Hardware Products
Every number below comes with a plain-English explanation of what it actually means.
RTX 4090
Desktop card. What most people mean by "a GPU."
$1,800
Price
About the price of a decent used car down payment — this is the cheapest way into serious AI compute.
24GB VRAM
Memory (how much "AI model" it can hold)
24GB of VRAM is the card's short-term memory for AI data. It's enough to run small-to-medium AI models (up to roughly 15-20B parameters) entirely on one card.
450W
Power draw
450W is about what a hair dryer or a microwave draws. One card is fine on a normal home outlet.
In real-life terms: 450W is
0.38 average U.S. homes' worth of power,
or enough to add 0.0050 of a full EV battery charge every hour
(about 200 hours for a full charge).
RTX 6000 Ada
Workstation card. The 'prosumer' step up.
$6,800
Price
Roughly 4x the price of an RTX 4090 — you're paying for double the memory and data-center-grade reliability.
48GB VRAM
Memory (how much "AI model" it can hold)
48GB VRAM means one card can hold twice as much model as an RTX 4090 — enough for many 30-40B parameter models solo.
300W
Power draw
300W — less power than the RTX 4090 despite doing more, because it's built for efficiency in workstations.
In real-life terms: 300W is
0.25 average U.S. homes' worth of power,
or enough to add 0.0033 of a full EV battery charge every hour
(about 300 hours for a full charge).
H100 SXM
Data-center GPU. What companies actually rent by the hour in the cloud.
$30,000
Price
$30,000 buys one card, not a computer — it still needs an expensive server chassis to plug into.
80GB VRAM
Memory (how much "AI model" it can hold)
80GB VRAM — built for training and serving large models, and multiple H100s can pool memory together over NVLink/NVSwitch.
700W
Power draw
700W per card — that's about 1.5 hair dryers running nonstop, per GPU.
In real-life terms: 700W is
0.58 average U.S. homes' worth of power,
or enough to add 0.0078 of a full EV battery charge every hour
(about 129 hours for a full charge).
DGX H100
A complete 8-GPU AI server, built and sold by NVIDIA as one box.
$200,000
Price
$200,000 for the whole server: 8 H100 GPUs, CPUs, networking, and cooling, all pre-integrated.
640GB VRAM
Memory (how much "AI model" it can hold)
640GB VRAM = 8 H100 GPUs (80GB each) whose memory is pooled together via NVLink, acting like one giant memory pool.
10,200W
Power draw
10,200W — as much as roughly 8-9 average American homes use at once, all in a single server rack unit.
In real-life terms: 10,200W is
8.50 average U.S. homes' worth of power,
or enough to add 0.1133 of a full EV battery charge every hour
(about 9 hours for a full charge).
GB200 NVL72 rack
A full server rack: 72 Blackwell GPUs wired together as one machine.
$3,000,000
Price
$3,000,000 for one rack — this is the scale companies buy in when they're training frontier AI models.
13,824GB VRAM
Memory (how much "AI model" it can hold)
13,824GB VRAM — nearly 14 terabytes of GPU memory, all connected by NVLink so it behaves like one enormous shared pool.
120,000W
Power draw
120,000W — as much electricity as a small apartment building, from a single rack the size of a large refrigerator.
In real-life terms: 120,000W is
100.00 average U.S. homes' worth of power,
or enough to add 1.3333 of a full EV battery charge every hour
(about 1 hours for a full charge).
How we calculate this: homes powered = watts ÷ 1200 (an average U.S. home draws about 1200W continuously).
EV charges per hour = watts ÷ 1000 ÷ 90 (converts watts to kWh, then divides by a ~90kWh EV battery).