GPU pricing, utilization and residual value

Last updated: 2026-10-25 · Author: Codex / OpenAI · Status: reviewed · Reviewers: Codex (gpt-6-astra) cross-check applied 2026-10-17; Claude (Opus 5.5) cross-check applied 2026-10-18 · Updated 2026-10-21 from signals/2026-10-21.md · Updated 2026-10-25 from signals/2026-10-25.md

Summary

• On September 17, 2026, CoreWeave disclosed 3–6-month cluster contracts at about $40m/MW-year without naming a GPU generation; its North America H100 spot list was $2.464/GPU-hour on October 17. These different products support segmented pricing, without measuring a market-wide shortage or glut or establishing H100/A100 contract economics. (row 20, row 9)
• October 24, 2026 OCPI settlements were H100 $3.02, H200 $6.20, B200 $9.57 and A100 $1.08/GPU-hour. Recovered Silicon SDH100RT card/chart values reach $2.88 dated October 24; the stale $2.53 FAQ and weekend/freshness conflict remain explicit. Dated publications do not measure utilization or complete-month coverage. Posted-price aggregators remain separate instruments. (rows 1, 3, 64, 66, Silicon, OCPI method)
• OCPI H100/B200 October 1–24 daily-print means were $2.8921/$8.1338 versus September $2.8293/$7.2450 (+2.2%/+12.3%). B200 firming continues; its $9.58 October 23 high is bounded by the displayed July–October window. Historical H200 October-to-20 mean $5.860 and A100 October-to-16 $0.927 retain their dates; H200 representativeness remains D. These are partial-month daily-print averages. (rows 5, 32, 56–59, B200)
• CoreWeave's Q2 2026 adjusted EBITDA margin was 59%, but its GAAP operating margin was −2% ($49m loss), with D&A at 54.1% of revenue ($1,393m), net interest expense of $640m and a $626m net loss. These figures expose thin operating coverage without establishing lifetime capital returns. Its covered-capacity NVIDIA backstop and NVIDIA's $36bn AI-cloud commitments can protect receipts without measuring independent end-user demand; their overlap is unresolved. (rows 18, 51, row 44, row 45)
• SemiAnalysis's one-year H100 midpoint rose from $1.70 in October 2025 to $2.35 in March 2026, while its composite stayed at $2.80–2.84 over those months, down from $6.62 in 2H 2023 (−57.7%). The recovery is in term prices within a C-grade survey; it does not validate company-wide useful lives. (rows 16, 52–53)
• American Compute's October 2026 appraisal, using data through August, reports a typical realized used H100 SXM sale at $14,250 and projects 2028 residuals of 5–7% of original list conservatively, or 7–14% moderately. This C-grade vendor appraisal needs independent sale records; Silicon Data's modeled residuals are not completed sales. (row 54, row 30)
• Paid GPU-hours and comparable renewal prices remain the main missing denominators: CoreWeave's 98% committed-revenue share in Q2 2026 and SF Compute's September 9, 2026 average 25% resale rebate measure neither productive utilization nor fleet occupancy. (row 18, row 15)

Key claims

Rental units: USD/GPU-hour. “Observed” is the access date, not settlement. A grades the published statement/print, not market representativeness. Proposed thresholds are D judgments. Original sources accessed 2026-10-17; Claude-review corrections and new evidence re-fetched 2026-10-18, as specified below. Historical observations retain their original dates. The October 21 update re-fetched the linked Ornn market pages, H100/H200/B200 public histories and methodology, Silicon H100 page and Price of Compute display; new observations are separated from the October 17 information set. The October 25 update re-fetched the four OCPI market pages/public histories, Silicon H100 dated card and embedded chart, and Price of Compute; only new October 21–25 observations extend the record.

Daily settlements and recovered Silicon observations from data/gpu-index-daily.csv, re-confirmed against H100, H200, B200, A100 and Silicon card/chart on October 25. Units are USD per GPU-hour. October 17–20 were already recorded; October 21–24 extend the record. A dash means no observation added here, not zero or no publication. These are retrospectively retrieved publisher histories, not independent same-day captures; Saturday Silicon data carry no verified freshness/revision flag.

