Contracts ·
Ten terms that move the price more than the rate does.
The dollar-per-GPU-hour figure is the number everyone negotiates. Over a three-to-five year term it is rarely the number that decides what you actually paid.
Written for the people who sign or approve GPU cluster and GPUaaS agreements — CIOs, CFOs, heads of AI platform and the procurement and legal teams who will live with the paper for three to five years.
A pattern we see often enough to describe. The rate came down, everyone shook hands, and the cluster went live. Eighteen months later finance is asking why the workload costs materially more than the model said, and the answer is not the rate. It is a ramp that started billing before the cluster was usable, a take-or-pay clause that billed two idle months, and a hardware generation that the market has since repriced downward while the contract has not.
A GPU contract is priced on one line and paid on thirty. The rate is visible, comparable and easy to argue about, which is why buyers spend most of their leverage on it. The ten terms below are where the effective cost of a cluster or a GPUaaS agreement is actually set. Every one of them moves on a dedicated neocloud cluster. Several move on hyperscaler paper too, if you ask.
- Term length. The largest single lever, and a mechanical one. A longer commitment lowers the supplier's funding cost, which is why five years prices materially below three on identical hardware. It is not a discount; it is the cost of money passed through. Hyperscalers structure the same idea as one- and three-year commitments — Azure reservations and Google Cloud committed use discounts both work this way. Decide the term on the workload, then price it, not the other way round.
- The ramp schedule. When billing starts relative to when the GPUs are usable, and how the committed volume steps up. A ramp aligned to your workload's growth is worth real money. A ramp aligned to the supplier's cash flow is a hidden price increase, and it is usually the default.
- Take-or-pay versus committed spend. Whether an idle month is still a billed month. Flexibility is available and it has a price. Know what you are paying for it, and whether the burst or the reduction you might need is actually written in.
- The deposit. Its size, what it secures, whether it is refundable and on which triggers. On build-to-suit clusters a deposit is the entry condition — at our desk today, roughly 30 per cent at three years and 20 to 25 per cent at five — and it is a term rather than a formality. The circumstances under which it comes back to you belong in the contract, not in an email.
- Hardware generation and refresh rights. A three-year contract on today's generation spans at least one refresh cycle. NVIDIA's rack-scale platform has already moved from GB200 NVL72 to GB300 NVL72, and the next one, Vera Rubin, is not a roadmap slide — it is already ramping into production supply. A contract signed on GB200 or GB300 today will see that successor generation before it ends. A right to migrate to the next generation at a defined price, on defined notice, is worth more than a further two points of rate. Without it you are paying a fixed price for hardware whose market rate is falling.
- Network fabric and storage. InfiniBand or Ethernet, the topology, storage throughput and capacity, and the interconnect between them decide whether the cluster performs as bought. If they are not specified in the contract they have not been bought, whatever the sales deck said.
- The remedy behind the SLA. Not the percentage — the credit mechanism, the cap, what counts as an outage, and whether a termination right exists at the outer bound. A credit capped at one month's fee is not a remedy for a workload that costs more than that per day, and both parties usually know it.
- Exit, assignment and change of control. What happens if you finish early, if you are acquired, or if the supplier is. The last is the one buyers forget, and it is the one that matters most with a newer counterparty.
- Expansion and first-refusal rights. Adjacent capacity, at what price, on what notice, and for how long the option holds. Growth is the point of most of these contracts, and the paper rarely says anything about it.
- Benchmarking and price adjustment. On longer terms, a mechanism that lets the rate move with the market rather than freezing a 2026 price into 2030. Suppliers resist it; some will accept a bounded version. It is worth asking, and the answer tells you something.
Have a term sheet in front of you? Send it to sales@thechiefnegotiators.com. We mark which of the ten it settles, which it leaves open, and what the market is signing instead. One business day, no cost to the buyer, MNDA on request.
A worked example
Take a 512-GPU dedicated cluster at an illustrative four dollars per GPU-hour on a three-year term. At full utilisation that is roughly 18 million dollars a year before anything else is considered. A ramp that bills the full cluster from day one while the workload takes four months to reach it costs something on the order of a quarter of a year's spend for capacity you did not use. A generation refresh in year two, without migration rights, leaves you at the same rate while the market for that hardware has moved. And a remedy capped at one month's fee on an outage that stalls a training run for a week returns a fraction of the loss. None of those figures are the rate. All of them are the price.
Reading the paper
Not all ten carry the same weight. Term, ramp and remedy set most of the effective cost and should be settled first. Refresh rights come next, because they decide whether the price you agreed stays sensible in year three. The rest are where a good contract separates from an adequate one.
Most of these terms are not adversarial. A supplier who can deliver would rather write them down than argue about them later, and the willingness to do so is one of the better signals you will get about who you are dealing with.
The rate is what you negotiated. The terms are what you bought.
Asked about this
Which term has the biggest effect on price?
Term length. It lowers the supplier's funding cost, which is why five years prices materially below three on identical hardware. Ramp and the remedy behind the SLA are the next two.
What are hardware refresh rights?
A contractual right to migrate to the next GPU generation at a defined price and notice period during the term. They protect against paying a fixed rate for hardware whose market price is falling, and are usually worth more than a further two points of discount.
Can these be negotiated on a hyperscaler contract too?
Several can — commitment shape, ramp, term, egress treatment and capacity-assurance language move if you ask. SLA definitions, liability caps and core terms of service generally do not. All ten move on a dedicated neocloud cluster contract.
Sources
- What are Azure Reservations? — one- and three-year commitment structure.
- Google Cloud committed use discounts — commitment terms and discount structure.
- NVIDIA GB300 NVL72 and GB200 NVL72 — successive rack-scale generations.
- NVIDIA Vera Rubin platform — the generation after GB300, now ramping into production.
- The Chief Negotiators, build-to-suit availability — current term, deposit and indicative rate bands seen at our desk.
Related: GPUaaS & AI cloud contract negotiation, buy-side negotiation advisory, and how hyperscaler and neocloud contracts differ.
Get these notes in Google. Add The Chief Negotiators as a preferred source and our insights are marked as such in Top Stories, AI Overviews and AI Mode.