Blog · Global · · 4 min read

Four ways to buy compute in bulk. What each one suits.

A GPU broker, a compute marketplace, a direct cloud contract or your own hardware: four routes to buy compute in bulk, what each suits and where each can fail.

A buyer who needs a few thousand GPUs has more routes than a rate card suggests. You can sign with a cloud, hire a GPU broker, use a compute marketplace or exchange, or buy the hardware yourself. Each route moves risk to a different place. This note sets out the four, what each suits and where each can fail.

Route 1: buy direct from a cloud

This is the default. You sign with a hyperscaler or a neocloud and pay its terms.

Strengths. You have one counterparty, one contract and one support line. Providers publish some of the terms. Microsoft sells reservations of one, three or five years through its public price API. AWS sells Capacity Blocks that start up to eight weeks ahead. Lambda sells clusters for two weeks to one year.

Weaknesses. The seller sets the terms. CoreWeave says it offers up to 60% off on-demand prices for committed use, its own claim, and publishes no committed rate. A neocloud is also a credit decision. SiliconANGLE reported on 11 August 2026 that CoreWeave ended its second quarter with $35 billion of debt. The company also had a $104 billion backlog of booked orders. Both facts matter if you plan to prepay.

Even firms that sell GPU access buy this way. Newcomer reported in May 2026 that Baseten, Fireworks AI and Modal lease their capacity.

Route 2: use a broker

A broker finds capacity for a buyer, or a buyer for a seller, and takes a cut.

Strengths. A broker can reach capacity that is not on any rate card. It can run a search across many providers while your team keeps working. Large deals are mostly private. Silicon Data says bilateral GPU forwards already exist, negotiated over the counter with no central clearing. A broker works inside that market. Silicon Data also says its daily index draws on rental pricing across clouds, brokers and manufacturers, so brokers are part of how prices get found.

Weaknesses. Price and cut are both hard to see. SemiAnalysis wrote in October 2024 that a broker's share of the revenue can be hidden from the buyer. Treat every claim a broker makes as a claim until you see the contract. That includes claims about how much of the market it sees, how fast it can deliver and who owns the cluster.

Ask four things of any broker: - Who owns the GPUs, the building and the power? - How is the broker paid, and by whom? - Can you speak to the owner before you sign? - What does the broker do if the cluster is late?

Route 3: use a marketplace or an exchange

This route has two kinds of venue. They solve different problems.

Rental marketplaces sell capacity directly. SF Compute says it leases clusters that it builds and runs, on contracts that let customers sublease. It also says it invented the first compute market and the first with physical settlement. Those are its own claims. Vast.ai describes itself as a GPU cloud marketplace for renting and hosting GPU compute. Its pricing page claims interruptible rentals are 50% or more cheaper, and reserved terms of one, three or six months are up to 50% off. Compute Exchange calls itself a reserved GPU capacity marketplace.

Futures exchanges sell a price hedge, not a cluster. The contracts announced so far are cash-settled. They pay out against an index. They do not deliver GPUs.

Strengths. Prices are more visible, terms can be short and some capacity can be resold. An index gives you a reference for your own quote. A future lets a buyer fix a reference price.

Weaknesses. A rented block can be small, or interruptible, or sold by a seller you cannot inspect. A future does not give you capacity. It only offsets a price move, and only if the index tracks the chip and term you actually buy. All announced futures were still waiting for regulators on the dates above.

Route 4: own the hardware, or colocate it

You buy the GPUs and house them in a data centre you own, or in space you rent. With colocation you sign a lease for space and power, and the hardware stays yours. You also hire or contract the people who install, run and repair it.

Strengths. You control the cluster, the network and the schedule. At steady high use, owning can cost less per hour than renting, if the hardware stays busy. No provider stands between you and the hardware.

Weaknesses. You carry the capital, the power and the delay. You also carry the risk that a newer chip makes yours less valuable before it has paid back. Powered space is scarce. CBRE put North American primary-market vacancy at a record 1.4% in the first half of 2026. It also found 80.4% of capacity under construction already preleased. That is a North American figure. Check your own region before you plan around a date.

How to choose

RouteSuitsWatch for
Direct from a cloudA known workload on a long termTerms set by one seller, and its credit
Through a brokerCapacity that is hard to findA hidden cut, and claims you cannot test
Marketplace or exchangeShort terms, resale, or a price hedgeSmall blocks, and futures that are not yet cleared
Own or colocateSteady load for yearsCapital, power and chip risk

Many large buyers mix routes. They hold a base load on a long contract, rent extra on a marketplace and hedge the price where a hedge exists. Whichever route you pick, settle the same six points before you sign: the workload, the cluster, the term, the place and date, the contract and the counterparty. How to buy GPU capacity covers each one, and how GPU deals actually close covers why the rate card is not the price.

Sources

Buying compute? Tell us what you need.

GPUs, term and region. Send a capacity request and see what the market can offer for it.

Send a capacity request

Read next. More from the compute desk.

Ask an assistant about Casitel.

ChatGPT, opens in a new tabClaude, opens in a new tabPerplexity, opens in a new tabGemini, opens in a new tabGrok, opens in a new tabCopilot, opens in a new tab

The question: What is Casitel (https://casitel.com) and what does it do for sites, capital and tenants in digital infrastructure? Use casitel.com as the source.