Analysis

Reading the market without mixing products

Rents are segmented by generation and contract terms; the evidence does not measure aggregate cluster availability. Silicon Data standardizes records for term, geography, platform and variant, then weights providers. Its weekly-recalibrated proprietary basis model is not externally reproducible. The portal required sign-in. The previously established H100 observation was SDH100RT $2.82 dated October 17, 2026; the summary panel's other GPUs were undated. On October 21 the H100 page showed readings temporarily unavailable. The October 25 fetch recovers explicitly dated October 21–24 values from its embedded chart and the October 24 card ($2.88). The same-date October 24 OCPI/Silicon H100 ratio is 1.049; sampling differences remain. Saturday chart dating and the stale FAQ prevent treating every recovered point as a verified fresh business-day publication (rows 1, 56, 64). (1, H100, 2, 42)

Ornn says OCPI uses active on-demand transactions, volume weighting and winsorization, excluding reserved contracts. Contributor identities and the winsorization parameter remain private. Its methodology makes the administrator's published benchmark primary evidence, with unverified market representativeness. Public history was retrieved despite documentation emphasizing API keys. Row 5 equally averages daily 4 p.m. ET observations of a rolling one-hour benchmark, not all monthly transactions; disruption rules permit stale carry-forward prints. Publication and quality flags are needed for fresh-observation coverage (row 47). (3, 4, 5, 43)

FastGPU records offers, not transactions or volumes. Its floor composite averages proportional changes geometrically; frozen fixings separately publish provider medians. The composite has much shorter history than its downloadable provider series. It averages endpoints over three days and plots a seven-day trailing average; the frozen daily fixings are a separate output (row 48). Historical observations in the CSV remain tertiary, reported unverified; direct current pages provide spot checks. (6, 7, 41, 44)

CoreWeave's spot discount buys interruptibility; Lambda's instance and cluster products carry different connectivity and availability promises. Shadeform's pricing page loaded without numeric catalogue data. SF Compute's wholesale price URL redirected to its homepage; Autoresearch retail rates are not wholesale clears. Vast's rendered cards do not establish a mode-specific transaction series. Inventory was not tested by ordering. Nebius's October 1 list increases of 16.9–21.0% across H100/H200/B200/B300, alongside separate cheap preemptible floors, strengthen the case for segmented asking-price power, without measuring transactions or occupancy (row 41). (Nebius pricing)

OCPI's H200 is an outlier: $6.34 on October 16 exceeds CoreWeave's $6.305, Nebius's $5.40 and Hyperstack's $3.99 on-demand lists and is 80.6% above Silicon's $3.51. H200/H100 is 2.31 on OCPI versus 1.02–1.25 across the four comparators (rows 56–57). Different contracts, contributor mix or thin coverage could explain this; none is established. Silicon's +6.0% weekly H200 move supports direction, not OCPI's level or its 17.8% October-to-16 mean increase. H200 market representativeness is D. (H100 API, H200 API, B200 API, A100 API, Silicon, CoreWeave, Nebius, Hyperstack)

On October 16/17 inputs, OCPI divided by CoreWeave's on-demand list is 0.94 for B200, 1.01 for H200, 0.45 for H100 and 0.35 for A100 (row 55). The older generations show the largest discounts. These cross-product ratios are useful monitoring signals, not matched-provider realized discounts: newer-generation ratios falling persistently would be more consistent with broad excess supply, while old-generation-only falls can reflect aging. B200 monthly means rose $6.14 August → $7.25 September → $7.74 October-to-17 → $7.87 October-to-20; its +15.5% 30-day change on October 16 hid a level almost unchanged from September 29. By October 20 B200 was $8.83 and H100 had rebounded to $3.12, with publisher weekly changes of +12.8% and −1.6%, respectively. By October 24 they were $9.57/$3.02, with weekly changes +12.9%/+8.6%; B200’s October-to-24 daily-print mean rose to $8.1338. These dated windows do not certify an all-history record. H100's historical daily log-return volatility was 4.6%, favoring monthly means over endpoint claims (rows 58–59). (H100 API, H200 API, B200 API, A100 API, list)

Contract economics, utilization and losses

The CoreWeave note's revenue/MW quotient combines quarterly revenue with ending active power. Ramping and hardware mix distort it. The Nebius note's $20–25m yield describes four landmark deals, with an overall Q2 average above $20m/MW-year; most associated capacity arrives late 2026 and contributes mainly in 2027. These numbers are not comparable tariffs. CoreWeave's subsequent short-term disclosure independently supports a premium for near-term clusters, but names no GPU generation; it cannot establish H100 or A100 contract economics. (18, 19, 20, 21)

Rate = annual revenue per MW × kW per GPU / (1,000 × 8,760) at full-year paid availability. Retrieved DGX maximum powers produce row 21's conditional examples. Networking, storage, headroom, PUE, actual draw and billing occupancy alter them. These are dimensional checks, not discovered customer tariffs; there is no evidence these exact DGX configurations underpin Nebius's deals. IREN's $9.7m/MW-year is a contract-term average with phased delivery, not a 2026 run-rate; its release lacks a GPU count, preventing an exact GPU-hour calculation. (specifications, B200 guide, IREN release) Ornn's reserve listings are indicative offers, not executed contracts.

