AltStreet Research: The Same H100 Was Offered at Nearly a 7x Spread Across a Five-Day Live Panel

AltStreet published a new primary-data research analysis today: seven and a half months of daily GPU rental price collection, scoped strictly to live-collected observations — 11,962 API and scrape observations recorded on 227 of 231 days between November 27, 2025 and July 15, 2026, yielding two insight sets: a nine-month time series for two venues (Vast.ai, RunPod) and a seven-venue cross-market live panel since a July 11 collection upgrade. The central finding is structural: there is no single GPU price. Within the strict five-day July panel, identical H100 SXM5 configurations were offered at prices differing by nearly 7x, and across GPU models and markets the cost of one normalized unit of dense FP16 compute spanned 13.7x.

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

What the live data shows

• —Same-SKU, same-week spread: in the July 11-15 live panel, the H100 SXM5 80GB was offered at a median $7.28/hr on Azure, $4.31/hr on CoreWeave, $4.14/hr on Lambda, and $1.05/hr on Vast.ai — approximately 6.9x for the same H100 SXM5 accelerator SKU across differently bundled venue offerings. The Vast.ai strict-window median is based on three observations; its July-to-date median of $1.21 produces an approximately 6x spread. The A40 spanned 6.3x ($0.30/hr RunPod vs. $1.89/hr Paperspace) in the same panel
• —Normalized spectrum: one dense FP16 PFLOP-hour ranged from $0.92 (RTX 5090 on SaladCloud) and $1.02 (RTX 4090 on Vast.ai) to $12.60 (A40 on Paperspace) across the contemporaneous live panel — 13.7x. The cheapest unit of FP16 compute is consumer silicon, priced down by its real constraints (limited VRAM, no NVLink)
• —Higher-end datacenter silicon is not necessarily cheaper per dense FP16 FLOP at observed prices: within the panel, H100 at its cheapest venue ($3.23 per PFLOP-hr, RunPod PCIe) costs more per dense PFLOP than lower-priced A40 and L40S alternatives (~$2.00-$2.02)
• —The two full-window live venues diverged on the identical card: from November 2025 to July 2026, RunPod's observed RTX 4090 median rose approximately 11% ($0.465 to $0.515/hr) as higher-priced posted tiers entered its available mix, while Vast.ai's fell approximately 58% ($0.402 to $0.169/hr) — the cross-venue ratio widened from 1.2x to 3.1x
• —A customer-price hardware-recovery calculation — explicitly not a host-profitability model — differed by approximately 2.5x-3.8x across the displayed same-SKU venue pairs (2.5x for H100 SXM5 across Azure and Vast.ai bases; 3.8x for RTX 4090 across RunPod and Vast.ai)

The methodology finding

The analysis opens its correction log with an artifact caught in AltStreet's own pipeline. The collection's 16,337 rows reconcile as 11,962 live observations plus 4,375 manually recorded rows excluded from all findings — 4,367 static hyperscaler reference prices entered at collection setup and echoed daily until the July 11 upgrade replaced them with live feeds, plus eight first-day-only rows from providers never re-collected. The pre-July hyperscaler rows look exactly like nine-month observed series of unchanged list prices; treating them as observations would have manufactured a list-price-persistence finding out of collection design. The analysis therefore makes no such claims and states its scope rule — live source kinds only — in the reproducibility appendix, alongside a day-weighted sensitivity check on every exhibited figure. The general lesson is stated in the piece: a constant series in a price panel is a claim about the collector until proven otherwise.

Two further artifacts are documented: an apparent ~45% July drop in cloud H100 pricing (a composition change across the collection upgrade, not a price cut) and an apparent ~30% rise in Vast.ai H100 offers (low-observation endpoint months; within fixed configuration bands the series is rangebound around $1.5-$1.7/hr).

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. The dataset supports different directional conclusions depending on the venue and pricing regime — within AltStreet's observed panel, the pronounced decline was concentrated in consumer-card offers on Vast.ai, while higher price levels appeared in posted tiers and live cloud offers. It does not support treating either a broad GPU-price collapse or a universal compute shortage as a market-wide fact.

The supply-side arithmetic is the second implication. At observed marketplace price levels, the customer-price hardware-recovery equivalent extends from roughly 1.5 years to several years at lower utilization — before marketplace payouts, full-system costs, hosting, and financing are considered — a result AltStreet interprets as consistent with some marginal supply operating on already-owned hardware, although the dataset does not observe host cost basis or utilization. The equivalent runs approximately 2.5x-3.8x faster at higher-priced venues for the same SKU. That dispersion, 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; providers named have no commercial relationship with the research. Only live-collected rows enter any exhibit. Figures are offer-weighted medians with counts disclosed per exhibit; the July cross-venue panel covers five collection days and is labeled as such. Observed prices are asking prices, not completed transactions, and bundle differing CPU, memory, networking, reliability, and support around the same accelerator. The hardware-recovery exhibit is a customer-price equivalent, not a host-profitability model. The dataset and analysis will be revised quarterly. This is not investment advice.