AI and Compute
A GPU reservation can bill every hour whether the cluster runs or sits idle. We find the capacity your workload needs, in the commitment it can carry, and settle the terms before you sign.
Book a conversationWhere AI Compute Comes From, and When to Buy It
GPU capacity comes in three forms. A major cloud rents it on demand, as a reservation held in one zone, as spare capacity it can take back, or as a short block booked for a future date and paid up front. Specialist GPU clouds also sell on demand, and at one, most capacity goes through multi-year take-or-pay contracts, typically with a prepayment before access. Or you own the GPUs and place them in a data center with the power and cooling to run them.
When you need it
- A training run or fine-tuning job needs GPUs for days or weeks, starting on a date you can plan around.
- You own or are buying GPU systems and have no space with the power and cooling to run them.
- A provider has offered a multi-year capacity contract with a prepayment, and you need to know what it binds you to.
- The next generation of hardware you plan to buy needs more power and cooling per rack than your data center supplies.
When you don't
In our judgment, hold off on buying compute when:
- A model API, a cloud AI service or an AI feature in software you already own meets the need. That choice starts with AI readiness and selection.
- You are running a pilot, or demand you can't yet forecast. A multi-year take-or-pay commitment is the wrong first purchase. Short blocks, flexible reservations or on-demand capacity keep your options open until the workload is proven.
- Your planned systems are air-cooled and fit the density your facility already supports. A liquid cooling retrofit can wait, as long as the generation after them will still fit.
If you don't need to buy compute, we say it before anyone asks you to commit.
What Decides Whether the Capacity Is There When You Need It
The GPUs are the easy line on the quote. Power, cooling and the networks around them decide when work can start.
Capacity and price are separate purchases
At one major cloud, a capacity reservation holds GPUs in a zone but carries no discount, and the discount plans hold no capacity. Buy the one you need, or both, on purpose.
Short blocks end on time, whether your job has or not
At one major cloud, a dated GPU block is priced on supply and demand when you book it, and paid up front. It ends at a fixed time, and the provider starts shutting instances down before the end. Training needs checkpoints that fit the block.
Spot capacity comes with little warning
Spot GPUs carry no service level and no promise of availability. Providers reclaim them with seconds or minutes of notice, and sometimes none. They suit only work that can stop and restart.
Power sets the start date
One analyst's 2026 market report finds power constraints extending data center build timelines. New generation takes years to connect to the grid. At least one utility writes commitments into power itself: one data center tariff carries a minimum bill on contracted capacity, a long term, an exit fee and collateral.
Dense AI racks need liquid cooling, or a plan for it
One chip maker's flagship rack-scale system is liquid cooled, while its 8-GPU system of the same generation is air-cooled. How many of those fit in a rack depends on the power and cooling the rack gets. Rear-door heat exchangers, which cool exhaust air with water at the back of the rack, are one way to raise what an air-cooled room can handle.
An AI cluster runs on more than one network
Reference designs keep the GPU-to-GPU fabric apart from storage, management and the front-end network that carries data in and checkpoints out. A quote that prices only the GPUs and their fabric leaves the rest for later.
InfiniBand or Ethernet
Both run AI training at very large scale. InfiniBand is built as a lossless fabric. One large operator chose Ethernet with RDMA for most of its AI capacity, found some tuning more delicate on it and had to rework standard load balancing for AI traffic. An Ethernet specification for AI and HPC reached version 1.0 in 2025, with compliance programs announced as in progress at release.
One chip maker, one supply chain
One specialist GPU cloud gets all its GPUs from a single chip maker and lists supply delays among its risks. It has also disclosed delays in delivering certain data centers. Ask where the capacity you are buying stands today, not only when it is promised.
The Commitment, Sized to the Workload
For IT and infrastructure leaders: capacity on the date you planned
- Delivery dates with remedies, and acceptance tests before any tranche is billed.
- Power, cooling and the networks around the GPUs priced before the quote is signed, not left for later.
- Your infrastructure team plans the cluster while TREU competes the capacity offers.
For finance: a commitment you can carry
- Take-or-pay sized to the demand you can show.
- Prepayment weighed against a liability cap that may be a fraction of the fees.
- Substitution and refresh rights settled before the next GPU generation ships.
What to Watch For in a Compute Contract
Two GPU capacity agreements filed publicly with the SEC show what can be negotiated. Your provider's paper may start somewhere else.
-
Take-or-pay.
In one published GPU capacity agreement, the buyer pays for each accepted tranche whatever it actually uses. At one specialist cloud, reserved GPUs bill every hour, running or idle. Commit to the demand you can show.
-
Delivery dates with remedies.
