AltStreet Research: Identical H100s Offered at Prices Differing by 10x — and the Cheapest Compute Is a Consumer Card

AltStreet published a new primary-data research analysis today: seven and a half months of daily GPU rental price collection — 16,337 offer observations across nine venues, recorded on 227 of 231 days between November 27, 2025 and July 15, 2026. The central finding is structural: there is no single GPU price. Identical H100 SXM5 configurations were offered at prices differing by roughly 10x across venues, and across GPU models and markets the cost of one normalized unit of dense FP16 compute spanned 12.9x.

The full analysis, including exhibit-level tables with observation counts and a reproducibility appendix, is published in AltStreet's AI Infrastructure & Compute research section.

What the data shows

• —Same-SKU spread: the H100 SXM5 80GB was offered at $12.29/hr at AWS list and $10.35/hr at Google Cloud list throughout the window, against monthly median observed offers of $1.21–$2.94/hr on Vast.ai — up to roughly 10.2x for the same configuration
• —Normalized spectrum: one dense FP16 PFLOP-hour ranged from $1.02 (RTX 4090, Vast.ai offers) to $13.13 (A100, AWS list) — 12.9x across models and venues. The cheapest unit of FP16 compute in the dataset is a consumer card, priced down by its real constraints (24GB VRAM, no NVLink)
• —Newer datacenter silicon is not cheaper compute at observed prices: H100 at its cheapest venue (~$2.90–$2.97 per PFLOP-hr) costs more per dense PFLOP than lower-priced A40 and L40S alternatives (~$2.00)
• —Hyperscaler list prices contained no price discovery: AWS and Google Cloud H100/A100 series each showed exactly one distinct price across 432 daily observations — the observed list price did not change during the collection window
• —Two marketplaces diverged on the identical card: RunPod's observed RTX 4090 median rose from $0.34 to $0.515/hr as higher-priced posted tiers entered its available mix, while Vast.ai's median observed offers fell roughly 58% from their November level — the cross-venue ratio widened from near-parity to ~3x
• —A customer-price hardware-recovery calculation — explicitly not a host-profitability model — shows the recovery arithmetic differing by roughly 5x across venue price levels for comparable hardware classes

Background

The dataset's providers resolve into three pricing regimes: hyperscalers publishing static list prices (AWS, Google Cloud, Azure), venues posting administered tier prices that change by decision (RunPod, Lambda, CoreWeave, Paperspace), and marketplaces carrying continuously repriced offers from independent hosts (Vast.ai, SaladCloud). The analysis treats these as different instruments measuring the same underlying hardware — and finds that which trend an observer reports (rising, falling, or flat) depends almost entirely on which regime they measure.

All figures are observed asking prices, not completed transactions: the dataset does not record whether an offer rented or what utilization it achieved. AltStreet documents this limitation, along with the unit of observation (daily offer observations, with serial correlation across repeated sightings of unchanged prices), in the analysis itself.

The methodology findings

Two apparent trends in the dataset are documented in the analysis as measurement artifacts rather than findings — a discipline AltStreet applies to its securities-filing coverage and extends here. An apparent ~45% July drop in cloud H100 pricing is a panel-composition change: Azure, CoreWeave, and Lambda H100 offers entered collection that month at lower price points while AWS and Google Cloud held unchanged. An apparent ~30% rise in Vast.ai H100 offers is an endpoint artifact of low-observation months; within fixed configuration bands the series is rangebound around $1.5–$1.7/hr.

A third correction was applied before publication: manufacturer spec sheets mix dense and 2:1-sparsity TFLOPS conventions across GPU generations, which would have distorted the cost-per-compute table by 2x. All throughput figures were re-normalized to the dense FP16 convention.

Why it matters

For allocators and operators evaluating compute costs or compute-adjacent investments, the venue spectrum reframes the standard question. A GPU price quote is meaningless without its regime: list prices measure what procurement departments pay for bundled infrastructure, compliance, and support; posted tiers measure marketplace operators' pricing decisions; repriced offers measure what independent supply will accept. Public claims that GPU prices are collapsing, or that compute shortages persist, are each true of exactly one regime and false of the others.

The supply-side arithmetic is the second implication. At observed marketplace offer levels, consumer-card hosts face hardware-recovery timelines of 1.5 years to several years at realistic utilization — before platform fees and system costs — consistent with marginal supply operating on sunk hardware. At hyperscaler price levels, the same arithmetic runs roughly 5x faster. That asymmetry, not any single price trend, is the durable structure of this market as observed.

Scope and limitations

The analysis is built exclusively from AltStreet's own collection pipeline; no vendor benchmarks or provider-supplied summaries are used, and providers named have no commercial relationship with the research. Figures are medians of equally weighted daily offer observations with counts disclosed per exhibit; observed prices bundle differing CPU, memory, networking, reliability, and support around the same accelerator, and comparisons measure the priced bundle, not the bare chip. Hyperscaler and neocloud figures are list prices that large buyers negotiate below. The hardware-recovery exhibit is a customer-price equivalent, not a host-profitability or payback model. AWS and Google Cloud collection paused July 11, 2026. The dataset and analysis will be revised quarterly with revision history noted. This is not investment advice.