind. | 288 GB via 12-high stacks — 50% more than B200 — with roughly 1.5× B200’s dense FP4 throughput3. Confirm per-GPU figures against an NVIDIA datasheet before contracting on them. |
| GB200 NVL72 | Rack-scale, 72 GPU | NVLink domain | Frontier training runs that need one coherent memory domain, not a network of them. | Sold as a rack, priced as a rack, and scheduled as a rack. Fragmenting an NVL72 destroys the thing you are paying for. |
| GB300 NVL72 | 20.7 TB HBM3e / rack | 130 TB/s NVLink | The current ceiling. Frontier training and the largest reasoning deployments. | 72 Blackwell Ultra GPUs and 36 Grace CPUs in one liquid-cooled rack-scale domain2. Availability, not price, is the binding constraint through 2026. |
Memory and bandwidth figures are vendor specifications, cited where quoted. Regional availability and contracted rates are on the pricing page.
Why a credit and not a contract
The problem with the current market is not price, it is shape. Capacity is sold in year-long blocks to buyers whose demand moves weekly, which is why so much reserved capacity sits idle while spot buyers queue.
A credit you no longer need can be sold back to the book at the clearing price. A reservation you no longer need is a sunk cost you explain to your CFO.
One venue, one price, visible to both sides. Today the same H100-hour trades across roughly a 4.7× range depending on who you ask6, which is what a market without a reference price looks like.
Forwards let a lab fix the cost of a training run before committing to it, and let an operator fix revenue before energising a hall. Both sides of that trade exist today and have no venue to meet on.
Data-center electricity demand is projected to more than double by 2030, to around 945 TWh7. Capacity that follows power is worth more than capacity that sits where power is scarce — a market prices that, a fixed contract cannot.
AI
Labs, enterprises and product teams consuming compute. What each one needs is not more capacity — it is a rate they can put in a budget and defend twelve months later.
Inference is a cost of goods sold, not a research expense. It scales with revenue, it recurs every month, and it is quoted to you at a rate that can move before your own pricing can. A team that has committed to a price per seat carries the gap.
Buying inference capacity through a book rather than a contract makes the rate an observable number with a settlement date attached. The position can be sized to the forecast and adjusted when the forecast changes, without renegotiating anything.
A training run needs contiguous capacity for a known length of time, and it needs it on a date. That is a forward, not a subscription: the quantity, the delivery window and the price are all known at the point of trade.
Runs slip. When one does, the position can be sold rather than absorbed — which is the difference between a reservation and a contract, and the reason a market exists for this at all.
One endpoint across frontier and open-weight models, billed per token from the same balance that holds GPU-hours. Models are addressed by name and version, so a change in routing cannot silently change which weights answered.
Accelerators by the hour, with no minimum term and no negotiation. The rate is the venue's rate — the same number the book is quoting — rather than a list price with a discount schedule behind it.
Compute
Operators with fleet to place and utilization to defend. Idle capacity is the only inventory that is worth nothing the moment it expires.
An operator with unsold hours has two options: build a sales function, or accept whatever a single offtaker will pay. The first is expensive and slow; the second is a price with no reference to check it against.
Listing into a book is neither. Capacity is offered at a price the operator sets, against demand it did not have to find, and the fill is cleared and settled without a contract negotiation per counterparty.
Settlement has to be computed from something. Telemetry is that something: per-node health, power draw and utilization, recorded continuously and retained as the record both sides settle against.
It is also the operator’s own instrument panel. The same series that proves an hour was delivered shows which racks are earning and which are drawing power for nothing.
Exchange
The side that makes the price. A venue with no professional liquidity is a price list that updates slowly.
A central limit order book in GPU-hours: price-time priority, visible depth, and the same rules for every participant. Buy and sell are symmetric — there is no side of this market that is a customer and no side that is a vendor.
The same accelerator does not clear at the same rate in every region, at every tenor, or in every configuration. Those spreads are the market’s own pricing error, and closing them is what makes the reference rate worth referencing.
Two exposures, opposite in sign. A buyer of compute is short the rate and wants it fixed; an operator is long the rate and wants a floor under next year’s revenue. Both are hedges, and the venue exists so they can face each other.
Futures here are cash-settled and perpetual: there is no physical delivery leg, so a hedge never obliges anyone to take hours they cannot use.
Two-sided quotes, in size, under an obligation — and a rebate schedule that is published rather than negotiated privately. A venue that pays for liquidity in secret has a different price for different members, which is the one thing an exchange cannot have.
Every figure above is numbered to an entry here. Links last read 27 July 2026.
The NVIDIA H200 is the first GPU to offer 141 gigabytes (GB) of HBM3e memory at 4.8 terabytes per second (TB/s).
ParaphraseGB300 NVL72 connects 72 Blackwell Ultra GPUs and 36 Grace CPUs in a rack-scale, liquid-cooled design with 20.7 TB of HBM3e and 130 TB/s of NVLink bandwidth.
NVIDIA Blackwell Ultra B300: Full Specs, 288GB HBM3e Memory, 15 PFLOPS FP4
ParaphraseB300 features 288 GB of HBM3e per GPU — 50% more than B200's 192 GB — via 12-high stacks, and delivers roughly 1.5× the dense FP4 throughput of B200.
Third-party compilation of vendor material. Per-GPU B300 figures should be confirmed against an NVIDIA datasheet before being used commercially.
Pricing — Claude Platform Docs
Claude Haiku 4.5 is priced at $1 per million input tokens and $5 per million output tokens.
ParaphraseAvailable at an introductory price of $2 per million input tokens and $10 per million output tokens through August 31, 2026, moving to $3 / $15 thereafter.
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.
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.
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.