The secondary market for GPU compute

Compute you already
paid for, now tradable.

GPU contracts cannot be resold, so we never resell one. The buyer is added underneath the seller’s contract with the provider’s written consent - and dead capacity becomes a market with a published price.

Launch app →
$2.95+14.0%
SL-H100-3M · $/GPU-hr · assessed 2026-08-31 · 4 instruments
Assessed 2026-08-31
H100$2.95+14.0%
H200$4.29-5.3%
B200$6.38+13.7%
B300$7.51+3.8%
Both sides of one trade

Sellers move the months they will not use. Buyers pay under the hourly rate.

Idle months earn nothing, so any price beats holding them — and the buyer skips the three-year commitment. One trade, end to end.

How one trade works25%
1Buys
A company buys three years of GPUs.

512 of them, paid up front. That is how compute is sold.

The seller’s fleet - 512 GPUs, three yearspaid up front, whether it runs or not
Running
Running
5120
Year 1Oct 2026 → Mar 2027Year 3
RunningSpare - different every month
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
Both sides, one price

Sell the months you will not use, for cash. Buy them today, under the on-demand rate.

Seller
Holds reserved GPUs it is not using.
paid $1.49 / GPU-hr
Silicon Liquid
Buys the reserved months. Sells them on-demand.
Buyer
Needs GPUs now, not in three years.
on-demand $2.97 / GPU-hr
GPU block →
← Cash
Seller gets cash for a month that was earning nothing.Buyer pays $1.49$2.38 and saves 20–50%

NVIDIA H100 reference rates. On-demand runs about 2× a one-year committed rate (Compute Exchange, 2026). Illustrative — not a quote.

The exchange

GPU capacity, priced and traded like power.

Brokers send our RFQ link instead of a WhatsApp message. Every reply comes back normalised to one instrument, and that is where the price comes from.

We start physically settled — every trade is a block of GPUs somebody actually runs on. The cash-settled commodity exchange comes after that, once the physical tape is long enough to price it.

not launched, maturing the benchmark right now

Under development$/GPU-hr · 256-GPU lots · 3-month equivalent-:-:-
SL-H100-256-IB-US-3M · term mark · daily
$2.95
Term mark · 6 prints
+16.4%
1M change
−0.27
Basis vs Ornn settle
2.96
1M high
2.40
1M low
2.502.753.00Aug 8SepOrnn 3.07Silicon Data 2.622.95
— the assessment, one price a day┄ trend┄ published index level
Closed deals H100Final negotiated price
DateBlockStruck atvs mark
31 Aug64 nodes 6 mo · EU$2.68+0.08
31 Aug32 nodes 12 mo · APAC$2.42+0.08
30 Aug64 nodes 12 mo · EU$2.44+0.07
30 Aug32 nodes 6 mo · APAC$2.71+0.07
30 Aug128 nodes 6 mo · US$2.70+0.06
28 Aug32 nodes 3 mo · US$2.820.00
28 Aug32 nodes 1 mo · US$2.740.50
27 Aug32 nodes 6 mo · US$2.46+0.05
26 Aug64 nodes 3 mo · US$2.76+0.19
26 Aug128 nodes 12 mo · US$2.22+0.14
25 Aug256 nodes 6 mo · US$2.62+0.04
Every row is a signed deal at its final price. The last column restates it as a 3-month equivalent and compares it to that day’s assessment, so a one-month block and a year sit on the same scale. Quotes are not on this tape.
H100 term 2.95 +1.9% 3MH200 term 4.29 +7.7% 3MB200 term mark withheld · 2 printsB300 term mark withheld · 1 printOrnn OCPI settles ICE futures and every Kalshi GPU marketSilicon Data H100 + B200 rental futures on CME · announced, pending regulatory reviewNext call auction 30 Sep 2026 · 256-GPU lots · physical delivery onlyH100 one-year contract +38% Oct 2025 → Mar 2026 · SemiAnalysisUnder development · the first auction opens with the RFQ boardH100 term 2.95 +1.9% 3MH200 term 4.29 +7.7% 3MB200 term mark withheld · 2 printsB300 term mark withheld · 1 printOrnn OCPI settles ICE futures and every Kalshi GPU marketSilicon Data H100 + B200 rental futures on CME · announced, pending regulatory reviewNext call auction 30 Sep 2026 · 256-GPU lots · physical delivery onlyH100 one-year contract +38% Oct 2025 → Mar 2026 · SemiAnalysisUnder development · the first auction opens with the RFQ board
How the price gets madeLeft to right. Each stage feeds the next.
  1. 01RFQ links
    Building

    Brokers send our link instead of a WhatsApp message. The quote comes back through it.

  2. 02Feed
    Building

    One instrument, one price, provenance on every print.

  3. 03Benchmark
    Next

    The published term mark per SKU.

  4. 04Venue
    Built

    The monthly auction where blocks change hands.

  5. 05Market makers
    In discussion

    Two-sided firm quotes at every auction.

Why nobody can copy this

The software is copyable.
The access is not.

A term market needs three things at once: deal flow to price from, a venue to trade on, and a seat both sides trust.

40+
operator and broker calls in 30 days
Inside the deal flow

Brokers send us their asks. Every conversation is logged, and the price is built from those notes.

1
call from a listing venue and deep-book market makers
Doors already open

Our advisors built exchanges and chains. These are people they know, not a cold pipeline.

0
book of our own
The neutral seat

A broker runs its own book, so it can never be neutral. The same firm buys one month and sells the next.

The precedent

Electricity could not be stored either.

Power solved it in a fixed order. Compute is doing it out of order.

usedspare, and then goneAn hour nobody uses cannot be carried into the next one. True of a megawatt-hour, true of a GPU-hour.

In power the exchange came first. In compute the futures did.

A contract cannot settle against a number nobody publishes. We are building the layer underneath, in the order power did.

No storage, no cost of carry.

A forward price is a forecast, not a financing cost. It is why the compute curve can fall while spot rises, and why the oil analogy misleads.

Today it is a private negotiation with no price and no exit.

No price

Four of the eight biggest providers won't publish a term price at any size. The four that do quote the same H100 from $3.19 to $6.16 an hour.

43%apart on the same H200, same day
Silicon Data $3.10 vs Ornn $4.42 · 31 Aug 2026
No depth

A listed price isn't a price you can fill. In April 2026, half the providers asked for even 64 GPUs were completely sold out.

3-4×what a short block costs against term
Nebius Q2 2026 shareholder letter
No exit

Non-refundable, non-cancellable, non-transferable. The best-documented buy-back pays in credits that expire; its own worked example recovered 2.8%.

+38%the move you're locked out of selling into
H100 one-year contract · SemiAnalysis, 2 Apr 2026
Why us

We have run an exchange before.

A prediction market is an exchange with the hard parts left in. We built and ran India’s second-largest one.

01
Arpit Dhamija

Markets and product. Founded India's second-largest prediction-market platform. Engineering at Amazon and Adobe.

02
Aayush Agarwal

Quant and operations. Quant at one of India's largest fintechs; founded a prediction-market hedge fund. CFA Level 3.

03
Siddharth Singhal

Capital markets and distribution. An L1 builder, now at Tezos in London. Berkeley and INSEAD. Brings the relationships that take a market from tokenisation to venue listing.

Previously at
AmazonAdobeGrowwTezosStanfordUniversity of California, BerkeleyINSEAD

Nothing here is a derivative.

Physical delivery only. No cash settlement, no net-out. The analogue is electricity, not oil.

Not Yet ;)

Launch app →