FinBlocks Tech
Markets · Capacity & access

Data-centre & compute capacity

An idle accelerator-hour is gone at the end of the hour, and a rack of unsold colocation power is gone at the end of the month. Both trade today through bilateral contracts, waitlists and a growing layer of brokers, at prices that differ several-fold on the same day for the same class of hardware depending entirely on who you know.

Instrument
Class × region × delivery window
Session pattern
Frequent calls, then continuous
Hard part
Delivery, not matching

01 · How this market works today

The same GPU-hour, at four different prices, on the same afternoon.

This market has enormous demand, genuine scarcity and no price. What exists instead is a set of private arrangements between individual operators and individual customers, plus single-vendor spot mechanisms where the seller is also the only counterparty and sets the rules.

Single-vendor spot is an auction with one seller

Where spot pricing exists at all it is run by the operator selling the capacity, who is simultaneously the price-setter, the sole counterparty and the party deciding when your workload is interrupted. That is a useful product. It is not price discovery, and it produces no reference anyone else can use.

Prices are dispersed several-fold with no visible reason

The same class of accelerator, in the same region, in the same week, transacts at wildly different rates depending on contract vintage, relationship and negotiating position. Nobody in the market can tell whether their price is competitive, so nobody can plan around it.

Buyers hoard, which makes scarcity worse

With no dependable route to short-term capacity, every serious buyer over-reserves and holds a safety margin idle. That behaviour is individually rational and collectively expensive: a large fraction of the world's most contested hardware sits reserved and unused because releasing it is not safely reversible.

Colocation power is a long contract with a private price

Underneath the compute, the actual constraint is a kilowatt in a building near a substation, sold in multi-year contracts at confidential rates. Secondary transfer is a contractual negotiation, so a tenant with unused capacity and a tenant who needs it rarely find each other.


02 · What an auction changes

Price discovery, in public, with the volume behind it.

Every structural condition for a real market is present — a fungible-enough unit, extreme perishability, many buyers and a growing number of sellers. What is missing is a neutral venue where the seller is not also the referee.

A neutral venue is structurally different from a vendor's spot market

When many independent sellers meet many buyers on one book, the clearing price is discovered between them rather than announced by one of them. That is the whole difference, and it is the reason a single operator cannot produce this outcome no matter how sophisticated their pricing.

A standard unit, defined tightly enough to be fungible

Accelerator class, interconnect, host memory, region and delivery window together define a book. Buyers care about those differences, so they belong in the instrument definition rather than in a footnote — which means more instruments and a market whose prices actually mean something.

A short clock for prompt capacity, a curve for forward

Frequent call auctions clear near-term delivery windows on a schedule matched to how fast the capacity perishes, while scheduled auctions further out build a forward curve. IOC and FOK validity let a buyer take prompt capacity or not take it, without leaving a resting order they cannot use.

A published price lets buyers stop hoarding

The reason to hold idle capacity is that reacquiring it is uncertain. A liquid book with a visible price makes release reversible, which frees the reserved-and-unused overhang — the single largest pool of supply in this market and the cheapest to unlock.

Pre-trade limits and netting make many-to-many workable

A market with many small sellers and large buyers needs credit control before execution and netted settlement afterwards, or the operational cost of a thousand small trades eats the benefit. Both are configuration rather than a build.

Auction price discovery and the Indicative Equilibrium PriceA chart of cumulative demand and cumulative supply against limit price across eight ticks from 148.10 to 148.45. Cumulative demand falls from 14,400 to 400 as the price rises; cumulative supply rises from 500 to 11,700. Bars behind the curves show the executable volume at each price, which is the lesser of the two. The tallest bar, and therefore the Indicative Equilibrium Price, is 6,300 units at 148.25, leaving a surplus of 900 on the demand side.03,0006,0009,00012,00015,000IEP 148.256,300 executable900 surplus on the demand side148.10148.15148.20148.25148.30148.35148.40148.45Cumulative quantityLimit price · tick 0.05Cumulative demandCumulative supply
Cumulative demandCumulative supplyExecutable volume at that price
Figure 1How an auction finds its price. Cumulative demand falls as the price rises and cumulative supply rises with it. The engine examines every price at which bids and offers overlap and uncrosses at the one where the greatest volume can be matched — the same mechanism whichever market the book happens to hold.

03 · What Mehrex supplies

The parts that are configuration rather than a project.

