QUESTIONS TO WATCH OUT FOR

Renting GPU Compute: What Actually Decides the Deal

Any company that moves past AI experiments hits the same wall. Running models needs powerful chips, and buying them takes months, so most teams rent.

Two kinds of sellers answer that need. The large global providers, who have everything and sell compute as part of a much bigger relationship. And a group of specialists such as Neysa, CoreWeave, Lambda and others, who focus on compute and the software around it.

The choice is usually presented as price. It rarely turns on price. Here is what it actually turns on.

1. Will the chip be there when you need it?

Renting is only cheaper if the chip is available when your project starts. Saying we have H100s is not the same as saying you will have eight H100s on 14 October.

My reading, stated as opinion: this is where the large global providers are weakest and least transparent. Their capacity is real, but it is allocated, and their biggest customers get first call on it. A mid sized buyer can discover this at the point of need rather than at the point of signing.

As of 20 August 2026, I could find no public material from the large providers committing a named chip to a named date for a standard customer.

Ask for a written commitment covering the exact chip, the named availability window and the consequence if that commitment is missed. Ask this of everyone and notice who can answer without escalating.

2. Does the saving hold up on your job?

Specialists in this market usually claim savings of between 40 and 60 percent compared with the global providers. Neysa publishes a figure in that range.

My reading: the reason a gap of that size can exist at all is that the global providers do not really sell compute at list price. They sell it behind committed spend, while also charging for taking your data back out.

The headline comparison is therefore always a comparison against a price that almost nobody actually pays.

So the advertised number is where your test starts, not where it ends. Run your own benchmark on your own job, using the exact chip, before signing anything long term.

3. What does year two cost?

Terms in this market range from on demand to 36 months, with longer commitments generally carrying better pricing. Neysa published terms sit within that range.

My reading: chip generations are turning over faster than three years. A long lock is therefore a bet that today's chip will still be worth having in three years. The buyer is the one placing that bet, not the seller.

That may pay off. It should be a conscious decision rather than a default.

Ask what happens to your rate when the next generation of chips arrives.

4. How do you leave?

Getting models, data and infrastructure setup out is the part nobody discusses at signing.

My reading: the whole category competes aggressively on getting customers in. The global providers have also built exit friction into their pricing through charges on data leaving their platforms.

As of 20 August 2026, I could find no provider in this market, large or small, publishing a simple plain language exit procedure.

Ask for exit terms in writing before the discount conversation, not after.

5. Who answers at 2am?

A training run dies overnight. Ask who picks up, how quickly they respond and what it costs.

Get the name of the team, not the name of the support tier.

My reading: this is one place where specialists have a structural advantage and buyers systematically ignore it. With a large provider, you are one account among millions. With a specialist, you may be one account among hundreds.

Where a specialist fits

Taking Neysa as the worked example, the price story is the easy pitch. The more important asset is the managed layer above the chips: the platform, monitoring across clusters, ready made setups and security layer.

That is much harder to copy than a price, and it is what specialists should be judged on.

Data staying inside the country also matters to regulated buyers. That is a real difference, not simply a marketing line.

Where a specialist does not fit

Occasional, unpredictable bursts are where a global provider can genuinely be simpler.

If your compute need is small and highly variable, use a global provider.

Steady, planned and predictable workloads are where the specialist case becomes stronger.

And there is one rule that should apply regardless of provider:

Any buyer who cannot get a written availability commitment for their start date should not sign a long term commitment at any price, from anyone.

This is a piece of opinion — our reading of what buyers should ask, based on public material available as of that date. It is not a statement of fact about any company. No company mentioned pays for the mention. Any company named here can write to hello@analystlayer.com; we respond within three working days and update the piece where the input is factual, with the update dated on this page.