GPU cloud providers compared, from the buyer's side.
Every provider comparison you find online is usually written by a provider. This one is not. We source across all of them and take no position in which one you pick, so here is how the tiers actually differ and which questions decide it.
Hyperscaler, specialist cloud, or your own metal.
| Dimension | Hyperscalers | Specialist GPU clouds | Reserved bare metal |
|---|---|---|---|
| Examples | AWS, Azure, Google Cloud | CoreWeave, Lambda, Nebius, Crusoe, GMI Cloud, Spheron | Your hardware, or a named facility deployment |
| Relative price | Around 86% above specialist clouds for the same H100 | H100 SXM roughly $1.49 to $3.50 per GPU-hour on demand | Lowest per unit at sustained utilisation, negotiated on term |
| Access to the metal | Virtualised, managed services layered on top | Varies: some offer true bare metal with root and dedicated IP | The whole node and the whole fabric, nothing between |
| Time to start | Minutes, subject to quota | Minutes to days depending on capacity | Roughly 8 to 10 weeks for a racked deployment |
| Best for | Existing cloud estates, tight integration needs | Spiky or medium-horizon workloads, fast starts | Sustained baselines you can forecast two quarters out |
| Main risk | Cost at sustained utilisation | Capacity and priority when the market tightens | Commitment against an uncertain roadmap |
Pricing ranges from published provider pricing and the GetDeploying GPU Price Index, 2026. Hyperscaler premium from Compute Exchange, Best GPU Rental Platforms 2026. Provider capabilities from their own published documentation. Availability and rates move constantly, so treat every figure as indicative.
Four questions that settle it.
If you are evaluating a specific provider, the same reasoning applied to CoreWeave, Lambda, AWS, Azure and Google Cloud is set out separately.
What is the horizon? Under a quarter, use on demand. Two quarters or more that you can forecast, reserve it. That single question eliminates most of the shortlist before you compare a rate card.
Does the workload need the fabric? Large distributed training lives or dies on interconnect quality. Single-node fine-tuning and most inference does not, and paying for a premium fabric you cannot saturate is the most common overspend we see.
Where must the data sit? Residency narrows the provider list faster than price does, and it is cheaper to discover that now than during procurement.
What happens when the market tightens? Ask every provider what your priority is when capacity is short. The answer separates a commitment from a hope, and it is the question buyers ask least often.
Asked first, answered straight.
The specialist tier includes CoreWeave, Lambda, Nebius and Crusoe, alongside newer entrants such as GMI Cloud and Spheron. They sit between the hyperscalers and bare-metal ownership, and they compete mainly on price, availability and how close you get to the metal.
Materially, yes. Published comparisons put hyperscaler H100 pricing at roughly 86% above dedicated GPU clouds for the same silicon. The gap is not the hardware, it is the platform layer around it.
There is no single answer, which is why we do not publish a ranking. Best depends on fabric quality for distributed training, geography for inference latency, residency for regulated data, and the commercial terms underneath. We shortlist against your requirement rather than a leaderboard.
No. We hold no inventory and take no position. You transact with the provider on their direct terms and our compensation sits on the supply side under disclosed arrangements, which is why we can tell you when a provider is wrong for you.
Working through this on a real requirement?
Twenty minutes with the desk, no pitch and no quote at the end of it. Tell us roughly what you need and we will come back within one business day.
Sent. We are on it.
Your enquiry has landed with the desk. Acknowledged within one hour.