Pricing · published in full

We publish the fee. The price is whatever the book says.

An exchange does not have a price list for the thing it trades — it has a fee for trading it. What we can publish, and do below, is exactly what we charge, exactly what the rest of the market charges for the same silicon, and where every one of those numbers came from.

0–180 bpsexchange fee

0provisioning fee

0egress fee

T+0settlement

What the market charges

Published on-demand H100 list prices, July 2026

This is the benchmark we price against, and the clearest single argument for the venue. The same fungible GPU-hour carries roughly a 4.7× spread depending on which published list you read1.

Provider class	H100 / GPU-hour	Shape	What that rate includes
Peer-to-peer marketplaces	$1.49 – $1.60	Marketplace	Interruptible, mixed provenance, no InfiniBand guarantee
Specialized neoclouds	$1.50 – $2.50	On-demand	The competitive center of the market
RunPod (community / secure)	$1.99 – $2.89	On-demand	Community tier cheaper; secure tier carries the SLA
Lambda	$3.99	On-demand	Published single-GPU on-demand rate
CoreWeave (8×H100 node)	≈ $6.16	Per-node	Normalized from a node rate; InfiniBand fabric included
General-purpose hyperscalers	up to $6.98	On-demand	Top of the surveyed range

Other companies’ published on-demand list prices, as surveyed in sources 1 and 2. These are not quotes we have received. Committed, reserved and enterprise-negotiated rates are materially lower and are not public for any provider on this table, including us.

NoteSpot capacity has traded near $1.03 per H100-hour in this period2. A spot floor below every published on-demand rate is exactly the dislocation a book exists to close.

Our fee schedule

What Exascale charges

Basis points on notional, charged on execution. No provisioning fee, no egress fee, no minimum commitment.

Tier	Spot	Forwards	Applies to
Taker	180 bps	225 bps	Crossing the spread on the book
Maker	45 bps	68 bps	Resting liquidity that gets filled
Maker (top 10 by volume)	0 bps	23 bps	Rebated at the top tier
Physical settlement	—	90 bps	On delivery of capacity rather than cash

Exascale's own schedule. Effective 27 July 2026 and subject to change with notice to account holders.

What is not charged

No fee to provision, to move data out, to hold credits, or to cancel a resting order. Storage and cross-region transfer are passed through at operator cost with the operator’s invoice attached.

How sellers are paid

Per-second metering against the same ledger the buyer is billed from, netted daily, paid T+0 out of segregated escrow. The seller sees the buyer’s meter and the buyer sees the seller’s attestation.

GPU credits

What a GPU credit costs

A GPU credit is priced by the book, continuously, per SKU and per region. There is no rate card for it — that is the point. What follows is how to read the instrument.

Instrument	Priced by	Typical use	Reference
H100 · region	Continuous book	The reference contract. Deepest liquidity.	Benchmarked against the table above1
H200 · region	Continuous book	Memory-bound serving; long context.	141 GB HBM3e at 4.8 TB/s5
B200 / B300 · region	Continuous book	Blackwell training and reasoning inference.	B300 carries 288 GB, 50% above B2006
GB200 / GB300 NVL72	Allocated, rack-scale	Frontier runs needing one coherent domain.	Sold as a rack; scheduled as a rack

AI credits

What a unit of finished work costs

Each modality bills in the unit its buyer plans in. Rates track the market rather than a single vendor — for calibration, published frontier list prices currently run $1 / $5 per million tokens at the small end3 and $3 / $15 mid-tier4.

Modality	Billed in	Rate	How it is metered
Text & reasoning	1M tokens in / out	market	Routed to the cheapest qualifying model that meets your latency and quality floor
Code	1M tokens	market	Repo cache billed once, not per request
Image	per image	market	By resolution tier and step count
Video	per output second	market	By resolution and frame rate
Speech	per audio minute	market	Batch and realtime priced separately
Search & retrieval	per 1K queries	market	Embedding and rerank billed as separate legs
Documents & OCR	per page	market	Structure-preserving extraction priced above flat text
Tabular	per 1M rows	market	Volume tiers from the first million
Realtime vision	per stream-hour	market	In-region placement required; edge priced separately
Agents	per agent-hour	market	Wall-clock plus tool calls, itemised
3D & simulation	per sim-hour	market	Quoted per program
World models	per rollout-hour	quote	Frontier capacity, allocated rather than listed

“Market” means the rate is set by the book at time of execution and shown before you commit — not that it is unpublished. Live rates are in the API console.

Method

How to check this page

Competitor rates

Taken from the two surveys cited, both of which list their own methodology and collection dates. We have not restated any provider’s rate from memory or from a sales conversation.

Our rates

Uncited, because they are ours rather than facts about the world. They carry an effective date and change with notice.

What we have not published

Committed and reserved pricing, for us or anyone else. No provider publishes it, and a comparison table that pretended otherwise would be worthless.

If a number is wrong

Tell us and we will correct it with the correction dated on the page. Every figure here is a link away from being checked, which is the only reason to publish it.

Rate card

What each thing costs, by unit

Three price lists, because there are three units. Accelerator time, tokens, and the fee for using the venue itself.

Capacity

Accelerator Rates

Rates run from H100 through GB300, quoted per GPU-hour and metered per second. The rate you are provisioned at is the venue’s prevailing rate at that moment, not a list price with a private discount behind it — which is why the number moves and why it can be checked against the book.

The comparison table above sets those rates against the published rates of the hyperscalers and the specialized neoclouds, so the spread is visible rather than asserted.

Quoted in: USD per GPU-hour
Metered: Per second, no minimum term
Covers: H100, H200, B200, B300, GB200 NVL, GB300 NVL
Moves: With the book — this is a market rate, not a price list

Inference

Token API Pricing

Per-million-token rates across every modality the router serves: text, reasoning, agentic, vision, OCR, image, video, speech, music, embedding, rerank, tabular, time-series, world models and code. Input and output are priced separately, because they cost differently to serve.

Rates are drawn from the same credit balance that holds capacity. A team that buys GPU-hours and a team that buys tokens are drawing on one account.

Quoted in: USD per million tokens, input and output separately
Modalities: Eighteen, across 500+ models
Billing: Drawn from the credit balance
Console: Model catalog and rates

Venue

Trading Fees

180 basis points all-in on the notional of a filled order. Resting liquidity is not charged, and the maker rebate is published rather than negotiated — a venue that pays for liquidity privately is running a different price for different members.

All-in: 180 bps on filled notional
Maker: No fee to rest; rebate per the published schedule
Taker: Charged on fill
Not charged: Listing, cancellation, market data, or settlement

Sources

Every figure above is numbered to an entry here. Links last read 27 July 2026.

1
H100 Rental Prices Compared: $1.49–$6.98/hr Across 15+ Cloud Providers
IntuitionLabs · 2026 · Third-party estimate
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.
2
GPU Cloud Pricing Comparison 2026
Spheron · 2026 · Third-party estimate
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.
3
Pricing — Claude Platform Docs
Anthropic · Accessed July 2026 · Primary
Claude Haiku 4.5 is priced at $1 per million input tokens and $5 per million output tokens.
4
Introducing Claude Sonnet 5
Anthropic · 2026 · Primary
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.
5
NVIDIA H200 Tensor Core GPU
NVIDIA · Product page · Primary
The NVIDIA H200 is the first GPU to offer 141 gigabytes (GB) of HBM3e memory at 4.8 terabytes per second (TB/s).
6
NVIDIA Blackwell Ultra B300: Full Specs, 288GB HBM3e Memory, 15 PFLOPS FP4
server-parts.eu · 2026 · Third-party estimate
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.

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.