GPU Price Index methodology

Lantern Works GPU Price Index — methodology v1.0

What it is: a daily, per-GPU reference rental price for six NVIDIA GPUs, computed only from Lantern Works' own sourced price rows (each row has a vendor source URL, verbatim quote and check time in the feed).

Scope

Series	Feed `gpu` values accepted	VRAM
H100	H100, H100 SXM, H100 PCIe, H100 NVLink	80 GB
A100	A100, A100 SXM, A100 PCIe, A100 NVLink	80 GB (40 GB rows are excluded, never mixed)
H200	H200, H200 SXM	141 GB
L40S	L40S	48 GB
RTX 4090	RTX 4090	24 GB
RTX 5090	RTX 5090	32 GB

Form factors (SXM/PCIe/NVLink) are pooled in v1.0 for coverage; per-form-factor series are a v1.1 candidate once each has ≥3 vendors.

Which rows count

A snapshot row is included only if all are true:

status == "ranked" (held/stale rows never count).
Tier is on-demand class (on-demand, secure or community cloud, per-hour instances, including 8-GPU instances priced per GPU).
Not spot/interruptible/preemptible/bid/batch/serverless-preemptible (tier text matches spot|interruptible|preempt|bid|batch|serverless).
Not reserved/committed/monthly-dedicated (reserv|month|dedicated|commit|N mo).
Marketplace/aggregator rows count only if the vendor has ≥5 listings for that GPU, read from offer_count in api/v1/gpu-prices.json at the last commit on that date. Unknown count → excluded (not assumed).

Statistic

Per vendor, take its lowest qualifying per-GPU-hour rate (hourly_usd), so one vendor = one vote.
Index value = median across vendors, USD per GPU-hour, 4 decimals.
Also published: minimum, vendor count, low_coverage (fewer than 3 vendors; don't headline these).
Monthly basis = median × 730 hours (same as the feed's hours_per_month). It's a basis for comparison, not a quote.

Price change vs. composition change

The headline median moves when vendors are added. Each day after the first also carries a matched-vendor change: the median over vendors present on both days, plus lists of vendors added and dropped. Only matched changes may be described as "prices moved". A newly tracked cheap provider is a coverage change, not a market price cut.

History

Source: data/gpu_pricing/snapshots/YYYY-MM-DD.json (one per day, starting 2026-10-27). Each output day records the

snapshot file's SHA-256 and the feed commit used for listing counts.

Coverage grew a lot from Oct 27 to Oct 29 (e.g. H100 went from 4 to 9 vendors), so early movement is mostly composition.

Don't publish trend claims before 30 days of history (per T21), and show history length next to any chart.

Reproduce

git clone https://github.com/connormatthewdouglas/agent-work-stack-deals repo
python3 repo/scripts/gpu_index.py repo out   # writes out/gpu-price-index.json (audit trail: included/excluded row keys) + .csv

No price is entered by hand. Every value traces to a snapshot row with its source_url and verified_at.

License and attribution (CC BY 4.0)

Data: CC BY 4.0, https://creativecommons.org/licenses/by/4.0/ . Code: MIT.

Attribution text (use as-is):

Source: Lantern Works GPU Price Index (https://thelanternworks.com), licensed CC BY 4.0. Median on-demand rental price per GPU-hour across tracked vendors; methodology at https://thelanternworks.com/gpu/index/methodology/.

Short form: Data: Lantern Works GPU Price Index, CC BY 4.0.

Back to the index