Contracted occupancy, billed occupancy, uptime and productive utilization differ. Take-or-pay lets idle chips earn revenue. SF Compute's resale demonstrates spare customer hours without measuring fleet occupancy. The undelivered MW in row 23 are scheduled future supply; those disclosures alone do not show overdue power or stranded purchased GPUs. (15, 19, 26)

CoreWeave's Q2 2026 GAAP operating loss was $49m (−2%) despite a 59% adjusted EBITDA margin. D&A absorbed 54.1% of revenue ($1,393m), before $640m of net interest expense (24.9% of revenue); adjusted operating income was $128m (5%). A hypothetical six-to-four-year life change applied to all that D&A would add $696.5m per quarter, an upper-bound sensitivity that assumes all D&A is affected GPU equipment and ignores cohort/residual differences (row 51). These figures show little operating room for faster depreciation, while lifetime capital returns still require cohort-level replacement spending, financing cash flows and lifetime revenue. Prepayments finance construction but create delivery obligations. Equity absorbs disappointing renewals first; secured lenders depend on contracted cash flows and recoverable hardware, with unsecured and vendor lenders exposed according to claim priority. Customers risk interrupted service and lost prepayments. Soluna is a primary counterexample to universal GPU-rental profitability. (18, 19, 20, 25)

CoreWeave's September 18, 2026 convertible pricing adds unsecured financing exposure; planned capped calls and conversion rights mean its 2.875% coupon is not an all-in capital cost. The announcement alone does not verify closing (row 43). (Financing release)

NVIDIA's September 2025 agreement backstops residual unsold capacity covered by its CoreWeave contract through April 2032, subject to conditions (row 44). My inference is that this shifts some revenue risk to the supplier and weakens the link between paid capacity and independent final demand; it guarantees neither the entire fleet nor debt. NVIDIA separately discloses $36bn of AI-cloud commitments at July 26, 2026 (row 45); overlap is unresolved, so these amounts must not be added. (CoreWeave 8-K, NVIDIA 10-Q)

Depreciation and used equipment

The useful-life register and unit-economics note remain the central comparison. SemiAnalysis reports a 38.2% rise in its one-year H100 midpoint from October 2025 to March 2026, with typically 25% prepayment; April's range is $2.10–2.70, midpoint $2.40 (rows 16, 49). This is C-grade survey evidence of continued earning potential, not an independently audited renewal tape. (Dashboard) The same publisher's H100 composite stayed at $2.80–2.84 in October 2025–March 2026, versus $6.62 in 2H 2023 (−57.7%); only the one-year terms recovered. “Sold Out” appears for February–April 2026 in a table column whose term mapping remains uncertain (rows 52–53). GetDeploying's $1.84 mixed-variant A100 median now reproduces on a page dated October 18; the earlier $1.79 was from its October 2 page (row 26). Neither proves individual assets will earn their carrying values. Nebius's letter now directly confirms the register's previously unverified life extension. These policies cover broader equipment categories, not individually validated GPU economic lives; Amazon attributes its shortening to faster technological development. (16, 20, 26, 27, 28)

SXM modules require compatible hosts; servers include more than GPUs. Asking prices, completed sales and modeled valuations measure different things. American Compute's October 2026 vendor appraisal reports a typical used H100 SXM realized sale of $14,250 (25th percentile $13,200, limited sample), dealer asks at 1.16× realized and an HGX H100 node at $195,000. Its June forecast puts 2028 H100 residuals at 5–7% of original list conservatively or 7–14% moderately (row 54, C). This supplies a completed-sale-based claim and dated residual path, but the underlying ledger remains unverified. Opposing American Compute and Epoka-linked readings still need matched configurations, regions and sale records. (Appraisal) Silicon Data's residual product capitalizes projected rents: it cannot independently corroborate rental-driven hardware appreciation. (29, 30, 31, 32)

