Hardware price-performance growth (doubling time)

Doubling time of FP16 tensor throughput per release-price dollar across accelerators with a public price, fitted log-linearly from 2012 on Epoch's hardware table. Derived metric perf_per_dollar_doubling_days.

Why it matters. The floor under falling inference prices; the brief's reading is +49% a year.

Proxy types

price, model_release

Unit

days

Cadence

annual

Valve

invention to product

• derived metric perf_per_dollar_doubling_days (formula in the semantic layer)

Only accelerators with a public list price enter, and recent data-centre GPUs have none, so the series can end years before today.

20 observations. Hollow points are disputed (see counterevidence). Every point links to its observation.

Evaluator: 1282 (Epoch AI ML hardware, 2025-11-06) is inside the consistent band (lo=500 ). Auto-reason; band rationale: Normal = doubling no faster than every 500 days (roughly the +49% a year Epoch reports); fast = yearly or faster. Between is emerging.

The tracker's prior expectation was emerging; the evaluator reads consistent with normal. The evaluator wins until a reviewed override.

• 2025-11-06 · 1282 daysas of 2025-11-06(20 obs)

1. 2026-09-10○unmeasured →to ●consistent with normalconf — → 40 · evaluate

   Evaluator: 1282 (Epoch AI ML hardware, 2025-11-06) is inside the consistent band (lo=500 ). Auto-reason; band rationale: Normal = doubling no faster than every 500 days (roughly the +49% a year Epoch reports); fast = yearly or faster. Between is emerging.

   obs:0aa9e9dfobs:15c9bef1obs:2b8b94cbobs:2e743762obs:345c670eobs:4974aa55+14 more

40 / 95 — limited or vague evidence

Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; below 50 limited or vague.

• Inference price at fixed capability (halving time) ◐emerging
• Rent concentration in semiconductors ▲concentrating
• Crosswalk: Methods ⇄ Compute & physical (input to)