The Hardware

32 gigabytes of fast video memory. That's the workspace the GPU uses to hold model weights while generating tokens. Bigger = can run larger models without swapping to slow regular RAM. At full precision, this fits models up to about 25 billion parameters.

Thermal Design Power — how much electricity it draws under load. 575 W is roughly the same as running four gaming consoles at once, or about half an average home's continuous draw (which is ~1,200 W).

MSRP at launch (January 2025) was $1,999, but a GDDR7 memory shortage drove street prices to $4,300–$4,900 by August 2026. You're paying a supply-chain premium — if that gap concerns you, consider waiting or stepping up to the RTX PRO 6000 Blackwell for more memory at a more predictable enterprise price.

96 gigabytes — three times an RTX 5090. At full precision, this fits models up to about 80 billion parameters. For the models on our AI page you'd still need multiple cards or a server, but this card alone handles many useful open-source 70B models like Llama 3 70B.

600 W is about half an average home's continuous draw. Expect a dedicated 20 A circuit in your workspace.

Workstation GPUs carry a premium for ECC memory protection and pro drivers. Enterprise buyers often pay up to $13,000 for volume orders with support contracts.

HBM3e (High Bandwidth Memory) is a different type of memory than GDDR7 — it sits directly on the same package as the GPU chip, giving extreme bandwidth. Each H200 has 141 GB; eight of them connected via NVLink give you 1,128 GB of shared GPU memory. That's enough to run all three models on our AI page.

The whole server pulls 5.6 kilowatts — about 4.7 average homes running simultaneously. You'll need a dedicated power circuit, and data-center-grade cooling (CRAC units, raised floors, or liquid cooling). Not for a typical office.

This is professional infrastructure. At this level you're buying reliability, NVLink interconnects, enterprise support, and the ability to treat all 1,128 GB as one big shared pool of GPU memory. Compare to renting equivalent cloud compute, which runs $15–25/hr, meaning you break even in roughly 2–3 years.

20 terabytes of GPU memory — all accessible as a single address space through NVLink. For reference, our largest listed model (GLM-5.2 at 744B params) only needs 893 GB. This rack can run that model with 22× headroom to spare, or run dozens of large models simultaneously.

137 kilowatts is the electricity draw of about 114 average homes running at the same time. At $0.10/kWh, that's roughly $120,000/year in electricity alone. This requires a dedicated data-center power delivery system (PDU, UPS, generator backup) and industrial-scale cooling — liquid cooling is effectively mandatory at this wattage.

Price varies by configuration, memory capacity, and negotiated enterprise discounts. At the high end this is equivalent to hiring 30+ senior engineers for a year. Most organizations at this scale are replacing $10M+/year cloud GPU bills, making the economics compelling within 12–18 months.