Quick answer
“The GPU price” is not a number — it is a market-structure spectrum, because buyers are purchasing different bundles around the same accelerator. A cloud price wraps the chip in networking, storage, compliance, support, and provider margin; a marketplace offer may provide a less-standardized, independently hosted bundle with fewer enterprise services around the accelerator. In AltStreet’s live July 2026 panel, one dense FP16 PFLOP-hour was offered at $0.92–$1.02 on consumer-card marketplaces, $2.00–$3.23 at posted marketplace rates, and $4.19–$12.60 at live cloud offers. Which trend you see — rising, falling, or flat — depends entirely on which pricing regime you measure.
Data notes & sources
Collection & unit of observation. AltStreet’s pipeline recorded 16,337 rows between November 27, 2025 and July 15, 2026 — 227 of 231 calendar days, no gap exceeding three days. This analysis uses only the 11,962 live-collected observations (11,622 via provider APIs, 340 via scraping). The remaining 4,375 manually recorded rows are excluded from every finding: 4,367 static reference prices for hyperscaler list rates, entered at collection setup and echoed daily until a July 11, 2026 collection upgrade replaced them with live feeds (see Artifact 01), plus eight first-day-only rows from three providers never re-collected. A persistent listing observed on multiple days contributes multiple observations; daily panels carry serial correlation. Observations are equally weighted; see the reproducibility appendix.
Venues. Two venues have live collection across the full window: Vast.ai (continuously repriced offers from independent hosts; 1,910 observations, 375 distinct prices) and RunPod (administered posted tiers; 9,634 observations, 22 distinct prices). Five more entered live collection on July 11, 2026: Azure (API), CoreWeave, Lambda, Paperspace, and SaladCloud (scrape). The analysis therefore presents two insight sets and labels every exhibit accordingly: trend findings from the two full-window venues (Exhibits 1b, 3), and breadth findings from the seven-venue live panel (Exhibits 1, 2, 4). AWS and Google Cloud have no live coverage in this dataset; their rows are static manual reference prices and appear in no exhibit. Providers with first-day-only coverage (FluidStack, gpu.earth, Genesis Cloud) are excluded.
Normalization. Cost-per-compute figures use dense FP16 tensor TFLOPS. Published spec sheets mix NVIDIA’s dense and 2:1-sparsity figures across generations; unadjusted, this distorts cross-generation comparisons by 2x. All figures were re-normalized to the dense convention before any ratio was computed.
What these prices are — and are not. Cloud figures (Azure, CoreWeave, Lambda, Paperspace) are live-collected posted offer prices, which large buyers negotiate below; marketplace figures are administered posted tiers (RunPod) or continuously repriced offers from independent hosts (Vast.ai, SaladCloud). Offers are not transactions: this dataset does not observe whether an offer rented, how long it stayed available, or what utilization it achieved. Venue prices also bundle different CPU/RAM allocations, networking, storage, interconnect topology, reliability, support, and minimum terms around the same accelerator — comparisons here measure the priced bundle, not the bare chip. Not investment advice.
The Live Venue Spectrum: One SKU, Nearly 7x in the Same Five-Day Panel
The July 11 collection upgrade produced something rare: a fully live, same-week, cross-venue panel for the market’s benchmark accelerator. Every figure below is a live-collected observation — no list prices, no reference values:
| Venue | H100 SXM5 80GB | n | H100 PCIe 80GB | n |
|---|---|---|---|---|
| Azure (live API) | 7.281 | 16 | 4.192 | 14 |
| CoreWeave (live) | 4.309 | 18 | — | — |
| Lambda (live) | 4.140 | 36 | 3.290 | 9 |
| RunPod (posted tiers) | — | — | 2.440 | 16 |
| Vast.ai (repriced offers) | 1.050 | 3 | — | — |
Source: AltStreet GPU price pipeline, live API/scrape observations only. SXM5 and PCIe are different physical configurations of the same accelerator family and are not directly comparable. Vast.ai SXM5 n=3 in the strict panel window (July full-month median $1.21) — low-n flagged. Strict-panel SXM5 spread: ~6.9x (Azure $7.281 vs. Vast $1.050); ~6x against Vast’s July-to-date median ($1.21).