Frequent sessions, an API-first channel for a participant set that is entirely technical, and the risk and clearing that let strangers trade with each other at volume.

A first configuration — settled in the market model workshop, not fixed by the platform
SettingTypical starting point
InstrumentAccelerator class × interconnect × region × delivery window
Lot sizeOne device-hour, or one kW-month for colocation power
SessionsFrequent call auctions for prompt; scheduled auctions for forward terms
ValidityIOC and FOK for prompt capacity; GTD across a forward curve
Price bandA tolerance around the last uncrossing, halting rather than printing through
RiskBuying power per participant; prepayment for unrated counterparties
SettlementShort netting cycle; provisioning instructed outside the venue

The modules that carry it

01 · core

Matching Engine

Call auctions on a short published clock for prompt delivery windows, continuous trading where depth supports it, and IOC and FOK validity for buyers who need capacity now or not at all.

Read more →
06 · channel

Connectivity & Market Data

The primary channel here: HTTP/JSON with server-streamed events and snapshot-then-stream market data, so a scheduler can bid programmatically against the live book rather than through a screen.

Read more →
04 · core

Risk Management

Buying power and notional limits per participant evaluated before the order reaches the book — the control that lets an operator sell to a counterparty it has never met.

Read more →
05 · post-trade

Clearing

Short netting cycles per participant, fees per trade, and enriched trade-capture records that tell the provisioning layer what was actually sold and to whom.

Read more →

04 · What you still have to build

The honest half of the estimate.

Be blunt about the ratio here. The matching engine is perhaps a tenth of this market. Delivery is the other nine tenths, and none of it is something a trading system can do for you.

Provisioning — a matched trade does not start your job

The venue produces an entitlement to capacity. Turning that into a running workload means scheduling, tenant isolation, networking, storage, images, identity and quotas, across operators whose platforms have nothing in common. That orchestration layer is the product; the exchange is a component inside it.

Verification of what was actually delivered

Was the hardware the advertised class? Did it stay up? What happens when a node fails four hours into a twelve-hour run, and who pays? Without measurement and a remedy the buyer is trusting a stranger, and the market clears at the price of that mistrust.

Making the unit genuinely fungible

Two nominally identical accelerators can differ in interconnect topology, host memory, thermal behaviour and driver stack, and those differences change job throughput materially. Defining a unit tight enough that buyers are indifferent — and enforcing conformance to it — is standards work with a testing regime behind it.

Security, tenancy and data residency

A large share of the demand will not run on unfamiliar hardware at any price, because of confidential data, model weights, regulatory residency or customer contract terms. Attestation, isolation guarantees and jurisdiction constraints are eligibility rules you have to define, and they cap the addressable market.

Credit, prepayment and settlement for many small sellers

A market with a long tail of sellers and concentrated buyers needs prepayment, escrow or a clearing arrangement, plus a policy for a seller that takes payment and does not deliver. Pre-trade limits control exposure; they do not underwrite it.

Contract, service levels and remedy

Availability commitments, interruption rights, notice periods and what a buyer is owed when capacity is withdrawn mid-run. This is the rulebook, and in a market where the good is consumed as it is delivered, it is more load-bearing than in almost any other market on this site.


Questions

What evaluators ask first.

Is this the same as a cloud provider's spot market?

No, and the difference is structural. A vendor spot market has one seller who sets the price, decides the interruption policy and is the only counterparty. A venue has many independent sellers meeting many buyers on one book, so the clearing price is discovered rather than announced — which is also why no single operator can produce this outcome alone.

How is capacity actually delivered after a trade matches?

Outside the engine. Mehrex matches, applies pre-trade risk, nets the obligation and emits an enriched trade-capture record describing what was sold. Provisioning that capacity — scheduling, isolation, networking, identity — is an orchestration platform the venue operator builds, and it is the larger part of the programme.

Can compute capacity be traded forward as well as prompt?

Yes. Forward delivery windows are separate instruments with their own books and their own scheduled auctions, so a buyer can secure capacity for a future quarter while prompt windows clear on a short clock. The engine treats both identically; only the instrument definition and session schedule differ.


Other markets
Next step

Tell us about your data-centre & compute capacity market. We will show you the engine matching.

A live demonstration runs about an hour: the order book, the Control Panel, an auction uncrossing, and a replay of the journal that produced it.