What it costs to rent a GPU.

It is the typical price of renting one GPU on demand, where a listing with 8 GPUs counts 8 times and a single-GPU listing counts once, and community hosts (individuals renting out their own machines) are left out whenever at least three offers from datacenter or other vetted hosts exist. Spot and bid prices, which can be interrupted, are never included. "Cheapest" is the single lowest offer of any kind. "Offers" counts listings available to rent; "GPUs listed" adds up their GPUs.

Full rule: Each offer is counted once per 5-minute snapshot (the latest capture of each listing wins) and placed in a like-for-like segment by GPU model, variant (SXM, PCIe, NVL...), GPUs per node, host tier and pricing type; within a segment every price is weighted by the number of GPUs in the listing, so an 8-GPU node counts 8 times and a 1-GPU listing once, and p25/median/p75 are those capacity-weighted quantiles. The headline price for a GPU model is the capacity-weighted median over all of its on-demand offers (every variant and node size), leaving out community hosts whenever at least 3 non-community offers exist; spot and interruptible bids are never in the headline. Run node tape.mjs index --data data for every like-for-like segment.

Over the 247 snapshots of the last 24 hours, 2,072 machines lost capacity and stayed down for at least 30 minutes, about 4,715 GPUs in total. Another 1,905 came back within 30 minutes; that is mostly listings flickering or very short rentals, so they are shown but not counted. "Take price" is the typical (GPU-weighted middle) price those GPUs were listed at just before they went; compare it with today's asking index to see whether the cheap GPUs or the expensive ones are going. "Still watching" means the drop is too recent to call either way.

Caveat: these are estimates inferred from listings disappearing, not observed transactions; a host that takes a machine offline for maintenance looks the same as a rental. An inferred rental is counted when a machine's rentable GPUs (its largest bundle, counting separate same-size bundles together) drop and stay down for two consecutive successful snapshots of that source; one-snapshot blips, a host re-cutting its bundles into new listings of the same total size, source outages, missing snapshots and marketplace result lists that were truncated at the 64-offer cap are never counted. The GPUs taken is the size of the drop; the price is the capacity-weighted median of the machine's listings when it was last seen. A take whose capacity comes back within 30 minutes of observations is labelled brief (listing churn or a very short rental); one still down after 30 minutes is sustained, and only sustained takes are headlined. These are estimates inferred from listings, not observed transactions. Left out because the marketplace truncated the list: A100 SXM4|A100 PCIE|A100X; B200|B300|H200|H200 NVL|GH200 SXM; L40S|L40|RTX 6000Ada|RTX A6000; RTX 4090; RTX 5090; RTX PRO 6000 WS|RTX PRO 6000 S|RTX PRO 6000 Max-Q.