SSP, Strategic Supply Partners Live inventory
The desk
GPU procurementNodes and clusters, landed to your rack ● Live inventoryVerified lines, indicative pricing
By platform
B300 and GB300Current-generation Blackwell H200 and H100New stock and fleet rotation All platformsThe full catalogue
Rest of the stack
Memory, servers and storageThe kit around the accelerators Networking and opticsTransceivers, switching, cabling
Capacity without capex
GPU-as-a-serviceReserved, committed spend, on demand Bare-metal leasingDedicated clusters on a term
Decide first
Cloud vs bare metalWhere each one actually wins Reserved vs on demandThe break-even, calculated Cluster cost calculatorThe whole bill, not the GPU line
Space and power
Colocation and powerSingle racks to 100 MW Above 40 kW per rackWhere the shortlist collapses
Where
United States Europe Asia Pacific Middle East and Africa All regions →
Supply side
Have capacity to fill?Tell us what you hold
Analysis
InsightsWhat things cost, with the numbers Sizing guidesMemory, nodes, power, cost ComparisonsBuy, lease or rent
Start here
How an engagement worksThe first week, held to dates Find your fitThree questions, the right desk Glossary
Talk to the desk
Blog · From the desk

Rent, reserve, or own: how to match compute commitment to your workload

Every team buying AI compute faces the same three doors. Most pick the wrong one, not because they are careless, but because each door has a salesman standing in front of it. Here is how I run the decision when nobody is selling anything.

Strategic Supply Partners25 August 20266 min read
Rent, reserve or own: three option cards with the deciding criterion under each

The three doors

What the market actually charges

  • Renting an H100 SXM on a specialist GPU cloud runs roughly $1.49 to $3.50 per GPU per hour on demand, with H200 starting near $3.72 and B200 between $2.12 and $6.04 depending on tier and commitment.Source: GPU rental market survey, GetDeploying GPU Price Index and provider pricing pages, 2026
  • Hyperscalers charge around 86% more than dedicated GPU clouds for the same H100 silicon.Source: Compute Exchange, Best GPU Rental Platforms 2026
  • Owned H100 nodes in the current secondary market sit near $240,000 for an 8-GPU node, which is where the rent-versus-own arithmetic starts to bite at sustained utilisation.Source: SSP verified inventory, August 2026

Renting is on-demand cloud: pay by the hour, walk away whenever. Reserving is committed capacity: a defined rate for a defined term on dedicated hardware, usually bare metal in a named facility. Owning is exactly what it sounds like: the nodes are yours, along with everything that comes with them.

None of these is wrong. Each one is wrong for somebody, and the market is very good at putting each door in front of the wrong buyer. Cloud providers sell flexibility to teams with none of the variability that justifies it. Hardware channels sell ownership to teams who have never racked a server. And plenty of desks will sign you to a five-year reserve because the commission is better on long paper.

The question that actually decides it

Strip everything else away and the decision comes down to one thing: how much of your compute demand can you see, and for how long?

Not how much compute you want. How much you can see. A training roadmap with funded projects across the next two quarters is visible demand. A hunch that usage will grow is not. The three doors map onto three levels of visibility:

  • You can see almost nothing. You are experimenting, evaluating, or demand genuinely spikes and vanishes. Rent. The on-demand premium is real, but you are actually using the flexibility it buys. This is the only situation where on-demand is cheap.
  • You can see a baseline. There is a floor of usage that simply does not go away: the training queue is never empty, the inference traffic never drops below a level. Reserve that floor. At sustained utilisation, on-demand rates run multiples of a reserved rate for identical silicon, and every month you leave the baseline on hourly pricing you are paying that multiple for flexibility you never use.
  • You can see years, and the workload runs hard. High utilisation, long horizon, a team that can operate hardware or a partner who will. Now owning starts to win, because you stop paying anyone else's structure at all.
The takeaway

Match the commitment to the visibility, not to the ambition. Reserve what you can see, rent what you cannot, and only own what you can see for years.

Where teams actually get it wrong

In practice I see the same two mistakes over and over.

The first is staying on rent too long. The workload became a baseline eighteen months ago, but migration feels like a project, so the team stays put and the cloud bill quietly becomes the largest line in the budget. Here is the uncomfortable arithmetic: if your utilisation is sustained, the two quarters you spend deciding typically cost more than the entire move.

The second is owning for the wrong reasons. Ownership appeals to instinct: an asset, control, no landlord. But the asset depreciates on NVIDIA's release schedule rather than yours, and the control comes bundled with import, deployment, operations, spares and an end-of-life problem. Owning is the right answer for a specific buyer, and an expensive identity statement for everyone else.

The blend nobody sells you

The answer for most serious teams is not one door. It is a blend: the visible baseline on reserved capacity at reserved pricing, the genuinely unpredictable layer on-demand, and ownership only where the numbers survive a spreadsheet built by someone with no stake in the outcome.

Nobody sells the blend, because nobody makes money on the whole of it. Which is precisely why it tends to be right.

Straight answers

Asked first, answered straight.

Should I rent, reserve or buy GPUs?

Rent when you cannot yet see your demand curve. Reserve once a baseline is visible and stable. Buy when the hardware will run hard for most of its useful life and you can carry the capital. The mistake is choosing on price alone, because each route is quoted by someone paid to sell that route.

Is renting GPUs more expensive than buying?

Per GPU-hour, almost always yes. That is the wrong comparison. Renting prices flexibility, and flexibility is worth paying for while your utilisation is uncertain. Once utilisation is consistently high, the same premium becomes pure waste.

What utilisation makes buying worthwhile?

There is no universal threshold, because it depends on your cost of capital, the power price at your chosen facility and how long the generation stays competitive. Run the three routes on your own numbers over the same period rather than accepting a rule of thumb.

Working through this decision on a real requirement? Twenty minutes with the desk, no pitch, no quote at the end of it. We run your numbers, not ours.

Talk to the desk
Keep reading

The reserved leasing desk →

Buy vs lease, in one table →

Run the numbers on your cluster →

Talk to the desk

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.

Acknowledged within one hour, first sourcing pass within one business day.
✓

Sent. We are on it.

Your enquiry has landed with the desk. Acknowledged within one hour.