Anton's AI Hardware GPU
GB200 NVL72 — $2,750,000
A full liquid-cooled rack: 72 Blackwell GPUs and 36 Grace CPUs connected as one giant NVLink domain with roughly 13.8TB of combined GPU memory. This is what 'a cluster' means at the top end — one logical machine, not a pile of boxes.
From a desktop card to a full datacenter rack. Each machine gets its own breakdown — real specs, real price, and a plain-words explanation of what it's for.
The enthusiast desktop card
A top-end gaming card that doubles as a cheap way to experiment with small open-source models locally. 24GB of memory is enough for models up to roughly 15-18B parameters, or for fine-tuning small models. Not enough memory for any of the frontier open models on this site.
The newest desktop flagship
NVIDIA's newest Blackwell-based desktop card. 32GB of memory stretches to roughly 20-25B parameter models comfortably. Still a desktop card — no NVLink to other GPUs, so it cannot be meaningfully clustered for huge models.
The workhorse datacenter GPU
A previous-generation datacenter GPU still widely used for inference. 80GB of HBM2e memory and NVLink support mean multiple A100s can be pooled into real, high-bandwidth clusters for serious models.
The current-generation AI accelerator
The GPU most large AI labs actually train and serve on today. 80GB of HBM3 memory per card, with NVLink for true multi-GPU clustering. This is the building block behind the DGX systems below.
8 GPUs in one box
A single server holding 8 H100 GPUs wired together with NVLink, for a combined 640GB of GPU memory. This is real clustering: all 8 GPUs can address each other's memory at very high bandwidth, not just sit in the same box.
The most powerful AI hardware on Earth
A full liquid-cooled rack: 72 Blackwell GPUs and 36 Grace CPUs connected as one giant NVLink domain with roughly 13.8TB of combined GPU memory. This is what 'a cluster' means at the top end — one logical machine, not a pile of boxes.
Three real open-source models from the leaderboard. Here's exactly what it takes to run each one.
A 1.6 trillion parameter mixture-of-experts model — only 49B parameters are active per token, but every parameter still has to live in GPU memory.
Minimum setup: Startup Build — DeepSeek V4 Pro — 3× DGX H100 — $1,125,000 total
3 DGX H100 systems, networked together, give 1,920GB of pooled GPU memory — exactly enough headroom for DeepSeek V4 Pro's 1,920GB minimum.
A 753B parameter MoE model with 40B active parameters — a strong open model at roughly half the memory footprint of DeepSeek V4 Pro.
Minimum setup: Startup Build — GLM 5.2 Max — 2× DGX H100 — $750,000 total
2 DGX H100 systems give 1,280GB of pooled memory, comfortably above GLM 5.2 Max's 903.6GB minimum, at the lowest real unit count that clears the bar.
NVIDIA's 550B parameter MoE model, the smallest of the three but still far beyond what a single GPU — or even a single desktop card cluster — can hold.
Minimum setup: Startup Build — Nemotron 3 Ultra — 2× DGX H100 — $750,000 total
Nemotron 3 Ultra needs 660GB minimum — just over what a single DGX H100 (640GB) provides, so the cheapest safe build is 2 DGX H100 systems (1,280GB).
A cluster, in plain words: multiple GPUs wired together with a fast interconnect (NVLink) so they behave like one much larger GPU, instead of several small separate ones.
The honest part: a datacenter cluster (DGX systems, GB200 racks) uses purpose-built NVLink fabric so every GPU can read every other GPU's memory at hundreds of gigabytes per second. A pile of desktop cards (RTX 4090s or 5090s) has no such fabric — you can add up their memory on a spec sheet, but you cannot actually run one giant model split across them at usable speed. That's why every recommended build on this site below uses real datacenter hardware, never a stack of desktop cards.
| Build | Units | Combined Memory | Combined Power | Total Price |
|---|---|---|---|---|
| Startup Build — DeepSeek V4 Pro | 3× DGX H100 | 1,920 GB | 30,600 W | $1,125,000 |
| Startup Build — GLM 5.2 Max | 2× DGX H100 | 1,280 GB | 20,400 W | $750,000 |
| Startup Build — Nemotron 3 Ultra | 2× DGX H100 | 1,280 GB | 20,400 W | $750,000 |
| Mid-Size Growth Cluster — GB200 NVL72 | 1× GB200 NVL72 | 13,824 GB | 120,000 W | $2,750,000 |
Watts on a spec sheet don't mean much on their own. Here's what each build actually costs to run, translated into homes and electric-car batteries. Reference: 1 home ≈ 1,200 W around the clock · 1 EV battery ≈ 90 kWh.
30,600 W
That's as much power as 26 average homes running around the clock — or draining an EV battery every 2.94 hours.
20,400 W
That's as much power as 17 average homes running around the clock — or draining an EV battery every 4.41 hours.
20,400 W
That's as much power as 17 average homes running around the clock — or draining an EV battery every 4.41 hours.
120,000 W
That's as much power as 100 average homes running around the clock — or draining an EV battery every 0.75 hours.