Path A: Small startup
Run DeepSeek V4 Flash as cheaply as it can honestly be run.
Recommended build
4x RTX PRO 6000 Blackwell, running DeepSeek V4 Flash (284B parameters, needs 341 GB of GPU memory).
384 GB combined vs. 341 GB needed — comfortable headroom above the minimum, at 4 cards. Three cards (288 GB) would fall short of the requirement, so four is the smallest number that works.
What a "cluster" means here
Definition: a cluster is two or more separate machines or cards connected together and used as a single combined system.
This build is a cluster of 4 cards in one tower.
Be honest with yourself about this build: these RTX PRO 6000 Blackwell cards have no NVLink. They talk to each other over PCIe — the same general-purpose bus your desktop's other components use — which is much slower than a proper interconnect. Compare that to a GB300 NVL72 rack, where all 72 GPUs are wired by an NVLink switch fabric and the software sees one huge GPU with shared memory. This build is a pile of desktop-class cards, not a real cluster in that sense. It works, it is cheap, and it is slower at multi-GPU work than datacenter hardware. If that tradeoff is fine for your team's scale, this is a genuinely good, inexpensive way to run DeepSeek V4 Flash.
Power, in human terms
2,400 W is 2.0x an average home's continuous draw (1,200 W): 2,400 W ÷ 1,200 W = 2.0 homes.
Running this build at full load for 24 hours (one day) uses 2,400 W × 24 ÷ 1,000 = 57.6 kWh. An electric-car battery holds 90 kWh, so 57.6 kWh ÷ 90 kWh = 0.64x one EV battery's worth of energy, every day this build runs at full load.