Rubin is already a production competitor: on September 30, 2026, CoreWeave reported limited deployment and up to 4.8× GB200 inference token throughput per GPU at matched interactivity for Cognition's workload (row 42). This vendor/customer result raises renewal and residual-value risk despite firmer older-card rents; it is not a universal price-performance ratio. (Deployment account)

The MiniMax M2.7 comparison's $0.190/GB200 versus $0.421/H100 per million tokens at 91 tok/s/user uses hyperscaler-volume ownership TCO ($1.86/$1.17 per chip-hour, July 2026 survey), not rental tariffs. The interpolated throughput is 2,725.1/771.8 tok/s/chip; the config says FP4 while the summary says fp8, with no software version or run date supplied (row 61). At the same throughput, H100's $2.74 OCPI rent implies $0.99/M tokens versus GB200's $10.50 CoreWeave list implying $1.07/M; both at CoreWeave list give H100 $2.22/M versus GB200 $1.07/M (2.07×). This is illustrative C-grade re-costing with mixed rental bases, not a replicated service-price comparison (rows 60–62). (Benchmark, pricing)

Supply and a fiber-style glut

Epoch's shipment estimates include uncertainty fields. Cumulative shipments precede deductions for warehouses, retirements and unavailable power; they are not energized installed capacity. H100-equivalent peak operations differ from workload throughput; the aggregates sum component estimates/medians rather than estimate an aggregate median. September's Rubin deployment makes successor throughput a current effective-supply consideration (row 42). SemiAnalysis's customer-level shipment/install-base tables were not publicly retrievable. TrendForce's server forecast counts a different denominator from its GPU mix forecast. (36, 37, 38, 39, 40)

NVIDIA's $0.4bn H200 inventory/purchase-obligation charge in H1 FY2027 amid China restrictions shows why rising OCPI H200 rents do not establish healthy inventory economics everywhere (row 46). Restricted inventory is not necessarily available to these rental markets. (NVIDIA 10-Q)

A fiber-style glut would combine persistent price declines with immediately available connected clusters, failed renewals, distressed sales and unpaid debt despite growing compute consumption. Cheap isolated hours or efficiency-driven token-price declines alone are insufficient. Row 39's thresholds are monitoring judgments, not universal cost floors.

Conflicts and open questions

October 25 evidence versus the prior note: Silicon numerical access has recovered: October 24’s dated card/chart is $2.88, superseding October 17’s $2.82 as the latest observed value without erasing the October 21 access failure. Its undated FAQ still says $2.53. The Saturday chart point also conflicts with assuming all observations follow the previously documented business-day calendar; prefer the dated display for price, keep freshness/coverage suitability C. B200’s $9.58 October 23 high supersedes the prior $8.83 displayed-window high; October 24’s verified +12.9% weekly change does not retroactively verify the earlier record-streak claim. Price of Compute’s H100 row and introduction show different ranges, so neither is a clean matched spread measure. No complete-month, utilization or collapse conclusion follows. (Silicon, B200, Price of Compute)

October 21 update versus the prior note: the earlier H100 reversal is no longer a sufficient description of the latest tape: OCPI recovered to $3.12 on October 20. B200's $8.83 high is bounded by the public July–October window; it does not establish a B300 or all-history record. Price of Compute's October 21 posted-price readings differ from OCPI's October 20 settlements in date, sample and method, so neither disagreement nor agreement identifies a market-wide rental move. Silicon's unavailable public display also conflicts with treating October 17's $2.82 as current. Preserve that historical print; recover publisher observations and quality flags before assessing coverage or invoking a fallback. Four daily observations establish no complete-month sequence. (H100, B200, Price of Compute, Silicon)

Original note versus Codex cross-check judgment (2026-10-17): the original Summary assigned B confidence to local cluster scarcity/no market-wide glut. The reviewer agrees the evidence does not demonstrate a glut, but judges it insufficient to measure aggregate availability or shortage: no inventory was ordered, and row 23 is scheduled future delivery. The applied view is segmented pricing with demonstrated premiums in selected contracts and unproven lifetime capital returns; supplier commitments and Rubin-driven throughput make the reviewer less comfortable inferring economic durability from rents alone. Both readings remain bounded interpretations, not measured market-wide supply balances.