The full-window live series adds the time dimension for the venue that reprices continuously — and shows what the July panel’s snapshot cannot: the marketplace level is not a July anomaly but where this SKU has been offered all along:
| Month | Median $/hr |
|---|---|
| Dec 2025 | 1.62 |
| Jan 2026 | 0.60 (n=4) |
| Feb 2026 | 1.47 |
| Mar 2026 | 1.47 |
| Apr 2026 | 1.47 |
| May 2026 | 2.94 |
| Jun 2026 | 1.68 |
| Jul 2026* | 1.21 |
Source: AltStreet GPU price pipeline, live API observations. Monthly n: 4–108; low-n months flagged. *July partial month.
Buyers at the top of the panel are purchasing interconnect, reliability, compliance surface, and support around the accelerator; buyers at the bottom are purchasing a less-standardized bundle whose price is more heavily determined by access to the accelerator itself. The nearly 7x same-panel spread — and the 6.3x spread on the humbler A40 ($0.30/hr RunPod vs. $1.89/hr Paperspace, same panel) — measures the market’s price for everything around the chip.
Cost per PFLOP: The Cheapest Compute Is a Consumer Card
Normalizing live observed offers to dense FP16 throughput collapses every model and venue onto one axis — and the ranking inverts the intuition that higher-end datacenter silicon is cheaper compute per FLOP:
| GPU · venue | $/hr | n | $/PFLOP-hr | |
|---|---|---|---|---|
| RTX 5090 · Salad offers | 0.294 | 9 | 0.92 | |
| RTX 4090 · Vast offers | 0.169 | 9 | 1.02 | |
| RTX 4090 · Salad offers | 0.204 | 9 | 1.24 | |
| RTX 5090 · Vast offers | 0.470 | 31 | 1.48 | |
| A40 · RunPod posted | 0.300 | 64 | 2.00 | |
| L40S · RunPod posted | 0.740 | 48 | 2.02 | |
| RTX 4090 · RunPod posted | 0.515 | 16 | 3.12 | |
| H100 PCIe · RunPod posted | 2.440 | 16 | 3.23 | |
| H100 SXM5 · Lambda live | 4.140 | 36 | 4.19 | |
| H100 PCIe · Lambda live | 3.290 | 9 | 4.35 | |
| H100 SXM5 · CoreWeave live | 4.309 | 18 | 4.36 | |
| L40S · CoreWeave live | 1.618 | 18 | 4.42 | |
| H100 PCIe · Azure live | 4.192 | 14 | 5.54 | |
| A100 40GB · Lambda live | 1.990 | 18 | 6.38 | |
| A100 40GB · Azure live | 2.074 | 16 | 6.65 | |
| A100 80GB PCIe · Azure live | 2.176 | 16 | 6.97 | |
| H100 SXM5 · Azure live | 7.281 | 16 | 7.36 | |
| A40 · Lambda live | 1.190 | 18 | 7.93 | |
| A40 · Paperspace live | 1.890 | 8 | 12.60 |
Source: AltStreet GPU price pipeline joined to dense FP16 tensor specifications (see Data Notes). Live API/scrape observations only; every row is the Jul 11–15 window. Vast.ai H100 SXM5 excluded here (n=3 in window; see Exhibit 1b for its full series). B200 excluded — insufficient live observations.
Three structural facts fall out. First, the compute-cost frontier is consumer silicon on repriced marketplaces: SaladCloud’s RTX 5090 at $0.92 and Vast.ai’s RTX 4090 at $1.02 per dense PFLOP-hr undercut every datacenter option — a rational discount, since limited VRAM and no NVLink make them generally unsuitable for frontier-scale training; the market prices the constraint, not the FLOPs. Second, 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, RunPod PCIe) costs more per dense PFLOP than the lower-priced A40 and L40S alternatives (~$2.00–$2.02) — Hopper buyers pay for density, interconnect, and memory bandwidth, not FLOP efficiency. Third, the same silicon spans the whole spectrum depending on venue: the A40 alone runs $2.00 to $12.60 per PFLOP-hr across three live venues in the same five-day panel.
