Roman's AI Hardware Co.

The hardware

Six products, from a single desktop card to a full rack. Every number below is explained, not just labeled — read the "How to read these numbers" section first if any of this is new to you.

How to read these numbers

GPU memory — the number that actually decides whether a model runs

A large language model is, at its core, a huge list of numbers (its "weights") plus some working space to process your request. All of that has to sit in the memory that's physically soldered onto the GPU — not your computer's regular system memory, and not anything the CPU can reach quickly. If the model's weights don't fit in GPU memory, the model does not run. Not "runs slowly" — it fails to load, full stop. That's why every build on this site is sized around GPU memory first, and everything else second.

CPU — important, but not the bottleneck here

The CPU handles the surrounding work: loading files from disk, running the web server that takes your request, scheduling jobs. For AI inference on a GPU, the CPU is mostly idle while the GPU does the actual math. A faster CPU will not let you run a bigger model — a GPU with more memory will. That's why you won't see CPU specs driving any recommendation on this site.

System RAM — a different pool of memory, for a different job

System RAM is the memory attached to the CPU, used for the operating system, the application, and staging data before it's copied to the GPU. It's typically far cheaper and more plentiful than GPU memory, but the model's weights don't live there while it's running — they have to be on the GPU. Confusing system RAM with GPU memory is the single most common reason people buy a machine that can't run the model they wanted.

Watts — what it costs to keep the thing running

Power draw, in watts, is how much electricity the hardware pulls at full load. It matters for two practical reasons: your electricity bill, and whether your building's power and cooling can actually support the machine. A 700 W GPU running continuously uses roughly as much power as an average American home does around the clock. Multiply that by racks of GPUs and power becomes a real planning constraint, not an afterthought.

Desktop card vs. datacenter GPU — not just a price difference

A desktop card like the RTX 5090 or RTX PRO 6000 Blackwell plugs into a regular PCIe slot, has its own cooling fans, and — critically — has no fast link to other GPUs in the same machine. If you put four of them in one box, each one's memory stays separate, and the software has to split the model between them by hand, shuttling data over the relatively slow PCIe bus. A datacenter GPU like the H100 SXM or H200 SXM is built to be wired directly to other GPUs over NVLink, a much faster interconnect, so multiple GPUs can share work — and in systems like the DGX B200 or GB300 NVL72, share memory — far more efficiently. Datacenter GPUs cost more per GPU, but they're built for exactly the multi-GPU workloads the largest models require.

The catalog

RTX 5090 desktop card

GPU memory 32 GB
The most you can hold on this single card. It's enough for smaller, quantized models, but nowhere near enough for any of the three models this store advises on — even the smallest, DeepSeek V4 Flash, needs 341 GB just to load. One RTX 5090 is a development or hobby card, not a production AI server on its own.
Price $4,200
The cheapest product we sell. It's the right choice for prototyping or running small models locally — not for the models featured on this site.
Power 575 W
Draws about half of what an average home uses continuously (575 W vs. 1,200 W). Runs fine on a standard household circuit. Run flat-out for a full day: 575 W × 24 ÷ 1,000 = 13.8 kWh ÷ 90 kWh (one EV battery) = 0.15x an EV battery's worth of energy per day.

RTX PRO 6000 Blackwell desktop card

GPU memory 96 GB
Three times the memory of the RTX 5090, in the same desktop form factor. This is the largest amount of memory you can get on a single card that still plugs into a normal workstation — which is exactly why four of them, at 384 GB combined, is our cheapest way to run DeepSeek V4 Flash (341 GB needed). See the "no NVLink" note below before you commit to that plan, though.
Price $13,250
About 3x the RTX 5090's price for 3x the memory — the memory cost per GB is nearly identical between the two. What you're really paying extra for elsewhere in this catalog is the NVLink interconnect on the datacenter parts, not raw memory.
Power 600 W
Just above an average home's continuous draw (1,200 W). Four of these in one build pull 2,400 W — about 2 homes' worth. Run flat-out for a full day: 600 W × 24 ÷ 1,000 = 14.4 kWh ÷ 90 kWh (one EV battery) = 0.16x an EV battery's worth of energy per day.