The original row 16 graded the direction of H100 recovery B and said SemiAnalysis “directly verifies” it. The reviewer grades market levels and inferred recovery C because the observations behind this single survey are unavailable; A applies only to attribution, B to arithmetic. The observed survey rise is retained with C market confidence, plus its prepayment assumption and April range (rows 16, 49). Independent comparable primary contract data would resolve the grading disagreement. (Survey, dashboard)

The original row 23 treated undelivered power as an active-supply constraint; the reviewer reads the disclosed dates as a future pipeline without proof of delay. The original cash-margin/capital-destruction language also went further than the reviewer accepts: adjusted EBITDA and a GAAP loss alone do not establish operating cash generation or lifetime capital impairment. The narrower readings are applied; missed delivery dates and cohort-level cash returns remain open questions. (10-Q, results)

Vast's earlier H100 floor $1.73 and A100 median $0.65 (row 13) differ from the review's later October 17 display around 20:55 UTC: $1.33 and $0.80 (row 50), reproduced when applying this review. The earlier values lack a preserved timestamped snapshot; intraday movement versus transcription error remains unresolved. The CSV retains those original snapshot records as unverified historical provenance, not the later verified display. Hourly offer updates are not a transaction-price trend. (Vast pricing)

The mid-October “all-time highs” versus “price pressure” conflict partly dissolves on dates and products. The October 15 report repeats a purported B300 record streak; the original Compute Desk method/tape was not retrieved, so the record remains D. Thunder Compute's October-updated article says “This pricing pressure creates opportunities”; its cited AWS reduction is an older event. AWS's own announcement dates that cut. FastGPU's September publication instead says “prices did not fall.” None proves an October-wide collapse. (33, 34, 35, 45)

In the October 21 information set, there was also a contrary price move: OCPI H100 fell sharply after its mid-month peak, then rebounded from $2.74 on October 16 to $3.12 on October 20; its October-to-20 daily-print mean of $2.849 remained near September’s $2.829. A100’s October-to-16 mean weakened as newer cards strengthened; its later October 24 print rebounded to $1.08 (row 3). Thin contributor pools, changing mix and different interconnect promises can explain some cross-index differences, but constituent data are needed to quantify how much. Broad record-high claims fail both historical-comparability and cross-GPU tests. (3, 4, 5, 6, 7)

Claude versus earlier Codex cross-check (2026-10-18): Codex reported no settlement date on Silicon's public panel and C freshness. Claude found a separate H100 page explicitly dated October 17 at $2.82; that fetch now reproduces, so H100 is A as a dated print while other summary-panel GPUs remain C. The page's undated FAQ still says $2.53; the dated reading takes precedence. This resolves H100 dating, not freshness of underlying trading. (Panel, H100 page)

Codex accepted the OCPI prints at A and their representativeness at C, and judged newer-GPU firming supported. Claude reproduces every print but grades H200's market level D: its $6.34 is 81% above Silicon and above several posted lists, with H200/H100 2.31 versus 1.02–1.25 elsewhere. Both views are recorded; A print/B arithmetic survive, while D applies to H200 market representativeness and the size of its rise. Contributor-level configurations, volumes and matched executable quotes are needed to resolve this. B200's rising monthly means remain supported. (H100 API, H200 API, B200 API, A100 API, provider list, Silicon)

Codex's profitability conclusion was that 59% adjusted EBITDA and a $626m net loss cannot establish lifetime capital impairment. Claude agrees on that limit but emphasizes the −2% GAAP operating margin, D&A at 54.1% of revenue and modeled 2028 residuals of 5–14% of list as reasons for less comfort. The added operating line and C-grade residual appraisal strengthen the risk case without demonstrating cohort-level impairment (rows 51, 54). Likewise, Codex emphasized the C-grade one-year H100 recovery; Claude adds the same publisher's flat composite and its −57.7% earlier decline. The term recovery remains, with its narrower meaning explicit (rows 16, 52–53). (Results, dashboard, appraisal)

The earlier applied note could not reproduce GetDeploying's $1.84, having fetched $1.79 on an October 2 page. Claude's October 18 $1.84/47-provider page now reproduces directly. Cached web responses can still return October 17 or October 2 versions; this is a dated-page/coverage difference, not evidence of a matched-price move. The pooled variants cannot be compared directly with OCPI A100 SXM4 (row 26). (Live page, alternate)

Not retrieved or independently verified: Silicon portal history and the brief's historical low/rebound; original Compute Desk transactions; Shadeform dynamic catalogue; SF Compute wholesale order-book history; market-wide paid utilization; complete used-sale ledgers; SemiAnalysis shipment/install-base tables; exact GPU counts and net effective prices for the highlighted bilateral contracts.

