AI compute, explained in plain words.
We sell the real GPUs that real AI datacenters run on. Every spec on this page is followed by a plain-language note that tells you what it means and why it matters — no computer science degree required.
Every AI machine on this page — from a desktop card to a full datacenter rack — is described by the same five specs. Here's what each one really means.
The CPU runs the operating system, prepares data, and hands work off to the GPU. For AI, it's supporting cast, not the star: a slow CPU can leave an expensive GPU waiting, but a fast CPU can't make up for a weak GPU.
Where it shows up here: the GB200 rack even builds its own CPUs in (Nvidia's Grace chip) — that's how central this pairing is at datacenter scale.
A GPU has thousands of small cores that all do the same simple math at the same time — exactly the kind of math AI models are built from. That's why GPUs run AI models 10-100x faster than a CPU alone.
Rule of thumb: this is the single spec that most determines how fast — and how big a model — your AI work can run.
GPU memory (VRAM, or HBM on datacenter chips) is the GPU's own private, very fast memory. The entire model has to fit inside it while the GPU works. Run out, and the model slows to a crawl or refuses to load.
Rule of thumb used on this page: a model with N billion parameters needs roughly N × 1GB × 1.2 (20% working room) of GPU memory just to run. See the Model Advisor below.
This is the computer's general-purpose memory, separate from the GPU's own memory. It holds the operating system, the dataset, and everything waiting to be handed to the GPU.
Rule of thumb: real GPU servers ship with hundreds of GB to multiple TB of system RAM so the GPUs are never left waiting on data.
Every part draws electricity, and the GPU draws far more than anything else. Power has to be planned for and delivered reliably — this is a real, physical constraint, not a footnote.
Made real below: every watts figure on this page is translated into "how many homes" and "how many hours of an EV battery" it's equivalent to.
These three figures are pre-verified — we use them exactly as given, not re-derived. The math: parameters (in billions) × 1GB × 1.2 for 20% working room.
284B parameters
minimum GPU memory needed
744B parameters
minimum GPU memory needed
1.6T parameters
minimum GPU memory needed
Two customer profiles, two honest recommendations — with the full math shown.
What's a cluster? A cluster is a group of separate machines wired together with fast networking so they work as one much bigger machine.
The honest difference: in a real datacenter cluster like the GB200 NVL72 below, GPUs are linked by NVLink into one shared, unified pool of memory with microsecond-level latency — 72 GPUs genuinely behave like a single machine. A pile of desktop cards connected over an ordinary network can split up work, but each card's memory stays separate — there's no unified pool, the interconnect is far slower, and consumer boards aren't built for continuous datacenter duty. It can help with small jobs, but it is not the same thing as a real cluster, and it hits a wall the moment a model needs more memory than any single card has.
The cheapest honest way to run one big model in production.
Targets DeepSeek V4 Flash (340.8 GB needed) — 43.2 GB of headroom.
2,000W of GPUs plus about 15% for host CPU, fans, and power-supply losses. That's about 1.9 average homes running continuously, and it would drain a fully-charged EV battery (90kWh) in about 39 hours (~1.6 days).
2 x $37,500 per-GPU, at full deployable-system scale (host CPU, networking, chassis) — priced from real 8-GPU DGX B200 systems (~$300k/8). Smaller node counts like this one often cost somewhat more per GPU due to fixed overhead, so treat this as a conservative estimate.
Two B200 chips linked over NVLink pool into 384GB of shared GPU memory — comfortably past the 340.8GB DeepSeek V4 Flash needs, without paying for a full rack you don't need yet.
Real capacity today, and room to grow into tomorrow's models.
Targets DeepSeek V4 Pro (1,920 GB needed) — 11,904 GB of headroom. Also covers: DeepSeek V4 Flash, GLM 5.1, DeepSeek V4 Pro.
Nvidia's nominal whole-rack spec (real-world load can run a bit higher). That's about 100 average homes running continuously, and it would drain a fully-charged EV battery (90kWh) in about 45 minutes.
The full rack, at its industry-estimated price — a single unit, not a custom multi-server quote.
13.5TB of pooled memory covers all three models at once with room to spare — run DeepSeek V4 Pro today and still have over 11TB free for growth, more concurrent models, or longer context.
Desktop card → datacenter GPU chip → full datacenter rack. Every spec below is followed by a plain-language note.
A real Blackwell GPU you can put in a desktop tower today.
$1,999
MSRP at launch (Jan 2025). Retail prices have run well above this due to demand — treat it as the honest list price, not a street-price guarantee.
Enough to load small AI models entirely on one card, but consumer memory isn't pooled with other cards — each card is an island.
Needs a serious desktop power supply (1000W+ recommended) — this alone draws more than most kitchen appliances. That's about 0.5x the continuous draw of one typical home, and it would drain a fully-charged EV battery (90kWh) in about 157 hours (~6.5 days).
Pairs with any consumer desktop CPU (e.g. 6-8 cores) — the CPU just needs to keep up with feeding the card, not do the AI math itself.
32-64 GB of ordinary desktop RAM is plenty to stage data for a single card like this.
Good for: Learning, prototyping, and running small open models on your own desk.
Buy this when: Buy this to learn on, or to prototype before committing to real datacenter spend.
The actual building block every AI datacenter is made of.
$35,000
Nvidia doesn't publish list pricing for datacenter GPUs — this is a mid-point of industry-reported estimates ($30,000-$40,000 per GPU).
6x a desktop card's memory, and — unlike a desktop card — it can be linked to other B200s over NVLink so several chips act like one big pool of memory.
Built to run flat-out, 24/7, in a datacenter with real cooling and power delivery — not a home outlet. That's about 0.8x the continuous draw of one typical home, and it would drain a fully-charged EV battery (90kWh) in about 90 hours (~3.8 days).
Deployed with a datacenter-grade host (e.g. an NVIDIA Grace CPU or a server-class x86 CPU) that keeps the GPUs continuously fed.
Paired with hundreds of GB of server RAM per node so large datasets never leave the GPUs waiting.
Good for: A real, small AI server built to run one large model reliably.
Buy this when: Buy this when a desktop card's 32GB genuinely isn't enough room for the model you need to run.
72 B200 GPUs and 36 Grace CPUs, engineered to behave like one machine.
$3,000,000
Industry-estimated at roughly $2M-$3M per rack; Nvidia sells these through OEM partners rather than a public price list.
72 GPUs x 192GB, linked by NVLink into one shared pool — the whole rack can address memory as if it were a single giant GPU.
Nvidia's nominal spec for the whole rack; real deployments have measured 130-132 kW under full load. That's about 100 average homes running continuously, and it would drain a fully-charged EV battery (90kWh) in about 45 minutes.
36 built-in Grace CPUs (one per two B200s) — the CPU layer is part of the rack, not something you add yourself.
Roughly 17 TB of CPU-side system memory ships across the rack, on top of the 13.5 TB of GPU memory above.
Good for: Running the largest open models available today, with room to run several at once.
Buy this when: Buy this when you need real headroom to grow — not just to fit today's model, but next year's.
No payment required. Tell us your name and how to reach you, and we'll follow up. You'll get a request number, and every request is logged at /requests.