An 80GB SXM H100 is the same silicon wherever it racks. Yet across 21 providers quoting it, the same accelerator rents for anywhere between $1.20 and $22.02 per GPU-hour: an 18.3x spread on an identical part. And the H100 is one of the tighter markets. The A100 spans 96.5x across 15 providers. The RTX 4090 spans 100x.
Those numbers come from Gerra's research on the GPU rental market, built on the cross-provider panel behind the GPU Price Index. This post explains what the panel measures, why dispersion this wide persists in a market for a homogeneous good, and what the data is actually for.
What the panel measures
The research panel holds 6,216 price observations across 38 cloud providers and 32 GPU SKUs, spanning May 2022 to May 2026, collected by polling each provider's public pricing surface (hourly where available) and backfilled from archived snapshots. The productized feed currently tracks 25+ providers across 22 canonical SKUs at hourly resolution.
The load-bearing step is normalization. Providers describe the same hardware in incompatible ways, so the many marketplace spellings of an 80GB SXM H100 collapse to one canonical identifier, every quote is converted to USD per single GPU-hour, and each observation keeps its price_type (on-demand versus spot or marketplace terms), region, provider, and a provenance flag distinguishing live polls from archived snapshots. Without that step, cross-provider comparison is not measurement; it is anecdote.
How wide the spreads actually are
Taking the latest observed price per provider for each SKU listed by at least three providers:
| SKU | Providers | Cheapest | Median | Priciest | Spread |
|---|---|---|---|---|---|
| H100 SXM 80GB | 21 | $1.20 | $2.05 | $22.02 | 18.3x |
| B200 SXM 192GB | 17 | $2.50 | $3.50 | $62.60 | 25.0x |
| H200 SXM 141GB | 16 | $1.45 | $2.30 | $13.25 | 9.1x |
| A100 SXM 80GB | 15 | $0.10 | $1.22 | $9.59 | 96.5x |
| RTX 5090 32GB | 13 | $0.25 | $0.50 | $1.20 | 4.8x |
| RTX 4090 24GB | 10 | $0.16 | $0.31 | $16.00 | 100x |
| L40S 48GB | 9 | $0.60 | $0.87 | $12.31 | 20.5x |
Why the dispersion persists
Several mechanisms, none exotic, compound into these spreads.
There is no central exchange. Compute has no consolidated tape and no clearing price. A buyer who wants the market price of an H100 must poll dozens of pricing surfaces that do not even agree on what to call the product. Search costs alone sustain a wide band.
Quotes mix contract forms. A posted price may be on-demand, a committed longer term, or interruptible spot capacity. Spot is structurally cheaper because it is reclaimable: on AWS, spot instances are spare capacity sold at a discount that fluctuates with long-term supply and demand, subject to reclamation on a two-minute notice (AWS EC2 documentation). A $1.20 interruptible hour and a $22.02 on-demand hour with enterprise support are different products wearing the same SKU.
Reliability is priced in, unevenly. Providers differ widely in uptime, support, and operational maturity. Part of a premium quote is paying for the cluster to still be there tomorrow; part of a marketplace discount is bearing the risk that it is not.
Interconnect and cluster context differ. A lone consumer card on a marketplace and eight SXM GPUs on a unified fabric quote the same per-GPU hour while offering very different machines for real training workloads. The panel preserves the raw instance description alongside the canonical SKU precisely so this can be controlled for.
Regional supply varies. Capacity is unevenly distributed across regions, and prices follow local tightness.
Capacity cycles move the tails. The feed's capacity layer snapshots total, available, and rented instances per SKU, a direct read on supply tightness, and the panel's composition shifts as providers enter and exit. The research treats this explicitly: trends are read within a fixed provider set, because a median that ticks up when new, pricier providers join the panel is an artifact, not a price move.
The cost curve underneath
Under the dispersion, the direction of travel is down. On the consumer segment, where the panel holds a continuous multi-year history from a stable provider set, the RTX 3090 fell from $0.150 to $0.105 per hour between July 2024 and March 2026 (-30%) and the RTX 4090 from $0.296 to $0.240 (-19%).
That is the rental-market expression of a much older hardware trend: Epoch AI's analysis of 470 GPU models from 2006 to 2021 finds FLOP/s per dollar doubling roughly every 2.5 years. Depreciating hardware plus growing supply of last-generation accelerators shows up as steadily cheaper rentals, with datacenter SKUs still too young in the panel to trend with confidence.
The equity signal is a null, and it is published
The tempting pitch for this dataset would be alpha: GPU prices as a leading indicator you can trade against NVDA, AMD, AVGO, TSM, ASML, and the rest of the AI-infrastructure complex. The research tested exactly that, and reports the result plainly: it does not survive honest validation.
Walk-forward models on price and dispersion features looked strong under a naive train/test split. Under purged k-fold cross-validation with an embargo, tested against a label-shuffled null and a naive-momentum baseline, the apparent edge disappears, with a shuffle-test p-value near 1.0. The paper's own assessment is that the naive backtest was fitting a sustained 2024-2026 AI uptrend, not extracting predictive content, and its verdict on the equity signal is not_ready. A tradeable claim would require twelve-plus months of forward-collected datacenter-SKU history, at least one non-trending regime, and a demand-side throughput signal.
Publishing that null is deliberate. A data vendor's positive claims are only worth what its negative claims cost, and the standard a buyer should hold every quantitative dataset to is the same one outlined in How to Evaluate Retail Sentiment Data Before You Backtest It: leakage-safe validation, stated baselines, published limits.
What the data is for
Stripped of the alpha claim, the panel is compute-cost and infrastructure intelligence, and that is where its value is real today:
- Procurement. With order-of-magnitude spreads on identical SKUs, naive single-provider procurement leaves large savings on the table. A normalized price surface turns provider selection into a measurable decision.
- Cost modeling. The true market price of every GPU class over time is the input for budgeting training runs and benchmarking internal compute spend against the market.
- Supply monitoring. Capacity and utilization snapshots show where the market is tightening before that tightness reaches announcements.
- Market structure research. Dispersion itself is a measurable feature of an immature market, and watching it compress, SKU by SKU, is watching compute become a commodity in real time.
The GPU Price Index ships the normalized surface hourly, with the equity, options, and DePIN reference layers attached as context for analysts. The research conclusion is the honest frame: descriptively valuable now, a trading signal only if the data one day proves it.