Frequently asked questions
Plain-language answers, no jargon left unexplained. If something below doesn't cover it, every number on the catalog page has its own tooltip explainer too.
Can I actually buy something here?
Every product listed is real hardware at a real, sourced price — this isn't a mockup catalog. What you can't do here is pay online: there's no checkout, no accounts, and no payment processing. Click Request a quote on any product or build, and it's saved with a request number you can look up later on the Requests page.
How is the "memory required" number calculated?
A model's parameter count in billions, multiplied by 1 GB per billion parameters, plus 20% extra headroom for things like the KV cache and activation buffers that a model needs while it's actually generating output. So a 70B model needs roughly 70 GB × 1.2 = 84 GB of GPU memory. We never recommend hardware that doesn't clear that number.
Why GPU memory and not CPU or RAM?
A big AI model's weights have to load into GPU memory before the GPU's own cores can run it — that's the hard requirement. Regular CPU and system RAM mostly just keep the machine's operating system running and shuttle data around; more of either won't let an undersized GPU hold a bigger model or make it run faster. That's why every spec on this site leads with GPU memory and watts, not CPU or RAM.
Where do the recommended models come from?
The three models on the Model advisor page are pulled from arena.ai's leaderboard for agentic tasks, filtered to open-source models — not picked because they're popular or because we sell hardware sized for them. If you have a different model in mind, the calculator at the bottom of that page runs the same real math for any parameter count you type in.
Is a "cluster" here a real cluster, or just several cards plugged in?
A real one. When a model needs more memory than any single product offers, we combine multiple units of the same clusterable server or rack — the kind of hardware that ships with NVLink already wiring its GPUs together, and that real datacenters network together with InfiniBand for multi-server setups. That's a different world from plugging several desktop cards into one PC over ordinary PCIe, which is slower and not something a datacenter would actually do at this scale. Cluster pricing here also doesn't include the extra InfiniBand networking gear a real multi-server cluster needs on top of the machines themselves — we say so every time a cluster build comes up.
What do the power numbers ("1.2x a home," "0.6 EV batteries") actually mean?
Watts measure electricity draw, not compute power. To make that concrete, every watt figure is also shown against two reference points: 1200 W, a rough estimate of an average home's continuous draw, and 90 kWh, a rough estimate of an electric car's usable battery capacity. Run a piece of hardware flat-out for 24 hours and we convert that to kWh, then show how many home-equivalents or EV-battery-equivalents that represents.
Are the prices exact?
They're sourced from real listings — NVIDIA's own product pages and major retailers (B&H, Micro Center, NVIDIA's marketplace) — as of July 2026, and we say so on every price tooltip. One exception, disclosed here honestly: NVIDIA doesn't publish rack-level pricing for the GB300 NVL72 publicly, so that figure is a widely reported estimate rather than a published list price. Everything else reflects an actual current listing.
What happens after I request a quote?
Your request is saved with a unique request number, which you can look up any time on the Requests page. Nothing is charged and no account is created — it's simply a record that you're interested in a specific build at a specific price, the same way you'd request a quote from any B2B hardware vendor before purchasing.