A40

Value ranks each provider on price for performance. We take this GPU’s FP32 throughput (TFLOPs), divide it by the hourly price to get raw compute per dollar, and scale that across the providers shown, so the best-priced option scores 100 and the most expensive scores 0. The $/TFLOP column expresses the same trade-off as an absolute dollar cost per TFLOP-hour.

* Live availability isn’t published by every provider. Where a provider doesn’t expose per-GPU stock, we mark it with an asterisk.

Community GPUs come from independent operators rather than vetted data centres, which is what makes them cheap. They can be reclaimed or drop offline with little warning, and you should treat any data on them as exposed. They're fine for interruptible or non-sensitive runs. For anything sensitive or long-running, rent a secure GPU instead.