Compute Futures — Figure

The cost of intelligence, in Qi

Qi is defined as hashpower × time, so its “per watt” price is just the inverse energy peg. One Qi is a fixed quantity of running compute — the only question is how much thinking that compute buys.

Cost of intelligence — power basis
0.0020 Qi per watt-day
Equivalently, one Qi powers 500 watt-days of running compute.
What one Qi becomes
1 Qi
Money
the energy-pegged unit
500 W·days
Power
= 12.0 kWh of electricity
121 EFLOP
Computation
raw floating-point work
43 M tokens
Thinking
generated reasoning
η  mining efficiency · W per MH/s
5.0
Network-marginal, set by the most efficient miner — not your rig. Tuned RTX 3060 Ti ≈ 5; industrial/ASIC < 3.
ε  compute intensity · TFLOP per joule
2.8
Dense FP16 frontier accelerator ≈ 2.8 (H100-class); newer parts higher.
J/token  energy per output token
1.0
Served inference, wide spread: ~0.3–3 J/token with model size and batching.
1 Qi = 100·η W × 24h = 2.4·η kWh · cost/W·day = 1/(100·η) Qi · FLOP/Qi = ε × 8.64×10⁶·η
The model in one line: the power-basis price is fixed by the peg; the real variable is how far down the ladder a joule travels. As ε climbs faster than η falls, the same Qi buys more thinking — intelligence deflates against energy money.

Caveats: 100 MH/s·day is the protocol reference; live conversion runs through the K-Quai controller and post-BIP-302 SOAP affects realized QUAI↔Qi, not the Qi peg. The J/token input is the soft link — swap in measured values.