Important: this card has no NVLink. Multiple RTX PRO 6000 Blackwell cards in one machine talk to each other over PCIe, which is much slower than the NVLink links datacenter GPUs use. Fine for holding a model in memory; noticeably slower than datacenter hardware for the communication multi-GPU inference requires.

H100 SXM datacenter GPU

GPU memory 80 GB
The "SXM" in the name means this GPU doesn't plug into a normal PCIe slot at all — it's a module that slots into a purpose-built server chassis with a direct NVLink connection to the other GPUs in that chassis. 80 GB is enough to hold plenty of mid-size models, but on its own it's short of every model this store advises on.
Price $30,000
Roughly 2.3x the RTX PRO 6000 Blackwell's price for less than the memory (80 GB vs. 96 GB) — the premium here is the NVLink fabric and datacenter-grade reliability, not memory capacity.
Power 700 W
Just over half an average home's continuous draw. Datacenter GPUs like this one are always deployed in groups of 8 in a server, so in practice you're planning for multiples of this number, not one. Run flat-out for a full day: 700 W × 24 ÷ 1,000 = 16.8 kWh ÷ 90 kWh (one EV battery) = 0.19x an EV battery's worth of energy per day.

H200 SXM datacenter GPU

GPU memory 141 GB
The H200 is the H100's memory-upgraded successor — same NVLink class of GPU, 76% more memory per card (141 GB vs. 80 GB). That extra headroom per GPU means fewer cards needed to reach a given total, which is why it's the building block inside systems like the DGX B200 (8 GPUs of a similar Blackwell-generation part reaching 1,440 GB combined).
Price $38,000
About $270 per GB of memory — the most expensive per-GB price in the catalog among single GPUs, reflecting the newest memory technology and datacenter-grade NVLink hardware.
Power 700 W
Same draw as the H100 SXM for significantly more memory — more capability per watt. Run flat-out for a full day: 700 W × 24 ÷ 1,000 = 16.8 kWh ÷ 90 kWh (one EV battery) = 0.19x an EV battery's worth of energy per day.

DGX B200 complete system, 8 GPUs

GPU memory 1,440 GB combined
This is one machine containing 8 GPUs at 180 GB each, wired together with NVLink so software can treat them as one large pool of memory. 1,440 GB is enough, by itself, to hold GLM 5.2's full 893 GB requirement with plenty of room to spare — no hand-splitting the model across separate, disconnected cards required.
Price $515,000
About $358 per GB of combined memory — more per GB than a single H200 SXM card ($270/GB), because the price includes the chassis, the NVLink switch fabric, networking, and CPUs that turn 8 separate GPUs into one coherent system. You're paying for the integration, not just the memory chips.
Power 14,300 W
11.9x an average home's continuous draw (1,200 W) — running one of these full-time uses about as much electricity as 11.9 houses. Run flat-out for a full day: 14,300 W × 24 ÷ 1,000 = 343.2 kWh ÷ 90 kWh (one EV battery) = 3.81x an EV battery's worth of energy per day.

GB300 NVL72 rack complete system, 72 GPUs

GPU memory 20,736 GB combined
A full rack: 72 GPUs joined by an NVLink switch fabric so the entire rack behaves like one enormous GPU with shared memory. This is more memory than any model this store advises on needs — even DeepSeek V4 Pro's 1,920 GB requirement uses less than 10% of it. It's the right choice when you need serious headroom for growth, multiple models, or many simultaneous users, not just to fit one model.
Price $3,900,000
About $188 per GB of combined memory — the cheapest per-GB price in the whole catalog, because you're buying at rack scale. The catch is the entry price: $3.9M is a lot to spend if you only need a fraction of the capacity.
Power 132,000 W
110x an average home's continuous draw — running this rack full-time uses about as much electricity as 110 houses. Run flat-out for a full day: 132,000 W × 24 ÷ 1,000 = 3,168 kWh ÷ 90 kWh (one EV battery) = 35.2x an EV battery's worth of energy per day.

Not sure which of these adds up to what you need? See the memory math for each model we advise on, or skip straight to Help Me Choose for a direct recommendation.