The newest-generation cloud GPU cohort (B200, B300, MI300X, RTX 5090) saw on-demand median hourly rental rise from $2.12 in October 2024 to $4.72 in September 2026, while the Modern cohort (H100, H200, A100, L40S, RTX 4090, A10G, T4, L4) moved from $1.35 to $1.45 over the same period [S4].
The Legacy silicon cohort (V100, P100, K80, M60, P40) saw its on-demand median hourly rental fall from $1.77 in October 2024 to $0.95 in September 2026 [S4]. The divergence is the dominant signal in the 2026 GPU pricing landscape.
Cloud rental price index by generation cohort
The AIMultiple index groups 17 tracked SKUs into three fixed cohorts to keep the time series comparable even as individual listings churn [S4]. Each GPU contributes equally to its group median, so a single provider's price cut does not skew the cohort. The Last released cohort (2024 and later) has the steepest curve: B200 and B300 medians now sit at $6.52 and $7.87 per GPU-hour on-demand, while MI300X runs at $2.91, undercutting H100 at $3.25 [S4]. RTX 5090 at $0.63 and RTX 4090 at $0.44 reflect consumer-card economics entering the rental pool, with the 4090 still widely available at 68% of tracked listings and the 5090 at 79% [S4].
MI300X reserved pricing reverses the usual discount pattern: its reserved median of $3.39 sits above the $2.91 on-demand median, a signal that long-term MI300X commitments are scarce enough to carry a premium, while B300's reserved median of $5.71 against its $7.87 on-demand still delivers a conventional 27% discount for committed terms [S4]. H100's reserved and on-demand gap is only $0.13 ($3.10 versus $3.23), reflecting mature supply [S4].
Retail-side price tracking and MSRP versus street
Tom's Hardware maintains a daily-updated GPU price index tracking lowest street prices across NVIDIA, AMD, and Intel cards, designed to help buyers navigate the gap between MSRP and retailer-listed price [S1]. Semianalysis publishes a separate GPU Pricing Index built as a composition-jump-resistant weighted mean of rental prices across the cloud market, intended to give a single robust number that survives provider entry and exit [S2]. The two indices answer different questions: Tom's tracks consumer purchase cost; Semianalysis tracks compute cost per GPU-hour.
Pagecrawl's December 2025 retail-tracker guide identified five price-shaping forces acting on street price: MSRP anchor, new-generation launch cycles, mining demand waves, tariff and supply-chain impacts, and seasonal retailer promotions [S5]. The guide cited a $999 MSRP RTX 5080 selling at $1,149 or higher across major retailers, with a transient $1,029 dip at one outlet lasting under 24 hours, an example of the $100-300 swing band that price-drop alerts are designed to catch [S5].
Provider-level dispersion: H200 ranges from $2.09 to $13.78
Provider choice adds a second axis of price variation. September 2026 H200 on-demand listings span $2.09 per GPU-hour at Beam to $13.78 at Microsoft Azure, a 6.6x spread that reflects differences in instance configuration, stock state, and data-residency tier rather than raw hardware markup [S4]. IONOS prices H100 and H200 as flat $3,990 monthly dedicated servers with EU data residency, a billing model that hides the hourly rate but is competitive against the $3.25 to $4.40 hourly H100/H200 medians for buyers needing residency guarantees [S4].
Cloud GPU pricing has bifurcated between hyperscalers (Azure, AWS, GCP) anchoring the high end with full-managed services and neoclouds (Beam, RunPod, Vast, Lambda) anchoring the low end with bare-metal or spot capacity. The price gap is widest on H200, narrowest on commodity cards like T4 and L4 where multiple providers converge near cost-of-power-plus-amortization [S4].
Supply availability signals and MI300X scarcity
Confirmed availability as a share of all tracked listings is lowest for MI300X at 7%, followed by B300 at 11% and B200 at 16% [S4]. H100 confirms availability on 33% of listings, while RTX 4090 and RTX 5090 reach 68% and 79% respectively, the consumer-card ceiling where almost every listed instance is actively rentable [S4]. However, 88% of MI300X listings carry unknown stock status, so the 7% confirmed-available figure cannot be inverted to mean 93% sold out, the denominator includes available, unknown, waitlisted, and unavailable entries across all billing tiers [S4].