Proposed P1 definition: retain SDH100RT's neocloud reading and the register's threshold. Average each complete calendar month's official daily publications equally; require the coverage in row 38 and test consecutive months wholly inside the forecast horizon. Use published corrections available at resolution, record methodology versions and publication/quality flags, and never forward-fill missing prices. Exclude publisher-carried-forward or disruption-affected prints from fresh-observation coverage; do not assume that a dated print represents fresh trading. No qualifying complete-month sequence is demonstrated here. Silicon’s previously retrieved methodology described business-day publication; the recovered Saturday chart point now requires calendar/freshness clarification (row 64). Its earlier index page advertised 90-day history behind its portal; its plans page separately lists 30 days for Basic and full history for Pro. OCPI publishes calendar-day observations; its public-3mo API returned 92 days (July 18–October 17) on October 18 and July 21–October 20 on October 21. These public windows do not provide a durable forecast-resolution archive (row 64). Prospective daily capture must preserve observation date, retrieval time, methodology and quality flags; missing archive days count as missing coverage. The October 21 update retained October 17–20 publisher histories without creating a prospective capture schedule. As of October 25, daily archiving uses scripts/archive-gpu-prices.sh under the update playbook, and October 21–24 observations are retained. This five-column capture still lacks intraday retrieval times and quality/revision flags; its Silicon series-name alias also needs normalization when evaluating coverage (row 68). Neither the script nor recovered chart certifies complete-month coverage or 60 fresh overlap pairs. (Access terms, panel, H100 API, H200 API, B200 API, A100 API)

If that series becomes unavailable, apply row 40's predeclared overlap-based conversion to OCPI. Use only fresh, quality-flagged observations for the overlap pairs and successor coverage; if freshness or sufficient overlap cannot be established, mark P1 unresolved rather than silently substitute an incomparable series. FastGPU floors are corroborating signals, not the successor: a handful of cheap hosts can move a minimum without repricing the contract book. This is a proposal only; the register was not edited.

P1 age interpretation and proposed P1b: At the historical October 17 SDH100RT baseline of $2.82, the $1.50 P1 threshold implied a 46.8% fall (row 63); this is not a distance calculated from the latest print. At roughly 6.4 years since the A100 family's May 2020 launch, its Silicon $1.53 and OCPI $0.94 readings are already near/below that threshold, while the H100 composite fell 57.7% from 2H 2023 to March 2026 (rows 52, 63). The 2028 vendor residual path reinforces aging risk, without proving a glut. P1 is substantially an aging test at the long horizon; a Yes alone cannot identify overbuild. The $1.50 threshold is 1.28× SemiAnalysis's $1.17 new-purchase owning TCO versus October 16's $2.74 rent at 2.34×; that estimate is not the sunk cost or marginal floor of a 2023-vintage fleet. (Launch, benchmark, appraisal)

Claude proposes P1b (row 65): a B200 complete-calendar-month daily mean below $4.50 for three consecutive months wholly inside the horizon, with ≥80% fresh calendar-day coverage, no interpolation and exclusion of carried-forward/disrupted prints. That is 62.1% of September's $7.245, a 37.9% fall. For the 24-month horizon, roll on October 17, 2027 to the newest NVIDIA data-centre GPU with ≥180 days of OCPI history, using 62% of its 2027-Q3 mean. If OCPI is unavailable for a complete month, use Silicon's corresponding series (initially SDB200RT, panel $6.12), scaling the threshold by the median OCPI/Silicon ratio over the preceding 90 overlapping days with ≥60 fresh pairs; express a Silicon threshold by dividing the OCPI threshold by that ratio. Otherwise void. Probabilities 3%/10%/20% at 6/12/24 months are Claude's indicative D judgments. P1 Yes/P1b No would favor aging; both Yes would strengthen glut evidence, subject to other signals. This remains a proposal: the register and scheduling are outside this edit's scope. (Claude review, H100 API, H200 API, B200 API, A100 API)