| GPU · venue | $/hr | n | $/PFLOP-hr |
|---|---|---|---|
| RTX 4090 · Vast offers | 0.169 | 57 | 1.02 |
| RTX 5090 · Vast offers | 0.469 | 222 | 1.47 |
| A40 · RunPod posted | 0.300 | 920 | 2.00 |
| L40S · RunPod posted | 0.740 | 690 | 2.02 |
| H100 PCIe · RunPod posted | 2.190 | 230 | 2.90 |
| H100 SXM5 · Vast offers | 2.936 | 108 | 2.97 |
| RTX 4090 · RunPod posted | 0.515 | 230 | 3.12 |
Source: AltStreet GPU price pipeline, live API observations, May–Jun 2026 medians (n 57–920). Rankings match the July panel for every overlapping row. Sensitivity: offer-weighted and day-weighted medians agree within rounding for all cells here except Vast.ai H100 SXM5 (offer $2.94/$1.68 vs day-weighted $4.40/$2.06 in May/Jun, consistent with high-listing days at lower prices); both methods show the May tightening followed by a substantial June reversal, with the July series returning below the pre-spike range.
Administered vs. Repriced: The Same Card, Opposite Directions
Both full-window venues are live API feeds — and a distinct-price-values test separates their regimes cleanly: Vast.ai’s offers show near-continuous variation (375 distinct prices; independent hosts repricing), while RunPod’s 877 RTX 4090 observations contain exactly three distinct prices — posted tiers, administered by decision. Over the window, the two regimes moved in opposite directions on the same card:
| Month | RunPod posted | Vast.ai offers | Ratio |
|---|---|---|---|
| Nov 2025 | 0.465 | 0.402 | 1.2x |
| Dec 2025 | 0.340 | 0.315 | 1.1x |
| Jan 2026 | 0.465 | 0.308 | 1.5x |
| Feb 2026 | 0.465 | 0.189 | 2.5x |
| Mar 2026 | 0.465 | 0.176 | 2.7x |
| Apr 2026 | 0.465 | 0.150 | 3.1x |
| May 2026 | 0.515 | 0.169 | 3.1x |
| Jun 2026 | 0.515 | 0.269 | 1.9x |
| Jul 2026* | 0.515 | 0.169 | 3.1x |
Source: AltStreet GPU price pipeline. RunPod monthly n: 58–122; Vast.ai monthly n: 10–74. RunPod operates multiple concurrent tiers per SKU (community/secure, configuration variants); its median reflects the buyer-facing tier mix, and a rising median can reflect tier composition as well as repricing. *Partial month.
The tier record shows the administered side explicitly: RunPod introduced a higher RTX 4090 tier ($0.69/hr) on April 17, 2026 and a higher H100 PCIe tier ($2.89/hr, +21% over the prior top) on May 20, 2026, while trimming L40S and A40 tiers slightly. From November 2025 to July 2026, RunPod’s observed median increased approximately 11%, from $0.465 to $0.515/hr, as higher-priced tiers entered the available mix, while Vast.ai’s fell approximately 58%, from $0.402 to $0.169/hr. The cross-venue ratio on the identical card widened from 1.2x to 3.1x — a spread that persists because the venues sell different bundles: RunPod sells reliability tiers and orchestration; the independently hosted venue sells less-standardized compute under continuously repriced offers.
Customer-Price Hardware Recovery: A 2.5x–3.8x Illustrative Same-SKU Venue Gap
| GPU · venue price basis | $/hr | 100% util | 50% util | 30% util |
|---|---|---|---|---|
| RTX 4090 · RunPod posted | 0.515 | 142 | 289 | 495 |
| H100 SXM5 · Azure live (Jul) | 7.281 | 173 | 347 | 580 |
| RTX 5090 · Vast offers | 0.469 | 202 | 416 | 722 |
| H100 SXM5 · CoreWeave live (Jul) | 4.309 | 295 | 592 | 991 |
| H100 SXM5 · Lambda live (Jul) | 4.140 | 307 | 616 | 1,032 |
| RTX 5090 · Salad offers (Jul) | 0.294 | 352 | 740 | 1,324 |
| H100 PCIe · RunPod posted (Jul tier) | 2.440 | 433 | 869 | 1,454 |
| H100 SXM5 · Vast offers | 2.936 | 436 | 877 | 1,471 |
| RTX 4090 · Vast offers | 0.169 | 537 | 1,159 | 2,156 |
| L40S · RunPod posted | 0.740 | 591 | 1,194 | 2,017 |
Method: GPU-module MSRP ÷ (24h × utilization × rate − electricity), electricity = full TDP during rented hours + 20% TDP idle otherwise, at $0.10/kWh. Cloud rows use Jul 11–15 live medians; Vast/RunPod rows use May–Jun medians. Excludes provider margin, platform fees/host payout share, full-system capex, hosting, cooling, financing, failures.
