Guided path · mid-size company
Build for headroom, not just today's traffic.
You have real, growing usage — more simultaneous users, and a product that can't afford to be slow or go down. The cheapest build that technically fits the model is the wrong answer here: you need spare memory for traffic spikes, room to run more than one model, and a machine that keeps serving customers if part of it fails.
Step 1 — which model are you building around?
Step 2 — the recommendation
1x NVIDIA GB300 NVL72 Rack
- DeepSeek V4 (Pro) needs 1,920 GB just to load. A startup's minimum build gets you exactly that much room — fine for one model, one modest audience, zero margin for error.
- DeepSeek V4 Pro is already the biggest of the three models, so a mid-size company should skip the minimum entirely. One GB300 rack gives 20 TB of shared memory — over 10x what the model needs — leaving room to also run smaller models, handle traffic spikes, and grow for years without buying another rack.
- Headroom, quantified: this build carries 20,480 GB against a 1,920 GB requirement — 10.7x the model's footprint, versus 1.2x for the bare-minimum startup build.
- Total price: $3,850,000. More upfront than the startup build — the trade you're buying is capacity you won't outgrow in a quarter.
- Total power: 135,000 W — about 112 homes running around the clock. Running continuously, that's a full EV battery's worth of energy (90 kWh) every about 40 minutes. At this scale, talk to your facility or colocation provider about power capacity before you talk about the invoice.