Analysis from the data
Everything here is computed from the same open dataset the site serves — no takes without numbers, and every number traces to a source or the API.
The same GPU costs up to 2× more depending on where you rent it
H100 PCIe: $2.00 on one provider, $3.29 on another — today, same silicon. Where the spread comes from, why it persists, and how to use it.
GPU prices moved 7% in one evening. Daily charts would never show it.
The night our live index printed 106.7 against a 100.0 close — what hourly marketplace data reveals that daily medians smooth away.
Where a GPU-hour's price actually goes
Electricity is 5% of an H100-hour and 11% of an RTX 4090-hour. The other ~90% is the machine itself. The arithmetic, from verified public inputs.
3.5 years of GPU rental prices, recovered from the archive
We reconstructed Lambda's list prices back to January 2023 from dated Internet Archive snapshots: the H100's cut to $1.99, an A100 price that went UP, and an A6000 that hasn't moved in three and a half years.
Every lab prices output tokens at 5–6× input. Why?
Across Anthropic, OpenAI, and Google, output-to-input price ratios cluster tightly at 5–6×. What the pattern means, who deviates, and why a scatter plot of token prices looks like a straight line.