Compare the options honestly.
Every comparison in this market is usually written by someone selling one side of it. We are buyer-side and paid the same whichever way you go, so here is the reasoning without the pitch.
Buy vs lease GPUs
Upfront capital, time to live, generation risk, exit terms and where the break-even actually sits.
Cloud vs bare metal
Pricing shape, performance, queueing and data gravity, plus the baseline-plus-burst split that usually wins.
GPU cloud providers
Hyperscalers, specialist GPU clouds and reserved bare metal compared on price, access, speed and risk.
H100 vs H200
Same architecture, very different memory. The one question that decides it, with current pricing.
B200 vs B300
Blackwell headroom against present price, and the facility requirement that gates both.
How to buy GPUs
Direct, reseller, buyer-side desk or rent: the four routes compared, including where each beats us.
Cluster cost calculator
Model GPU count, configuration and deployment assumptions into an initial view of cluster cost.
Nobody publishes a like-for-like number.
The GPU market is opaque by design. Rate cards omit egress, storage and support. Hardware quotes stop at the factory door and leave freight, duties and import handling for you to discover later. Reserved pricing looks superb until you notice the term length required to unlock it. None of this is unusual commercial behaviour, but it means the headline numbers you are given are almost never comparable.
These pages exist to normalise the comparison. Each one takes a decision buyers genuinely face, sets out what is actually being traded, and names the costs that tend to arrive after the decision is made.
Structure first, then vendor.
Start with the structural questions, because they constrain everything after. Buying versus leasing is a question about your cost of capital and your confidence in the roadmap, not about price per hour. Cloud versus bare metal is about whether you are paying for flexibility you are actually using.
Then the practical questions. How to buy GPUs sets out the routes to market and what each demands of you. Comparing GPU cloud providers covers the three tiers of the market and how to read a rate card that has been designed not to be read.
Hardware choices come last, once the structure is settled: H100 versus H200 and B200 versus B300. And if you are already on a hyperscaler or a specialist and want to know what else exists, the alternatives set covers CoreWeave, Lambda, AWS, Azure and Google Cloud.
Asked first, answered straight.
Per hour, renting is almost always more expensive. Over the life of a well-utilised cluster, owning usually wins. The honest comparison runs both routes on your own utilisation and cost of capital over the same period, on a landed-cost basis.
Normalise everything onto the same sheet before comparing. Rate cards routinely exclude storage, egress, support and minimum commitments, and the excluded lines are frequently where the real difference between providers sits.
Answer the structural questions first. Whether you buy or lease, and whether you run bare metal or cloud, constrains the hardware shortlist far more than the specification differences between individual accelerators do.
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.
The two-quarter question.
Steady demand earns the sustained rate. Bursty demand should be renting on purpose, not by accident.
One forecast question forks the whole cost curve