Public benchmarks

Numbers in the pricing guides link back to this page. Format:

[<label>](./benchmarks#<slug>)

Cost-floor unit: $/compute-hour. Ask price unit: $/1,000 tokens.

Accessed 2026-08-22 unless noted.

Power

EIA industrial electricity

• Source: U.S. Energy Information Administration, Electric Power Monthly
• Figure: U.S. industrial average $0.082 / kWh (2024 annual)
• Scope: national industrial average; coastal ISO peaks run higher (use $0.16–$0.22 / kWh for CAISO-like profiles)

Hydro / surplus industrial

• Source: EIA average retail price by state
• Figure: surplus-hydro industrial bands around $0.045–$0.060 / kWh
• Scope: Pacific Northwest / Quebec-class industrial contracts, not residential

Hardware amortization

A100 80GB street price

• Source: ServeTheHome GPU pricing coverage and public secondary listings
• Figure: $10,000–$15,000 per A100 80GB SXM, mid $12,500
• Scope: 2024–2025 secondary market; new OEM list is higher

H100 SXM street price

• Source: public cloud residual and secondary listings summarized by SemiAnalysis
• Figure: $25,000–$32,000 per H100 80GB SXM, mid $28,000
• Scope: excludes NVLink fabric and host

Useful life

• Source: hyperscaler GPU refresh commentary in SemiAnalysis
• Figure: 36 months useful life, 24 hours/day, ~26,280 compute-hours
• Scope: production inference boxes, not lab cards

Spare capacity

Marketplace off-peak discount

• Source: public Vast.ai and RunPod spot vs on-demand spreads
• Figure: off-peak / interruptible asks clear 30–50% below 24/7 on-demand
• Scope: observed public list prices, not this gateway's book

Overhead

AWS inter-region egress

• Source: AWS EC2 on-demand data transfer
• Figure: $0.09 / GB internet egress in the lower usage tier
• Scope: AWS public list, us-east-1, 2026-08-22

Datacenter PUE

• Source: Uptime Institute Global Data Center Survey
• Figure: average PUE ~1.4; best-in-class ~1.2
• Scope: colocation / enterprise, not a single site

Staffing and maintenance load

• Source: operator reports cited in Uptime Institute survey notes
• Figure: 8–15% of hardware + power as annual opex for a one- to four-box node
• Scope: small operator, not a 1,000-GPU hall

Margin ranges

Public GPU marketplace asks

• Source: RunPod pricing, Vast.ai, Lambda on-demand
• Figure: A100-80GB on-demand roughly $1.10–$1.99 / GPU-hour; H100 $2.00–$3.00 / GPU-hour
• Scope: public list at access date; does not include Spot book prints

Target margin

• Source: implied by the spread between the cost-floor methodology on this site and those public asks
• Figure: 15–35% above fully loaded cost floor for a resting ask that still clears
• Scope: guidance range, not a platform rule