Help Me Choose the Right Build
Two guided paths — one for a startup on a tight budget, one for a company that needs real capacity. Each ends with a complete recommended build, total price, and honest power numbers.
Startup Minimum Build
Run one model, at the lowest honest cost
GLM 5.1 is the smallest of the three models at 754 billion parameters, which means it needs the least GPU memory and costs the least to run. It still needs 905 GB of GPU memory — far beyond any single card or any desktop setup.
What This Build Includes
| Component | Qty | Unit Price | Subtotal |
|---|---|---|---|
| NVIDIA H100 PCIe 80GB | 12 | $27,500 | $330,000 |
| Dual-socket server node (2 nodes, 6 GPUs each) | 2 | $25,000 | $50,000 |
| Total Build Price | $380,000 | ||
Combined Build Specifications
| Total GPU memory | 960 GB |
| Total power draw | 4,800 W |
- 4.0× an average home — an average American home draws about 1,200 W continuously. This build draws as much power as 4.0 of those homes running at the same time.
- Drains one EV battery every 18.8 hours — a typical electric car battery holds 90 kWh. Running this build for 18.8 hours uses the same amount of energy.
- 1.3 EV batteries per day — in 24 hours of continuous operation.
What a cluster is: A cluster is two or more computers connected by a fast network and working together as if they were one machine — the software sees a single pool of memory and compute spread across both nodes.
Cluster vs. stacking desktop cards: These server nodes use dedicated InfiniBand networking so they can share GPU memory across machines, letting a 905 GB model span both nodes seamlessly. Desktop RTX cards — even a roomful of them — cannot share their video memory at all: each card's 24 GB is completely isolated from every other card, and no software can bridge them into a single pool.
Mid-Size Company Build
Capacity for the largest model, with room to grow
DeepSeek V4 Pro is the most capable of the three at 1.6 trillion parameters. Three DGX H100 servers provide exactly the 1,920 GB minimum required — and you can add a fourth server when demand grows.
What This Build Includes
| Component | Qty | Unit Price | Subtotal |
|---|---|---|---|
| NVIDIA DGX H100 Server | 3 | $350,000 | $1,050,000 |
| Total Build Price | $1,050,000 | ||
Combined Build Specifications
| Total GPU memory | 1,920 GB |
| Total power draw | 30,600 W |
- 25.5× an average home — an average American home draws about 1,200 W continuously. This build draws as much power as 25.5 of those homes running at the same time.
- Drains one EV battery every 2.9 hours — a typical electric car battery holds 90 kWh. Running this build for 2.9 hours uses the same amount of energy.
- 8.2 EV batteries per day — in 24 hours of continuous operation.
What a cluster is: A cluster is two or more computers connected by a fast network and working together as if they were one machine — the software sees a single pool of memory and compute spread across all nodes.
Cluster vs. stacking desktop cards: Inside each DGX H100, the eight GPUs share memory via NVLink at 900 GB/s — so fast they act like one chip. Between the three DGX servers, InfiniBand connects them at up to 400 Gb/s — proper datacenter clustering, engineered and tested end-to-end by NVIDIA. Stacking desktop RTX cards gives no inter-GPU connection at all: each one's 24 GB stays isolated, and a 1,920 GB model cannot be distributed across them by any means.
Not sure which path fits you?
Fill out a quote request and describe your situation in the "build" field — we'll match you to the right configuration.
Open the Quote Form