Hardware Catalog
Every number below comes with a plain-language explanation. Prices are approximate 2026 street prices — ask us for a live quote before ordering.
Consumer / Hobbyist
Consumer / Hobbyist
NVIDIA GeForce RTX 4090
Memory
24 GB GDDR6X
This is the card's own dedicated memory. Think of it as the desk space the AI model gets to spread its notes out on while it "thinks" — the bigger the desk, the bigger the model that can fit.
Power draw
450 W
Under full load this card pulls about as much electricity as four microwave ovens running at the same time — but only while it's actually crunching numbers, not when it's idle.
Price
$2,500
About the cost of a good laptop. This is the cheapest way to get real GPU horsepower for local AI experiments.
Best for: Learning, prototyping, running small-to-mid sized open models at home.
Workstation / Prosumer
Workstation / Prosumer
NVIDIA RTX 6000 Ada Generation
Memory
48 GB GDDR6 (ECC)
Double the desk space of the 4090, with error-correction (ECC) — the memory double-checks its own math, which matters when a business is relying on the answer.
Power draw
300 W
Actually draws less than the "smaller" 4090 despite having twice the memory — it's built for efficiency, not gaming, so it does more useful work per watt.
Price
$6,800
Roughly the price of a used sedan. Common choice for a single serious workstation used by a small team.
Best for: Small business workstation, fine-tuning small models, running one mid-sized model reliably.
Datacenter Inference
Datacenter Inference
NVIDIA L40S
Memory
48 GB GDDR6 (ECC)
Same memory size as the RTX 6000 Ada, but built to live in a rack and run 24/7 in a server room instead of under a desk.
Power draw
350 W
Roughly what a large flat-screen TV plus a hair dryer draw combined, sustained continuously, all day, every day, in a data center.
Price
$11,000
Priced for businesses running AI as a product, not a hobby — includes data-center-grade reliability and support.
Best for: Serving AI chatbots/APIs to real customers around the clock.
Datacenter Flagship
Datacenter Flagship
NVIDIA H100 SXM5
Memory
80 GB HBM3
HBM3 is a much faster type of memory stacked directly on top of the chip, like a pantry built into the kitchen counter instead of down the hall. This is what lets it feed a giant model fast enough to answer in real time.
Power draw
700 W
About 10x a high-end gaming PC's power draw, in one chip. This is the card most large AI companies use to train and run frontier models.
Price
$28,000
Roughly the price of a new car, per chip. Large builds use dozens or hundreds of these.
Best for: Running large frontier open models like GLM 5.2 Max, DeepSeek V4 Pro, or Kimi K2.7 Code at real speed.
Datacenter Flagship
NVIDIA H200 SXM
Memory
141 GB HBM3e
Nearly double the H100's memory in the same size chip — like renovating the same kitchen to fit almost twice as much pantry space.
Power draw
700 W
Same power draw as the H100 despite the extra memory — you get more capacity for the same electricity bill.
Price
$35,000
A premium over the H100 that buys you fewer chips needed per model, which can offset the higher per-chip cost.
Best for: Fitting bigger chunks of a huge model onto fewer physical GPUs.
Datacenter Flagship
AMD Instinct MI300X
Memory
192 GB HBM3
The largest single-chip memory on this list — over double an H100. Fewer chips are needed to hold the same model.
Power draw
750 W
Slightly more than an H100 or H200, roughly like adding one more household hair dryer to the load.
Price
$18,000
AMD's answer to NVIDIA's H100/H200 — typically priced lower per chip, at the cost of a less mature software ecosystem.
Best for: Cost-conscious large-memory deployments for teams willing to work with AMD's ROCm software stack.
Server Infrastructure
Server Infrastructure
8-GPU Rack Server Chassis (dual EPYC, NVLink backplane)
Memory
2 TB system RAM (typical config)
This is separate from the GPUs' own memory — it's the "regular" RAM the server uses to manage the operating system, network traffic, and juggle requests.
Power draw
1600 W
This is the chassis, CPUs, fans, and RAM only — NOT the GPUs, which are priced and powered separately above. Add GPU power on top.
Price
$45,000
The "empty shell" cost before GPUs are installed — think of it as the cost of the building, before you move the expensive equipment in.
Best for: The physical box that holds 8 datacenter GPUs, wired together internally with NVLink for fast GPU-to-GPU communication.
Server Infrastructure
InfiniBand NDR Switch (400 Gb/s per port, networking for real clusters)
Memory
N/A — networking equipment
This device doesn't hold model weights — it's the high-speed "hallway" that lets multiple 8-GPU servers talk to each other fast enough to act like one giant machine.
Power draw
750 W
Comparable to one datacenter GPU, but it runs continuously to keep every server in the cluster connected.
Price
$95,000
This is what actually makes a "cluster" real instead of marketing talk — see our Cluster page for why this matters.
Best for: Connecting multiple GPU servers into one true, fast, coordinated cluster.
CPU / High-RAM Server
CPU / High-RAM Server
Dual AMD EPYC 9005 Server, up to 6 TB DDR5 ECC RAM
Memory
Configurable, 512 GB – 6 TB system RAM
Regular system memory (RAM) instead of GPU memory (VRAM). RAM is far cheaper per gigabyte than GPU memory, so it lets you hold a huge model in memory for a fraction of the GPU cost — the tradeoff is speed (see below).
Power draw
900 W
About as much as a home electric oven running continuously — far less than an equivalent bank of GPUs.
Price
$22,000
Price shown is for a 1 TB RAM configuration. Roughly 15–25x cheaper than the GPU memory needed to hold the same amount of model data.
Best for: Running huge models slowly-but-cheaply for low-traffic internal tools, where GPU-speed responses aren't required.
Not sure what to pick? Try the Startup Guide or Mid-Size Guide for a full recommended build.