The secondary market for GPU compute

Compute you already paid for,
now tradable.

Sell the GPU capacity you’re not using. Get GPUs for a month, set up and verified.

Bought in years.

A GPU contract is one solid block: every node, every month, paid for up front.

Capacity sits idle.

The capacity a team isn’t using sits under the line. Already paid for, and today there’s nowhere to sell it.

Now it moves.

We move that window, node by node, to a buyer who needs it. With the data centre’s consent, set up to spec, and verified before any money moves.

Launch app →How it works

Reference 2026-11-02

H100$3.27-1.6%

H200$5.04+0.5%

B200$8.19+0.3%

B300$9.84+0.3%

$/GPU-hr · on-demand median · public sources

The problem

GPUs are sold like a building and used like electricity.

You sign for three to five years and pay for all of it, every month. You only need that much in your busiest month.

What you pay for, every month

What you actually useMicrosoft’s own clusters averaged 52% utilisation · Philly trace

1
The gap is paid for, and it perishes.
An unused GPU-month cannot be stored.
2
Smaller buyers are locked out.
A startup cannot sign three years. Same H100:
$2.70
on contract
$3.32
neocloud hourly
$12.29
Azure
3
And the gap cannot be sold.
The contract forbids it. The machines are set up for whoever holds them.

How it works

Sellers move the capacity they are not using.
Buyers pay under the hourly rate.

The data centre agrees once. The seller lists a window on a link. The buyer takes it at the posted price, delivered and verified. The contract never moves. See a request link

How one trade works25%

1Buys

A company buys three years of GPUs.

512 of them, paid up front.

The seller’s fleet · 512 GPUs, three yearspaid up front, whether it runs or not

Running

5120

Year 1Oct 2026 → Mar 2027Year 3

RunningSpare, different every month

Seller

256 GPUs · 3 months→←$1.32M in cash

Buyer

The provider signs once · the contract does not move

The seller gets

$0.00M

cash, not credits that expire

The buyer gets

3 months

with no three-year commitment

The contract moves

Nowhere

we keep $0.40/hr in between, $224K on this trade

The premium

Compute is bought in years and used in months.

The short window is the scarce thing, so it carries the price. That gap is what a holder can sell and a buyer can still save on.

What one GPU-hour costs, by how you buy it

36-month contractCivo$3.79

On demand, neocloudmedian of 21 providers$6.49

On demand, hyperscalerAWS$14.24

+71%on demand over the contract3.8×hyperscaler over the contractgetdeploying.com, 27 Sep 2026

The seller

Cash for months that would bill for nothing.

The buyer

Under the hourly rate, with no three-year commitment.

The desk

The spread on a matched trade. The buyer signs first; we never hold capacity.

Not free money. The premium pays for carrying the months and delivering the block.

Delivery

How a GPU actually moves.

A block is not a login. The machines are rebuilt for the next tenant, measured against the spec, and accepted before the seller is paid.

01Consent
0 to 2 days
The data centre agreed once. Each sub-tenant is confirmed against it.
02Screening
3 to 10 days
KYB, ownership, sanctions and export checks.
03Partition and wipe
2 to 3 days
Fabric partitioned, disks wiped, firmware pinned to the listing.
04Attestation
hours
Serials, network and storage measured against the listing.
05Acceptance
48 hours
The buyer tests in-tenancy. If it fails, it walks.
06Record
on release
Hashed and timestamped. Funds release. The trade prints.

What is attested before acceptance

IdentitySerials, firmware, driver, node count and topology, as listed
FabricAll-reduce of 380 GB/s or better across 32 nodes, one island
Storage2 GB/s per GPU, sustained
HealthThe operator's node events plus the buyer's own probes

What counts as down, and what happens

Xid classes 13, 31, 48, 79
ECC errors over threshold
NCCL bus bandwidth below the floor
Thermal throttle

1 in 16

hot spare nodes, per lot

45 min

node replacement clock

Tested on two data centres. Nothing fee-bearing has settled yet.

The exchange

GPU capacity, priced and traded
like power.

Every window that clears prints a price for one instrument. Physical first. The cash-settled market comes once the tape is long enough to price it.

not launched, maturing the benchmark right now

Under development · simulated data$/GPU-hr · 256-GPU lots · 3-month equivalent-:-:-

SL-H100-256-IB-US-3M · term mark · daily

$2.95+16.4% 1MTerm mark · 6 prints

— the assessment, one price a day┄ trend┄ published index level

How the price gets madeLeft to right. Each stage feeds the next.

RFQ links
Building
Brokers send our link instead of a WhatsApp message. The quote comes back through it.
Feed
Building
Replies normalised to one instrument, with the source on each print.
Benchmark
Next
The published term mark per SKU.
Venue
Prototype
The monthly auction where blocks change hands.
Market makers
In discussion
Two-sided firm quotes at every auction.

Advisory

A compute desk for the companies that hold compute.

If your GPU bill floats and your prices do not, you hold a position in compute. We read the contracts and tell you what to sell, what to buy short, and what to leave alone.

Inference companiesData centres and neocloudsAI labsRobotics labsEnterprises with reserved capacity

Book a compute review List a window instead

The first review is free.

01Your book

Every contract, tenor and price on one page.

02Your windows

The months you will not use, and what they clear for.

03Your exposure

What a 40% move in GPU rates does to your margin.

04Your next buy

Sign long, take a window, or wait a quarter.

Why us