Even as a customer-price equivalent, the asymmetry is structural — and it varies by market segment. Using the stable marketplace medians and five-day cloud panel shown above: at full utilization, the recovery equivalent differed by approximately 2.5x for the H100 SXM5 across Azure and Vast.ai bases (173 vs. 436 days) and approximately 3.8x for the RTX 4090 across RunPod and Vast.ai (142 vs. 537 days) — before the deductions that fall hardest on exactly the venues with the slowest arithmetic. It is consistent with the supply behavior visible in Exhibit 3: hosts still listing consumer cards at $0.169/hr are plausibly running sunk hardware on cheap power, which is how offer medians can fall 58% while posted venues introduce higher tiers into the same market. A full host-economics model — payout shares, system capex, colocation, financing — is the natural sequel to this analysis, and this exhibit is its input, not its answer.
What This Data Does Not Show
Correction log — three artifacts we almost published
Artifact 01“Hyperscaler list prices never moved — 432 observations, one price.” The pipeline’s pre-July AWS and Google Cloud rows look exactly like a nine-month observed series of unchanged list prices. They are not observations at all: those rows are static reference prices recorded manually at collection setup and echoed daily until the July 11 upgrade replaced them with live feeds. Treating them as a series would have produced a list-price-persistence finding built on our own collection design — so this analysis makes no such finding, and every exhibit uses live-collected observations only. The general lesson generalizes: a constant series in a price panel is a claim about the collector until proven otherwise.
Artifact 02“Cloud H100 prices dropped ~45% in July.” They did not. Live cloud collection beganJuly 11 — Azure, CoreWeave, and Lambda entered the panel at $4.14–$7.28/hr, replacing higher static reference rows. The “cut” is a composition change across the collection upgrade, the same class of error as comparing amended and unamended regulatory filings.
Artifact 03“Vast.ai’s H100 offers are up ~30%.” They are not. The endpoints are low-n months (4–6 observations), and within fixed configuration bands the series is rangebound, with a May 2026 tightening to $2.94 that fully reverted by July. The honest characterization: datacenter offer medians on the marketplace are noisy around $1.5–$1.7/hr with no established trend — which contradicts both the “GPU glut collapse” and the “compute shortage” narratives for this tier. Within AltStreet’s observed panel, the clearest decline was concentrated in consumer-card offers on Vast.ai, while higher observed price levels were concentrated in posted tiers and live cloud venues.
Every exhibit carries observation counts and provenance for the same reason. Daily offer panels are unforgiving of composition changes, small samples, serial correlation, and mixed spec conventions — the dense-vs-sparse normalization in the Data Notes corrected a further 2x distortion that published spec sheets would have introduced into Exhibit 2.
Reproducibility Appendix
Per-observation fields: provider, GPU model string, region, vCPU, RAM, storage, hourly price (USD), pricing type (spot/on-demand), source kind (api/manual/scrape), price verification date, collection timestamp. Prices are as published by each venue in USD; no currency conversion applied. GPU quantity: single-GPU-equivalent rates throughout; multi-GPU instance prices divided where the venue prices per instance.
Scope rule: only rows with source kind api or scrape enter any exhibit — 11,962 of 16,337 rows. Rows with source kind manual (4,375) are excluded from all findings: 4,367 static reference prices with no verification date, plus eight manually recorded first-day-only observations excluded for insufficient repeat coverage.
Aggregation:each provider × model × window figure is the median across all qualifying offer observations in that window (offer-weighted); observation counts (n) shown in every exhibit. A sensitivity check recomputing every exhibited cell as the median of per-day medians (day-weighted) left all rankings and trend directions unchanged; the one materially sensitive cell is disclosed in Exhibit 2b’s note. Persistent listings contribute one observation per collection day; the pipeline does not currently assign stable cross-day listing identifiers, so unique-listing counts are not reported — “11,962” is a count of daily offer observations, not unique offers. A host with many identical listings can influence a daily median; the offer-weighted method means days with more listings carry more weight, which the day-weighted sensitivity check above bounds. July-panel figures (Azure, CoreWeave, Lambda, Paperspace, SaladCloud) cover five collection days and are labeled as such.
Specifications and cost assumptions: dense FP16 tensor TFLOPS re-normalized from manufacturer specifications (dense convention; sparsity figures halved); GPU acquisition costs use module MSRPs as recorded in AltStreet’s specification table; electricity at $0.10/kWh, full TDP during rented hours and 20% TDP idle otherwise.
