3.5 years of GPU rental prices, recovered from the archive

Our live pipeline started collecting in August 2026 — but list prices didn’t start existing then. The Internet Archive has been photographing provider pricing pages for years, so we built an ingester that parses Lambda’s archived pricing page across four separate page redesigns and both of its domains, monthly back to January 2023. 264 recovered observations, every one carrying its snapshot URL, run through the same validation bands as live data, and labeled archival. Months without a usable snapshot stay empty; rows that don’t say which form factor they were are skipped, not guessed.

What the recovered record shows:

The H100 got cheap fast. Lambda listed the H100 PCIe at $2.40/hr in May 2023 — months after launch — and cut it to $1.99 by July 2023. The H100 SXM ranged $2.59–$3.49 across the record as configurations shifted.

Prices go up, too. The 1× A100 40 GB listed at $1.10/hr in January 2023 and $1.29 later in the record — a 17% increase, from the era when A100 supply tightened as everyone retooled around the H100. Anyone modeling GPU prices as a one-way decline should look at this series.

And sometimes nothing happens at all. The RTX A6000 listed at $0.80/hr in January 2023 and $0.80/hr in the latest archived snapshot — three and a half years, zero movement, through the largest compute buildout in history.

Every chart on the site draws the archival region dashed with hollow points — visually distinct from live pipeline data, monthly resolution, fewer providers. The methodology page documents the rules; the raw series is free in the API (try /api/v1/history/h100-pcie, which now starts in May 2023). The one thing the archive can’t recover is marketplace pricing — offer books were never archived — which is exactly why the live pipeline exists.