The hardware lineup
Four real NVIDIA products spanning desktop to full datacenter rack. GPU memory is the number that decides whether a model fits at all; watts decide what it costs to run and how much electrical infrastructure it needs.
Desktop / Professional GPU
NVIDIA RTX PRO 6000 Blackwell Workstation Edition
A single high-end card that fits in a desktop tower. The entry point into serious AI hardware.
- GPU memory 96 GB GPU memory is the shelf space a model's weights sit on while it runs. If a model needs more memory than a GPU (or group of GPUs) has, it simply cannot be loaded — no amount of extra time fixes that.
- Power draw 600 W That's 0.50x the continuous draw of an average home (1,200 W). Running it for one hour uses 0.60 kWh — about 0.7% of a typical 90 kWh EV battery.
- CPU / RAM This is an add-in card, not a full computer — you supply the host PC/workstation (CPU + system RAM). Any modern workstation-class CPU and 64GB+ of system RAM is enough to keep it fed; the GPU's own 96GB of memory is what actually holds the AI model.
- Price $16,000 estimated market price NVIDIA's own online Marketplace listing price as of August 2026. This card launched at an MSRP of $8,565 in March 2025, but NVIDIA raised its own listed price to $13,250 in June 2026 and then to $16,000 in August 2026, reportedly due to a GDDR7 memory shortage. Independent retailers have shown prices anywhere from about $14,000 to $15,500. Treat $16,000 as a current snapshot, not a stable number.
High-Memory Server GPU
NVIDIA H200 NVL
A datacenter GPU built to be bridged together with others over NVLink — the real way multiple GPUs share one big pool of memory.
- GPU memory 141 GB GPU memory is the shelf space a model's weights sit on while it runs. If a model needs more memory than a GPU (or group of GPUs) has, it simply cannot be loaded — no amount of extra time fixes that.
- Power draw 600 W That's 0.50x the continuous draw of an average home (1,200 W). Running it for one hour uses 0.60 kWh — about 0.7% of a typical 90 kWh EV battery.
- CPU / RAM Sold as a PCIe card for a partner server chassis (not a full computer by itself). Built to be linked 2-way or 4-way over NVLink bridges at 900 GB/s per GPU, so a group of them can act like one much larger pool of GPU memory — something desktop cards cannot do.
- Price $32,000 estimated market price NVIDIA sells H200 mainly through OEMs/integrators, not at a public list price. Industry estimates place a single GPU roughly between $28,000 and $40,000 depending on configuration and volume; $32,000 is a commonly cited midpoint used here.
Datacenter GPU System
NVIDIA DGX H100
A complete, ready-to-run 8-GPU server — the GPUs are already wired together internally with NVSwitch, so it behaves as one coherent machine, not eight separate cards.
- GPU memory 640 GB 8 GPUs x 80 GB each, combined into one pool. GPU memory is the shelf space a model's weights sit on while it runs. If a model needs more memory than a GPU (or group of GPUs) has, it simply cannot be loaded — no amount of extra time fixes that.
- Power draw 10,200 W That's 8.50x the continuous draw of an average home (1,200 W). Running it for one hour uses 10.20 kWh — about 11.3% of a typical 90 kWh EV battery.
- CPU / RAM Includes everything: 2x Intel Xeon Platinum 8480C CPUs (112 cores total) and 2 TB of system RAM, plus 8x 400 Gb/s InfiniBand networking ports to link multiple DGX systems together. You are not assembling this from parts — it arrives as one machine.
- Price $350,000 estimated market price NVIDIA sells DGX systems through direct/partner channels with negotiated pricing — there is no public MSRP. Reseller and industry estimates for a fully configured DGX H100 commonly range from about $250,000 to $400,000; $350,000 is used here as a representative midpoint.
Full Rack-Scale System
NVIDIA GB200 NVL72
An entire liquid-cooled server rack — 72 Blackwell GPUs and 36 Grace CPUs wired together as a single giant machine over a 130 TB/s NVLink switch fabric. This is what real "AI supercomputer" hardware looks like.
- GPU memory 13,400 GB 72 GPUs x 186 GB each, combined into one pool. GPU memory is the shelf space a model's weights sit on while it runs. If a model needs more memory than a GPU (or group of GPUs) has, it simply cannot be loaded — no amount of extra time fixes that.
- Power draw 120,000 W That's 100.00x the continuous draw of an average home (1,200 W). Running it for one hour uses 120.00 kWh — about 133.3% of a typical 90 kWh EV battery.
- CPU / RAM 36 NVIDIA Grace CPUs (2,592 Arm cores total) plus 17 TB of aggregate CPU memory, all connected to the GPUs over NVLink-C2C so the CPUs and GPUs share work efficiently. Delivered as a complete rack, not individual parts.
- Price $3,000,000 estimated market price NVIDIA does not publish a rack-level list price — GB200 NVL72 is sold through OEM/ODM partners (Dell, HPE, Supermicro, Foxconn, and others) with configuration-dependent, often confidential pricing. Industry estimates cluster around $2 million to $3.4 million per rack depending on configuration; $3,000,000 is used here as a representative midpoint.