Compute Prices what a token and a GPU-hour cost, indexed daily

The DC Hardware index prices the inputs to a data center. This page prices what comes out of one: the per-token list prices of six frontier and open models on OpenRouter, and the rental price of a GPU-hour on vast.ai and sfcompute. Both composites are chain-linked equal-weight geometric means: each day's move averages the day-over-day price changes of the members priced on both days, so a SKU missing a day, joining, or retiring changes who is averaged but never jumps the index — and a deprecated model drops out instead of freezing a dead price into it. Collection began 2026-06-16: the history is short and says so.

Token price index
104.3
2026-07-15 = 100 · 30d 0.0% · as of 2026-10-02
GPU-hour index
94.0
2026-07-15 = 100 · 30d +9.2% · as of 2026-10-02
History
107d
daily since 2026-06-16 · 6 models · 7 GPU SKUs
Cite
CiteMacroGauge Token price index, 2026-10-02, 2026-07-15=100, 104.3 — https://macrogauge.vercel.app/compute

Models — list price per million tokens (OpenRouter)

ModelInputOutputBlended30dAs of90d
GPT-4o$2.50$10.00$4.3750.0%2026-10-02
Claude Sonnet$2.00$10.00$4.0000.0%2026-10-02
Llama 3.1 70B$0.40$0.40$0.4000.0%2026-10-02
DeepSeek$0.26$1.03$0.4500.0%2026-10-02
Gemini Flash$1.50$9.00$3.3750.0%2026-10-02
Mistral Large$2.00$6.00$3.0000.0%2026-10-02

GPU rentals — $ per GPU-hour

SKU$/GPU-hr30dAs of90d
H100 SXM (vast.ai) vast_h100_sxm$1.602−11.1%2026-10-02
H200 (vast.ai) vast_h200$3.556−13.4%2026-09-28
B200 (vast.ai) vast_b200$8.125+35.4%2026-09-30
A100 SXM (vast.ai) vast_a100_sxm$0.669+25.4%2026-10-02
RTX 4090 (vast.ai) vast_rtx4090$0.482+32.4%2026-10-02
H100 (sfcompute spot) sfc_h100$2.060+5.6%2026-09-24
B300 (vast.ai) vast_b300 · not in index$8.719—2026-09-30—

vast.ai rows are the marketplace median for the SKU; sfcompute is its spot average. List prices, not negotiated rates — the same caveat the DC Hardware index carries for OEM inputs. Series are config (config/series.json); a stale series (7-day limit) shows on /status and leaves the mean. vast.ai medians cover the whole on-demand market: the API caps a query at 64 offers, so a saturated SKU is re-queried in price bands and the bands merged. Rows marked “not in index” (B300, first collected 2026-09-29) are priced but kept out of the GPU-hour index, whose base is the first day every member was priced.