Ask for notice of expected delays, damages for late delivery and the right to end a late tranche and recover the upfront payment. One published agreement carries all three.
-
Acceptance before the term starts.
The service term should start when you accept the capacity. Write the acceptance tests down, decide what happens if only part of the cluster passes, and watch for acceptance that is deemed if you don't respond in time.
-
Financing conditions.
One published agreement creates no obligations until the provider has its financing, and delays caused by that condition carry no damages. Your delivery date can depend on someone else's lender.
-
Prepayment and the liability cap.
Committed GPU contracts can require a prepayment before access. In one published agreement, the provider's liability is capped at a fraction of the fees, so what you owe can far exceed what you could recover. The provider's financial strength is part of the delivery risk.
-
Credits, not damages.
One major cloud's compute service level makes credits the sole remedy for unavailability, and one GPU cloud's master agreement excludes damages for service interruptions.
-
No early exit.
Neither published agreement gives the buyer a right to leave for convenience. At one major cloud, short dated blocks can't be canceled, moved or split. Under one master agreement, if the provider ends it for your uncured breach, the fees for the rest of the term come due.
-
Hardware generation and substitution.
One published agreement fixes the GPU type for each tranche, with no clause letting the provider substitute models or move to a newer generation. Another lets the provider change services with reasonable notice. Settle in the contract whether substitution is allowed, against what test, and whether a refresh is offered mid-term.
-
Your data and model weights at the end.
Under one master agreement, the provider deletes customer data after a recovery period, and exporting it before the term ends is the customer's job. Plan the export of checkpoints and model weights before expiry.
-
AI-density colocation terms.
Define the critical power allocated to you, and the limits on raising density later. The rest of the colocation agreement is covered on our cloud and data center page.
-
First offer on future capacity.
A right of first offer on future capacity at the site, with price, delivery dates and term disclosed before anyone else is offered it, can secure expansion before you need it.
This list is a starting point for your review. TREU is not a law firm. Your counsel gives the legal advice.
How TREU Sources AI Compute
See how we work-
Supplier identification
We match the workload to the form of capacity it needs, then to the clouds, specialist providers and AI-density data centers that can deliver it, favoring the ones that can grow with you.
-
Buying cycle
We put every offer on the same basis and negotiate the commercial terms with your counsel: commitment, prepayment, delivery dates, acceptance, liability and exit.
-
Implementation oversight and accountability
Each tranche is tested before its term starts. We hold the provider to the delivery dates through installation and acceptance, then check the first invoice against the contract. GPU utilization and monthly bills sit with the providers we source to track them.
-
Client and supplier lifecycle management
After acceptance, we keep your terms and renewal dates on file, sit in on supplier business reviews, renegotiate at renewal and source the next generation of capacity, for the life of every agreement.
Questions Buyers Ask Before They Commit to GPU Capacity
What is the difference between on-demand, reserved and spot GPU capacity?
On-demand is pay as you go, with no promise the GPUs will be there at peak. A reservation holds capacity for you, and at some providers bills whether you use it or not. Spot is spare capacity the provider can take back with little warning, so it suits only work that can restart.
What does take-or-pay mean, and can we avoid it?
You pay for the capacity you committed to, whatever you use. Short dated blocks, flexible reservations and on-demand capacity avoid a long commitment. In our judgment, that is where a pilot belongs.
What happens if the provider delivers our GPUs late?
It depends on what you negotiated: notice of delays, damages for late delivery and a right to end a late tranche. Without them, your remedies are whatever the provider's paper offers. Check whether its obligations depend on its financing.
Do we need liquid cooling?
For the flagship rack-scale system from one leading chip maker, yes. Some smaller GPU systems are still air-cooled. The real question is whether the hardware you will buy next fits the power and cooling your facility can supply.
InfiniBand or Ethernet for our AI cluster?
Both run AI training at very large scale. InfiniBand is lossless by design, and Ethernet takes careful tuning under AI traffic. Price the storage and management networks too, not only the GPU fabric.
What if a newer GPU generation ships during our term?
In one published agreement, the GPU type is fixed for each tranche. Decide at signature whether substitution is allowed, against what test, and whether a refresh to a newer generation is offered mid-term.
Who keeps our hardware and data if the provider gets into trouble?
Protect both in the contract. For hardware in colocation, see our cloud and data center page. In a capacity contract, plan the export of data, checkpoints and model weights before the term ends, because the provider may delete them after a recovery period.
Will TREU watch our GPU utilization and bills?
No. We check the first invoice against the contract, then track terms and renewal dates and join business reviews. Our technology savings and cost optimization work covers utilization and AI spend, and we source the providers who track usage and invoices.