Signals to watch

Sources

Original access 2026-10-17. Sources 1, 5, 9, 12, 17, 18, 21, 28 and 46 re-fetched 2026-10-18 for this review; sources 51–58 accessed 2026-10-18. The October 21 update re-fetched Ornn H100/H200/B200 market pages and JSON histories, OCPI methodology, Silicon H100 availability and Price of Compute. The October 25 update re-fetched H100/H200/B200/A100 market pages and all four JSON histories, Silicon H100 card/embedded chart and Price of Compute. New H100/B200 partial-month means were checked against the API; the additional observations already present in the daily CSV were retained without duplicates. Exact means bridge the existing rental CSV with the 17–20 October daily CSV and were checked against the re-fetched histories. Dates below identify publication or observation. The linked Microsoft/Alphabet/Amazon/Fed/CoreWeave/NMED schedules were independently re-fetched October 25; these are existing future gates, not new realized events. SEC retrieval used the generic User-Agent research-bot contact@example.com.

1. Silicon Data, live GPU indices, undated: https://www.silicondata.com/products/silicon-index

2. Silicon Data portal, sign-in required: https://portal.silicondata.com/

3. Ornn Data, settled index panel, settlement 2026-10-16: https://data.ornn.com/

4. Ornn, H100 market/history, settlement 2026-10-16: https://data.ornn.com/markets/h100-sxm

5. Ornn public H100 JSON, through 2026-10-16; other GPU endpoint URLs in CSV: https://data.ornn.com/api/public-index/gpu/H100%20SXM/index-history

6. FastGPU fixings CSV, daily: https://fastgpu.co/api/v1/dataset/gpu-price-fixings.csv

7. FastGPU floor index, through 2026-10-16: https://fastgpu.co/price-index

8. Lambda pricing, live: https://lambda.ai/pricing

9. CoreWeave pricing, live: https://www.coreweave.com/pricing

10. Oracle USD price list, live, numbers read in Chrome: https://www.oracle.com/cloud/price-list/

11. RunPod pricing, live: https://www.runpod.io/pricing

12. Hyperstack pricing, live: https://www.hyperstack.cloud/gpu-pricing

13. Vast.ai live pricing, numbers read in Chrome: https://vast.ai/pricing

14. SF Compute Autoresearch pricing, live: https://autoresearch.sfcompute.com/

15. SF Compute, realized resale discount, 2026-09-09: https://sfcompute.com/news/realized-discount

16. SemiAnalysis, The Great GPU Shortage, 2026-04-02: https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity

17. SemiAnalysis pricing dashboard, history through Apr 2026: https://gpu-index.semianalysis.com/

18. CoreWeave Q2 results, SEC exhibit, 2026-08-11: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm

19. CoreWeave Q2 10-Q, period 2026-06-30: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm

20. Nebius Q2 shareholder letter, SEC exhibit, 2026-08-12: https://www.sec.gov/Archives/edgar/data/1513845/000110465926094568/tm2622968d1_ex99-2.htm

21. CoreWeave higher-price contracts, 2026-09-17: https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx

22. Nvidia DGX H100/H200 guide, undated: https://docs.nvidia.com/dgx/dgxh100-user-guide/introduction-to-dgxh100.html

23. Nvidia DGX B200 guide, undated: https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html

24. IREN–Microsoft announcement, 2025-11-03: https://irisenergy.gcs-web.com/news-releases/news-release-details/iren-secures-97bn-ai-cloud-contract-microsoft

25. Soluna annual-results presentation, March 2025, slide 15: https://www.solunacomputing.com/wp-content/uploads/2025/03/SLNH-10K-2025-PRESENTATION.pdf

26. CoreWeave FY2025 10-K, period 2025-12-31: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm

27. Amazon FY2025 10-K, period 2025-12-31: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm

28. GetDeploying A100 comparison, earlier page dated 2026-10-02; current direct fetch dated 2026-10-18: https://getdeploying.com/gpus/nvidia-a100

29. Caladan used-GPU asking-price survey, observation 2026-09-09: https://caladansemi.com/blog/used-nvidia-h100-a100-gpu-buying-guide-2026

30. American Compute resale index, October 2026 edition: https://www.amcompute.com/gpu-resale-index

31. Winston Su, Australia Data Centre Index, used-GPU report, 2026-10-07: https://certifiedstrategic.com/insights/used-gpu-resale-prices-australia-2026

32. Silicon Data residual-value methodology, undated: https://www.silicondata.com/products/gpu-residual-value

33. Carl Peterson, Thunder Compute market trends, updated 2026-10-01: https://www.thundercompute.com/blog/ai-gpu-rental-market-trends

