Public prices / measured performance
GPU Market
Five charts on GPU rental prices, output throughput, compute cost and measured energy efficiency.
GPU rental prices · six-month history
Loading sourced observations…
Daily median rental prices are drawn as smooth, continuous lines, each a centered ±3-day moving average so short-term noise does not hide the trend; every GPU variant keeps its own line across the full six-month window. History combines live collection with dated Internet Archive snapshots of the same public provider pages. A gap between observations is a straight interpolation, not new price data; hovering any day shows its exact daily median. An observed price does not reserve capacity.
Sources & data
Current rental rates · USD / GPU-hour
Loading sourced observations…
Compare the median and published price range for current on-demand, single-GPU offers. PCIe, SXM and NVL variants remain separate because hardware and rental terms differ. These are public observations rather than executable quotes; availability, region, taxes and additional charges may change the price you ultimately pay.
Sources & data
Output throughput · tokens / second / GPU
Loading sourced observations…
Representative eight-GPU MLPerf systems show measured output throughput for the Llama 2 70B Server workload, divided by GPU count. Precision, benchmark round and software remain visible because they affect the result. This is a comparison of submitted systems, not a universal GPU ranking or a prediction for another workload.
Sources & data
Estimated compute cost · USD / million tokens
Loading sourced observations…
Measured eight-GPU throughput is translated into an estimated output-token cost using matching fresh single-GPU median rental rates. This is compute-only arithmetic, not a quoted eight-GPU configuration: networking, software, memory and utilization can change achievable performance. Systems without a matching rental observation are omitted rather than assigned an assumed price.
Sources & data
Measured efficiency · output tokens / second / watt
Loading sourced observations…
Only eight-GPU Server results with published whole-system power measurements appear here. Output throughput divided by measured watts is tokens per second per watt (the same quantity as tokens per joule), and it includes the measured system rather than GPU power alone. The available systems use different quantization, so these measurements should not be read as an otherwise identical hardware efficiency test.
Sources & data
Public prices are observations, not quotes or capacity reservations. Benchmark evidence comes from submitted MLPerf systems; each chart discloses its sources, measurement boundaries and missing data.