What AI teams actually pay for GPUs.
Provider pricing pages tell you the sticker rate. They don't tell you what teams end up paying once idle pods, storage that never got detached, and the wrong card for the job are counted in. The CoreForge GPU Spend Benchmark measures the real number, from real invoices, and publishes it in full.
Optional summaries contribute anonymous figures.
When someone runs the free audit and requests an email summary, the figures from it — totals, the compute and storage split, GPU-hours, estimated overspend and the mix of GPU classes in use — join the benchmark. No name, no company, no invoice text, no workload names, and no link back to the person. The contribution is structurally anonymous, not anonymous by promise. How that works.
Live aggregates from anonymised contributions, updated as they arrive.
- The real effective $/GPU-hour teams pay, by card, versus the published rate.
- How much of a typical bill is compute versus storage and transfer.
- What share of spend shows the signature of idle or forgotten pods.
- Which GPUs are over-rented for the jobs actually running on them.
Want it when it lands? Run the free audit and tick the box — you'll contribute to it and get it first.