GPU Pricing in September 2026: What the Current Catalog Can Tell You

GPU prices are easier to misinterpret than they first appear. A lower number can reflect a different memory configuration, a spot offer, a cheaper region or a larger minimum instance.

This September 26, 2026 review replaces earlier unsupported claims about price declines, used-hardware bargains and supply conditions. We have a current catalog snapshot, not a controlled historical price index. That distinction limits what we can conclude.

What the September snapshot shows

In our nine-provider catalog, selected lowest listed on-demand rates are $0.340 per GPU-hour for RTX 4090, $1.390 for A100 80GB, $2.890 for H100 80GB, $4.500 for H200 and $6.690 for B200.

These are different products with different memory and configurations. They are not comparable measures of work completed. The B200 example requires an eight-GPU Lambda instance billed at $53.520 per hour; the other examples listed here use one GPU.

The relevant provider source versions are dated September 26. Google Cloud's source snapshot remains September 24. Read the detailed dated price table or the current price board for context.

More offers do not mean more providers

This month's catalog update changed the way NVGPU retains regional and memory configurations. It also moved collection from the older combined feed to individually versioned provider feeds.

The resulting increase in displayed offers does not, by itself, establish growth in GPU supply or a price reduction. A change in data collection can change the number of rows without changing the market.

A similar problem appears when a provider is renamed or when missing data returns after a failed collection. Count the underlying service consistently rather than treating every label as a new competitor. Our DataCrunch key consumes the Verda feed to avoid counting the renamed service twice.

What a defensible trend analysis needs

To measure a price trend, compare the same provider, GPU model, VRAM, instance size, region and pricing type at multiple dates. Record the source publication date separately from the date you downloaded it.

Keep discontinued or missing offers distinguishable from unchanged prices. If a provider's feed fails, carrying its old value forward should be visible rather than interpreted as a new observation.

For a useful historical dataset, retain:

• The provider and original offer or instance identifier.
• GPU model, quantity and reported memory.
• Region, currency, billing unit and spot status.
• Whole-instance price and normalized per-GPU price.
• Collection time, source version and any known availability limitations.

Without those controls, a falling average can simply mean more small GPUs entered the dataset.

Compare job economics, not just generations

A newer GPU may cost more per hour but finish a task sooner. A larger-memory GPU may avoid offloading or let you use a different batch size. Neither outcome is guaranteed by a model name.

Run the same workload and report total runtime, throughput, memory use and final quality. Include setup and idle time when calculating the bill. Use H100, H200 and B200 pricing as inputs to that calculation, not as benchmark rankings.

What about buying hardware?

This article does not publish a September retail-price index. A useful buying comparison needs current seller quotes, taxes, warranty, delivery and the complete host-system cost.

Compare that ownership cost with realistic utilization. Dividing a purchase price by the cheapest advertised cloud rate overlooks maintenance, electricity, resale uncertainty and the fact that the cheapest cloud configuration may not be equivalent.

What to do now

Use the current catalog to choose candidates, then verify capacity and rates at the provider. Save the quote and job results if you want to compare them later. Avoid inferring a market-wide direction from a single snapshot.

Our collection approach follows the independently versioned feeds documented by gpuhunt. See NVGPU's methodology for coverage limits and the affiliate disclosure for how commercial links are handled.