Data access: the live pricing interface is at /data. Aggregated per-exhibit data (provider × model × window medians with counts) is available on request while a downloadable derivative is prepared; this page will be revised quarterly.
Frequently asked questions
How much does it cost to rent an H100 in 2026?
+
It depends almost entirely on venue, not on the chip. In AltStreet's live July 2026 panel, the same H100 SXM5 80GB was offered at a median $7.28/hr on Azure, $4.31/hr on CoreWeave, $4.14/hr on Lambda, and roughly $1.05–$1.21/hr on Vast.ai — nearly a 7x spread on the same SKU within the strict five-day panel (about 6x against Vast.ai’s July-to-date median of $1.21), every figure a live-collected observation. The H100 PCIe variant — the same accelerator family in a different physical configuration, not directly comparable to SXM5 — ran $2.44/hr at RunPod's posted tiers to $4.19/hr on Azure.
What is the cheapest GPU compute available?
+
Normalized to dense FP16 throughput, the cheapest compute in AltStreet's live data is consumer silicon: an RTX 4090 on Vast.ai at about $1.02 per PFLOP-hour across May–June 2026, and in the July live panel an RTX 5090 on SaladCloud at roughly $0.92 — a third the cost of the cheapest datacenter H100 and an eighth of the most expensive live cloud offer. The discount prices real constraints: limited VRAM and no NVLink make consumer cards unsuitable for frontier training, but for inference and fine-tuning workloads that fit, they are the cost frontier.
Why do GPU rental prices differ 7x for the same card?
+
Because buyers are purchasing different bundles around the same accelerator, priced under different regimes. A cloud price includes networking, storage, compliance surface, support, and provider margin. Marketplaces like RunPod post administered tier prices that change by decision. Venues like Vast.ai and SaladCloud carry continuously repriced offers from independent hosts with varying reliability, interconnect, and configuration. In AltStreet's live July panel, the same A40 was offered at $0.30/hr on RunPod and $1.89/hr on Paperspace in the same week — a 6.3x same-SKU spread that measures the value of everything around the chip as much as the chip itself.
Are GPU prices falling in 2026?
+
It depends on which market you look at, which is the point of this analysis. Consumer-card offer medians on Vast.ai fell hard (RTX 4090: roughly $0.40 in November 2025 to $0.17/hr in July 2026, about 58%). Datacenter H100 offer medians on the same venue were rangebound with a brief May 2026 tightening that fully reverted. RunPod's observed medians rose as higher-priced posted tiers entered its available mix. A single 'GPU prices are collapsing' claim is unsupportable from this data.
Is hosting GPUs for rental income profitable?
+
This analysis cannot answer that directly — and flags analyses that claim to. Exhibit 4 computes a customer-price hardware-recovery equivalent: GPU acquisition cost divided by the end-customer price attributable to one GPU. Real host economics subtract marketplace commissions, full system cost beyond the GPU module, hosting, cooling, financing, and utilization losses; cloud prices include provider margin and infrastructure no independent host can capture. What the exhibit does show is that the recovery arithmetic differs by approximately 2.5x–3.8x across the displayed same-SKU venue pairs — the starting point for any host-economics model, not the conclusion.
How does AltStreet collect this data?
+
An automated pipeline records rental offers daily, capturing price, configuration, pricing type, and provenance fields (source kind, verification date, freshness) for every row. This analysis uses only live-collected observations: 11,962 API and scrape observations recorded between November 27, 2025 and July 15, 2026 — two venues (Vast.ai, RunPod) across the full 227-of-231-day window, and five more (Azure, CoreWeave, Lambda, Paperspace, SaladCloud) from a July 11 collection upgrade. A further 4,367 rows containing manually recorded static reference prices are excluded from every finding; see Artifact 01 and the reproducibility appendix.
Editorial independence & research methodology
This analysis is built exclusively from AltStreet’s own daily price collection — no vendor benchmarks, no provider-supplied summaries. Providers named here have no commercial relationship with this research. Figures are medians of live observed offers with counts and provenance disclosed; normalization conventions, collection upgrades, and small-sample windows are documented in the body rather than footnoted away. Wire coverage of this analysis: The Same H100 Was Offered at Nearly a 7x Spread Across a Five-Day Live Panel.