H100 cloud index opens October 5 at $3.49 per GPU-hour
Source report: 2026-10-05 · Editorial analysis published: 2026-10-05
GPU Economy’s provisional weekly index shows a $3.49 median from 31 listed rates across 29 providers. The data measures advertised on-demand H100 pricing, not booked capacity or a miner’s AI revenue.
Analysis and practical implications
This section is our analysis and illustrative calculations, separate from the source report.
A dated benchmark for advertised compute
GPU Economy’s Cloud GPU Price Index showed a median of $3.49 per NVIDIA H100 GPU-hour for the week beginning October 5, 2026. We checked the source snapshot timestamped October 5 at 12:40 UTC. It contains 31 publicly listed on-demand rates from 29 providers. The week is still running, so the figure is provisional rather than a finalized weekly result. This is a market-data update with a stated measurement window, not an announcement of a new ASIC or a promise that a customer can immediately reserve every advertised offer.
The timestamp also prevents a moving source page from being mistaken for a fixed historical quote. If rates or panel members change later in the week, a reader checking the live index may see a different number. A dated snapshot and its constituent list allow the current article to remain interpretable without presenting every later reading as a separate market event.
The sample describes listings, not physical capacity
The index uses public price quotations. A provider appearing in the panel does not tell us how many H100 cards are available, whether a requested region has stock or whether a large cluster can be delivered on a particular date. Quotation count and installed GPU count must remain separate. For a miner evaluating an AI conversion, this is a useful reference for what customers can see in public rate cards. It is insufficient by itself to forecast the operator’s contracted sales, achievable utilization or customer acquisition costs. Those inputs require actual offers and negotiated service terms.
A rate card may apply to a single instance while a customer needs a tightly connected cluster. The ability to rent one GPU is therefore a different procurement question from the ability to reserve many GPUs with a shared network and a synchronized start. Neither quantity can be recovered by multiplying the number of providers shown in the index.
Normalization makes prices comparable within limits
The source normalizes a node price to one GPU-hour by dividing by the card count, and takes the median rather than an arithmetic mean. Its panel uses the newest eligible quote for each provider-region pair. Such normalization helps compare headline listings, but the underlying service can still differ in CPU allocation, memory, storage, networking, support and billing conditions. Our assessment is that a buyer should compare the complete configuration before choosing an offer. An identical GPU model name is not sufficient evidence that two services will deliver the same training speed, reliability or total invoice.
For a workload comparison, throughput per billed hour can matter as much as the nominal price per card. Software settings, interconnect and storage performance can change how long the same task takes. A lower hourly quotation does not automatically mean a lower completed-job cost; buyers need a benchmark representative of their own workload.
A weekly median is not a pure price-change measure
The current $3.49 median is also the source’s finalized value for the preceding September 28–October 4 week. That equality does not prove every provider held its rate constant. A median can remain unchanged while individual quotations move, and a changing panel can shift a median without the same providers changing their prices. The source separately publishes an index intended to track like-for-like price changes. Readers should keep the level, the panel and the change measure attached to the appropriate reporting period. We do not infer a sector-wide pricing trend from one provisional point.
The source’s provisional weekly value can update as new observations arrive. It should not be described as a settlement price or a guaranteed contract rate. The finalized previous week offers a dated historical comparator, but a trend assessment still requires a consistent methodology over more than one or two observations.
An eight-GPU example illustrates billing scale
At the headline median, eight GPUs used for one hour would cost 8 × $3.49 = $27.92. For 24 hours, the corresponding arithmetic is $670.08. These are illustrative calculations using the published median, not a provider quote or an estimate of mining profit. Actual billing can depend on minimum duration, storage, data transfer, taxes, discounts and the service configuration. A node with eight GPUs may have a package price that differs from eight separate instances. The exercise is useful for keeping GPU-hours and node-hours distinct when comparing a compute proposal.
For illustration, halving usage from 24 to 12 hours would halve the GPU-hour subtotal under simple hourly billing, but not necessarily every cost on an invoice. Storage may remain allocated and a committed contract may still be payable. This is why an hourly example should identify its assumptions instead of implying that all cloud charges scale with active GPU time.
A listed customer rate is not operator revenue
An infrastructure owner cannot multiply its theoretical maximum GPU-hours by this index and call the result expected revenue. That approach assumes full utilization at a public customer-facing price while ignoring the selling channel, contract discounts, interruptions and collection risk. Our interpretation is that the index is a reference point for procurement discussions, not a replacement for an operating model. A credible forecast would use contracted net rates and a realistic utilization profile, with separate treatment of capital costs, power, cooling, repairs, software and staffing. The spread between a customer’s bill and a site’s net receipts can be material.
The relevant revenue unit for a site might be a reserved cluster, a managed service or a wholesale capacity agreement. Its pricing need not match a public retail card-hour. Operators should model the contract they intend to sell, including responsibility for networking and support, rather than substituting a readily available public price for an undisclosed commercial term.
Bitcoin hashprice and GPU-hour pricing measure different things
Bitcoin hashprice normally expresses expected mining revenue per unit of hashrate over time, commonly dollars per PH/s-day. A cloud GPU rate measures the advertised charge for renting a GPU for an hour. One is a mining revenue reference and the other a compute service price; neither can be substituted directly for the other. The workloads, hardware and customer obligations also differ. For a mining company considering AI infrastructure, comparisons should be made through complete project cash flows and service requirements. A higher-looking unit price alone cannot show that a conversion will produce a better return.
The hardware decision also requires checking conversion costs and residual equipment value. ASICs execute specialized mining work; GPUs support different computing tasks and software requirements. Existing power access may be useful in both settings, but it does not convert a mining facility into a fully provisioned AI service without additional engineering and commercial work.
Use the snapshot with its methodology and vintage
GPU Economy publishes constituent quotations, receipts and a methodology, and labels completed weeks separately from the current provisional one. Readers can use those materials to check how the headline is constructed. We cite the October 5 snapshot rather than presenting the number as a fixed price available indefinitely. The figures cover publicly listed on-demand H100 pricing and do not establish negotiated contract rates, spot availability or the revenue a particular data centre will earn. The practical result is a transparent reference for evaluating compute offers, provided its sample and measurement limits remain visible.
A reproducible benchmark is most useful when the reader can inspect which offers entered it and what was excluded. This source’s published receipts are a route to that check. For an actual procurement decision, the final step remains obtaining a current offer from the provider and confirming capacity, configuration, duration and all charges applicable to the requested service.
Source: GPU Economy ↗
Mining calculator ↗
