The hardware we sell
From a $2,500 desktop card up to a $3,000,000 datacenter rack. Every product below shows its real price, GPU memory, and power draw — plus what each number means and when it actually matters.
Desktop card
NVIDIA RTX 4090
A top-end gaming/workstation GPU. Great for learning, prototyping, and running small models — not built for giant open-source models.
| Price | $2,500 |
|---|---|
| GPU memory | 24 GB |
| Power draw | 450 W |
| Interconnect | PCIe (no NVLink between cards) |
What "price" means here: the one-time cost to own this
hardware outright. Bigger models need more of it, so price is really a proxy for "how much
capability can I afford" — compare price against GPU memory, not on its own.
What "GPU memory" means: the model's entire brain (its
parameters) has to fit in GPU memory before it can answer a single question. A model needs
roughly 1 GB of GPU memory per billion parameters, plus working room. Run out of GPU memory
and the model simply won't load — no amount of CPU RAM or clever code fixes that.
What "power draw" means: how many watts this hardware
pulls from the wall while running flat-out. It matters for two reasons: your electricity bill,
and whether your building's cooling and circuits can actually handle it.
450 W is about
0.38x an average home's continuous draw
(1 home ≈ 1,200 W).
Datacenter GPU
NVIDIA H100 SXM
The workhorse datacenter GPU. Sold as an individual accelerator that slots into a server baseboard alongside others, wired together with NVLink.
| Price | $32,000 |
|---|---|
| GPU memory | 80 GB |
| Power draw | 700 W |
| Interconnect | NVLink (900 GB/s between GPUs) |
What "price" means here: the one-time cost to own this
hardware outright. Bigger models need more of it, so price is really a proxy for "how much
capability can I afford" — compare price against GPU memory, not on its own.
What "GPU memory" means: the model's entire brain (its
parameters) has to fit in GPU memory before it can answer a single question. A model needs
roughly 1 GB of GPU memory per billion parameters, plus working room. Run out of GPU memory
and the model simply won't load — no amount of CPU RAM or clever code fixes that.
What "power draw" means: how many watts this hardware
pulls from the wall while running flat-out. It matters for two reasons: your electricity bill,
and whether your building's cooling and circuits can actually handle it.
700 W is about
0.58x an average home's continuous draw
(1 home ≈ 1,200 W).
Full system
NVIDIA DGX H100 (8x H100)
A turnkey server: 8 H100 GPUs, CPUs, networking, and storage in one box, fully wired with NVLink/NVSwitch. Plug it in and it works.
| Price | $350,000 |
|---|---|
| GPU memory | 640 GB |
| Power draw | 10,200 W |
| Interconnect | NVLink + NVSwitch (all 8 GPUs act as one pool) |
What "price" means here: the one-time cost to own this
hardware outright. Bigger models need more of it, so price is really a proxy for "how much
capability can I afford" — compare price against GPU memory, not on its own.
What "GPU memory" means: the model's entire brain (its
parameters) has to fit in GPU memory before it can answer a single question. A model needs
roughly 1 GB of GPU memory per billion parameters, plus working room. Run out of GPU memory
and the model simply won't load — no amount of CPU RAM or clever code fixes that.
What "power draw" means: how many watts this hardware
pulls from the wall while running flat-out. It matters for two reasons: your electricity bill,
and whether your building's cooling and circuits can actually handle it.
10200 W is about
8.50x an average home's continuous draw
(1 home ≈ 1,200 W).
Datacenter rack
NVIDIA GB200 NVL72 Rack
A full rack of 72 Blackwell GPUs wired together as a single giant accelerator. This is what companies buy to serve the largest models at scale, with room to grow.
| Price | $3,000,000 |
|---|---|
| GPU memory | 13,500 GB |
| Power draw | 120,000 W |
| Interconnect | NVLink switch fabric across all 72 GPUs |
What "price" means here: the one-time cost to own this
hardware outright. Bigger models need more of it, so price is really a proxy for "how much
capability can I afford" — compare price against GPU memory, not on its own.
What "GPU memory" means: the model's entire brain (its
parameters) has to fit in GPU memory before it can answer a single question. A model needs
roughly 1 GB of GPU memory per billion parameters, plus working room. Run out of GPU memory
and the model simply won't load — no amount of CPU RAM or clever code fixes that.
What "power draw" means: how many watts this hardware
pulls from the wall while running flat-out. It matters for two reasons: your electricity bill,
and whether your building's cooling and circuits can actually handle it.
120000 W is about
100.00x an average home's continuous draw
(1 home ≈ 1,200 W).