34. AWS GPU price-reduction announcement, 2025-06-05: https://aws.amazon.com/blogs/aws/announcing-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/

35. Yash Thakker, explainx.ai, B300 record claim, 2026-10-15: https://explainx.ai/blog/nvidia-b300-gpu-rental-prices-record-high-17-day-streak-compute-desk-claimed-vs-verified-2026

36. Epoch AI shipment timelines CSV, updated 2026-10-14: https://epoch.ai/data/ai_chip_sales_timelines_by_chip.csv

37. Epoch AI methodology, current: https://epoch.ai/data/ai-chip-sales-documentation

38. TrendForce, Nvidia shipment mix, 2026-04-08: https://www.trendforce.com/presscenter/news/20260408-13003.html

39. TrendForce, AI server pipeline, 2026-10-14: https://www.trendforce.com/presscenter/news/20261014-13269.html

40. Global X, Memory Is the New Bottleneck, 2026 report: https://assets-cms.globalxetfs.com/Memory%20Is%20the%20New%20Bottleneck%20in%20AI%20Semiconductors%20%5BFinal%5D.pdf

41. FastGPU open dataset, live; attribution: FastGPU GPU cloud pricing dataset, CC BY 4.0: https://fastgpu.co/dataset

42. Silicon Data index methodology, 2026-08-27: https://www.silicondata.com/documents/silicon-data-index-methodology.pdf

43. Ornn OCPI methodology, initially published 2026-07-24: https://data.ornn.com/methodology

44. FastGPU data provenance, live: https://fastgpu.co/data-provenance

45. FastGPU, Are Prices Falling?, 2026-09-15: https://fastgpu.co/blog/gpu-cloud-pricing-trends-2026

46. Nebius GPU pricing, live page and search-cached version showing changes effective 2026-10-01: https://nebius.com/prices

47. CoreWeave, Cognition becomes first Vera Rubin NVL72 customer, 2026-09-30: https://www.coreweave.com/blog/cognition-becomes-first-customer-for-nvidia-vera-rubin-nvl72-on-coreweave-cloud

48. CoreWeave, convertible senior notes pricing, 2026-09-18: https://investors.coreweave.com/news/news-details/2026/CoreWeave-Prices-Upsized-3-7-Billion-Convertible-Senior-Notes-Offering/default.aspx

49. CoreWeave 8-K, NVIDIA capacity agreement, filed 2025-09-15, agreement 2025-09-09: https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm

50. NVIDIA Q2 FY2027 10-Q, period 2026-07-26: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm

51. Silicon Data dedicated H100 index, readings dated Oct 17 and Oct 21–24, latest card/chart fetched 2026-10-25: https://www.silicondata.com/products/silicon-index/h100

52. American Compute, H100 appraisal, October 2026 Rack Report (data through August); residual projections June 2026: https://www.amcompute.com/appraisals/h100

53. SemiAnalysis InferenceX, GB200 NVL72 versus H100, MiniMax M2.7, interpolated benchmark, ownership-cost survey July 2026; run date unstated: https://inferencex.semianalysis.com/compare/minimax-m27-gb200-vs-h100

54. Silicon Data access plans, current terms: https://www.silicondata.com/pricing

55. Ornn public history APIs, through 2026-10-17: H100; H200; B200; A100 SXM4. Prices fetched directly using the four model-specific endpoints.

56. NVIDIA, A100 in full production and shipping, 2020-05-14: https://nvidianews.nvidia.com/news/nvidias-new-ampere-data-center-gpu-in-full-production

57. Claude independent cross-check, 2026-10-18 (review judgments, not an independent source for market figures): /reviews/gpu-pricing-and-utilization-claude-xcheck

58. GetDeploying alternate A100 URL, redirects to current daily page; direct destination fetched 2026-10-18: https://getdeploying.com/reference/cloud-gpu/nvidia-a100

59. Ornn H200 settled market table, October 17–20/24 settlements, latest accessed 2026-10-25: https://data.ornn.com/markets/h200

60. Ornn B200 settled market table, October 17–24 settlements, latest accessed 2026-10-25: https://data.ornn.com/markets/b200

61. Price of Compute posted-price display, dated 2026-10-21/25, latest accessed Oct 25: https://priceofcompute.com/

62. Ornn A100 SXM4 settled market table, observation 2026-10-24, accessed Oct 25: https://data.ornn.com/markets/a100-sxm4