Five tiers of real NVIDIA hardware

What this means: GPU memory is the shelf space a model's brain sits on while it's "thinking." A model that needs more memory than a card has simply cannot be loaded onto it — there's no partial credit.

What this means: TDP: how many watts the card can draw flat-out. This is what actually shows up on your electric bill and what your power supply has to be able to deliver.
In real terms: 0.48 of one home's continuous draw; running nonstop, it would burn through a typical EV battery's charge (90 kWh) every about 6.5 days.

Range seen in the market: $4,300–$5,000 street (as of July 2026). Launched at $1,999 in early 2025. A global GDDR7 memory shortage — the same chips going into AI datacenters are competing for the same factories — has more than doubled street prices.

What this means: HBM ("high-bandwidth memory") is a faster, more expensive kind of memory built right next to the chip. 80 GB sounds like a lot until you meet a 700-billion-parameter model.

What this means: Nearly 5x a desktop card's power draw — this is why datacenter GPUs live in rooms with industrial cooling, not on a desk.
In real terms: 0.58 of one home's continuous draw; running nonstop, it would burn through a typical EV battery's charge (90 kWh) every about 5.4 days.

Range seen in the market: $30,000–$40,000 (individual unit, OEM pricing). SXM-format GPUs aren't sold individually at retail; this is the per-GPU price when bought as part of a certified 8-GPU server.

What this means: 76% more memory than an H100 in the same power envelope — pure upside if the extra room is what you need.

What this means: Identical power draw to the H100 despite the memory bump — more capacity without a bigger electric bill per GPU.
In real terms: 0.58 of one home's continuous draw; running nonstop, it would burn through a typical EV battery's charge (90 kWh) every about 5.4 days.

Range seen in the market: $32,000–$45,000 (individual unit, OEM pricing). Like the H100, sold as part of certified 8-GPU servers rather than individually at retail.

What this means: Because NVLink lets the 8 GPUs pool their memory, a model can be split across all of them and treated as one 1,128 GB machine — something you can't do by just plugging 8 desktop cards into one PC.

What this means: 5.6 kW is just the GPUs; the whole system — CPUs, networking, storage, cooling fans — pulls closer to 8 kW at the wall. We show the full-system number since that's what your power bill sees.
In real terms: about 7 homes running around the clock; running nonstop, it would burn through a typical EV battery's charge (90 kWh) every about 11.2 hours.

Range seen in the market: $320,000–$420,000 (fully integrated, OEM list). This is a complete rack-mountable server, not just 8 loose GPUs — the price includes the NVLink fabric, CPUs, chassis, and integration that make the 8 GPUs behave like one machine.

What this means: 20 TB (20,480 GB) shared across all 72 GPUs as if it were one pool — enough headroom to run several of today's largest open models at once, with room for tomorrow's bigger ones.

What this means: 135,000 W continuous, up to 155,000 W at peak load — this is why racks like this live in purpose-built datacenters with their own substations and liquid cooling, never in an office.
In real terms: about 112 homes running around the clock; running nonstop, it would burn through a typical EV battery's charge (90 kWh) every about 40 minutes.

Range seen in the market: $3.7M–$4.0M bare hardware; $5M–$6.5M fully provisioned with cooling, networking, and support. The bare-hardware price buys you the rack itself. Actually running it — datacenter cooling, networking to the rest of your fleet, support contracts — typically adds another $1.3M–$2.5M.