Datacenter Math

Three small calculators that follow a rack from the grid to the price of a token: rack power and line current, KV cache concurrency, and dollars per million tokens. The formulas and default numbers come from the From Watts to Tokens course. Every number names its denominator. Back to rahilbhavan.com.

Rack to grid

Rack IT power
kW per rack
Line current, AC three-phase
A per line
Conductor current, DC
A per conductor
Facility power
MW at the meter
Facility energy
MWh per day
Water intake
L per day

Source: From Watts to Tokens, draft updated 2026-09-10. 100 kW at 480 V AC draws about 120 A per line; at 800 V DC, 125 A per conductor. 100 MWh IT per day at 1.25 L/kWh is 125,000 L per day.

KV cache and concurrency

KV cache size
KiB per token
Resident requests
full-context requests in the pool

Source: From Watts to Tokens, draft updated 2026-09-10. Llama 3.1 70B in BF16 is 320 KiB per token; a 64 GiB pool holds 25 requests at 8K context and 6 at 32K. Excludes prefix sharing, block rounding and weights.

$/token

Token cost
$ per million tokens

Source: From Watts to Tokens, draft updated 2026-09-10. Rental ranges are 25th to 75th percentile one-year terms per GPU-hour, dated as labelled. Compute cost only; excludes occupancy, networking and margin.