For a procurement team reading the index, the actionable signal is provider-count and listing-count, not the confirmed-available percentage in isolation. A 7% confirmed rate across 40 listings is stronger evidence of scarcity than a 7% rate across 5 listings, and the index does not currently publish the listing-count denominator publicly [S4].
Build-vs-buy cost driver: hourly rate versus workload completion
Hourly rental rate alone does not measure workload cost. The 2026 GPU index recommends pairing the rental table with a multi-GPU benchmark covering H100, H200, B200, and MI300X to compute time-to-completion on representative training and inference workloads [S4]. A 1.8x faster GPU at 1.5x the hourly rate is cheaper per job, a calculation that flipped H200's value proposition in 2025 once NVLink and FP8 throughput differences were measured on production transformer training rather than peak FLOPS [S4].
On the consumer side, Pagecrawl's tracker identified two purchase-timing windows within any GPU generation: 6-9 months post-launch when supply stabilizes and board-partner competition widens, and the pre/post-next-gen transition when retailers discount remaining inventory to clear shelf space [S5]. The 2025 RTX 5080 example, a $999 MSRP card that held $1,149 or higher street pricing, saw a transient drop to $1,029 at Best Buy as the only sub-$1,149 print captured in the public example [S5].
Methodology notes and known limitations
The AIMultiple index computes each median from listings present in the given month, so provider and SKU turnover can shift the median even without an underlying rate change [S4]. Physical variants sold under the same model name (A100 40GB versus 80GB, H100 PCIe versus SXM versus NVL) are combined unless the source explicitly separates them by memory or interconnect [S4]. Semianalysis applies a composition-jump-resistant weighted mean specifically to suppress this kind of turnover noise, at the cost of a less transparent per-GPU median [S2].
Reservation records aggregate across different commitment lengths, so the reserved-versus-on-demand gap is not a clean like-for-like discount from a single provider on an identical instance; it is a market-wide median comparison [S4]. The Tom's Hardware retail index is daily-updated and model-anchored to lowest listed price across tracked retailers, so it misses private bundles, OEM system pricing, and used-card markets unless those listings surface on the tracked retailers [S1][S5].
Reading the signals: a 2026 buyer framework
For a hyperscale or neocloud capacity buyer, the dominant 2026 signal is the 122% rise in newest-cohort on-demand pricing against flat Modern pricing, which means the cost of stepping up to B200/B300 has roughly doubled since late 2024 while the cost of staying on H100/A100 has not [S4]. The H200 6.6x provider spread ($2.09 to $13.78) is the next-most-actionable number, because a workload that runs acceptably on H200 can be cost-cut by 5-6x simply by changing provider, with residency and managed-service features as the trade-off [S4].
For a consumer or workstation buyer, the actionable signals are the 68-79% availability on RTX 4090/5090 listings (deep rental supply, weak secondary-market demand), the 7-16% availability on MI300X/B200/B300 (tight cloud supply, indicating vendor pull-through demand), and the two retail-timing windows at 6-9 months post-launch and at the generation transition [S4][S5]. Pairing the Tom's Hardware retail index with automated drop-alert tooling, as the Pagecrawl guide recommends, catches the 24-48 hour $100-300 dip windows that manual checking misses [S1][S5].
Trackable next nodes: the September 2026 weekly AIMultiple release for whether the newest-cohort rise has topped out past $4.72/hr [S4], the B300/MI300X listing-count denominators once AIMultiple publishes them, and the Semianalysis composition-jump methodology revision notes for whether the weighted mean shifts on provider entry [S2]. For a related reference on vendor positioning, see the NVIDIA GPU roadmap 2026-2028 and the AI server competitive landscape analysis, which together frame the supply side behind the rental price moves.
For component-level specifications, see construction machinery and equipment, lamps and light fittings, and lighting equipment and electric lamps.