TheAIThe Blackwell Fact Table
Every figure from “How Much Blackwell Actually Exists?” with its source: shipments, the B300 slice, capex and anchor deals, manufacturing ceilings, Rubin, the H100 precedent — plus methodology notes and unit conventions.
· 17 min read
Source appendix for How Much Blackwell Actually Exists? (TheAI Cloud, August 2026). Every figure in the article, with its source. Tracker pages accessed Aug 18, 2026 unless noted; document dates are publication dates. Figures marked derived are our arithmetic on top of the cited sources.
Unit convention. A Blackwell package contains two silicon dies; NVIDIA counts dies as "GPUs" in marketing (Huang: "each GPU die is a GPU"). Every row states its unit. FY = NVIDIA fiscal year (FY26 = Feb 2025 – Jan 2026).
Pricing & the rentable market
| Claim in the article | Figure | Source | Link |
|---|---|---|---|
| Median public on-demand B300 price, and its rise since Nov 2025 | $7.87/GPU·hr; +57% (from $5.00) | getdeploying B300 index (live FAQ), Aug 18, 2026 | https://getdeploying.com/gpus/nvidia-b300 |
| Decomposition of the +57%: same-offer repricing, not index composition | ~50 of 57 points = same providers repricing (Enverge $5.00→$7.50, Verda $5.74→$7.50, Oracle flat); ~7 points = composition; Aug 18 median without hyperscaler listings $7.50 vs $7.85 with them | Wayback snapshots of the same page (Mar 9 / Jul 2 / Aug 5, 2026) + live Aug 18; per-provider prices parsed from each snapshot's embedded JSON | https://web.archive.org/web/20260309131533/https://getdeploying.com/gpus/nvidia-b300 |
| The repricing wave ran March → early July; like-for-like flat since | 9 of 10 providers listed on Jul 2 unchanged through Aug 18 (only Vast moved: $6.88→$7.69→$7.50); median peaked in early July | same snapshots | https://web.archive.org/web/20260702130054/https://getdeploying.com/gpus/nvidia-b300 |
| Providers repriced on their own price pages, too | Nebius B300: $6.10 (May 7) → $7.85 on-demand (Jul 12), +29%; Nebius B200 $5.50 → $7.15; Lambda 8×B200 $5.74 (Mar 10) → $6.69 (Apr 1); Lambda cluster pricing $4.62 → $8.87–9.86 (May 16) | Wayback captures of nebius.com/prices and lambda.ai/pricing | https://web.archive.org/web/20260712235400/https://nebius.com/prices |
| B200 like-for-like: broad, one-directional | of 12 providers listed since March, 8 raised (+13% to +77%), 4 flat, 0 cut; median $5.25 → $6.37 | Wayback getdeploying B200 snapshots (Mar 9 / Apr 13 / Jun 20 / Aug 13) + live | https://web.archive.org/web/20260309131340/https://getdeploying.com/gpus/nvidia-b200 |
| Countable on-demand B300 stock on open marketplaces: tens of GPUs | Vast.ai API: 7 machines / 48 GPUs listed, 0 rentable at snapshot; RunPod out of capacity on both tiers; Shadeform: 2 clouds; SF Compute: "B300s coming to SFC this fall" | Vast.ai public API + tracker pages, Aug 20, 2026 | https://console.vast.ai/api/v0/bundles |
| Hyperscaler GB300 is GA but reservation-first | Azure ND GB300 v6 GA Nov 18, 2025 ("tens of thousands of GB300 GPUs" deployed, scaling to "hundreds of thousands"); AWS P6-B300 GA Nov 18, 2025 (on-demand via account managers / Capacity Blocks), P6e-GB300 UltraServers GA Dec 2, 2025; Google A4X Max reservation-only | Microsoft Tech Community (Nov 18, 2025); AWS What's New (Nov 18 / Dec 2, 2025); Google Cloud docs | https://aws.amazon.com/about-aws/whats-new/2025/11/amazon-ec2-p6-b300-instances-nvidia-blackwell-ultra-gpus-available |
| Contracted GB300 fleets are an order of magnitude 10⁵ — and not merchant | Microsoft contracted ~200,000 GB300s from Nscale (up to $14B); CoreWeave "largely sold out of 2026 capacity" (Q1 FY26 call); Nebius capacity auction cleared "15% above the highest price ever charged for Blackwell chips" (Q2 call, Aug 12, 2026) | Nscale press release (Oct 15, 2025); CoreWeave Q1 FY26 call (May 7, 2026); Nebius Q2 2026 call | https://www.nscale.com/press-releases/nscale-microsoft-2025 |
| Providers listing B300 at all; with a public on-demand price; with live stock | 23 / 15 / 10 | getdeploying B300 index, Aug 18, 2026 | https://getdeploying.com/gpus/nvidia-b300 |
| B300 is the thinnest supply segment | providers listing: B300 23 < B200 27 < H200 37 < H100 53 | getdeploying per-GPU pages, Aug 18, 2026 | https://getdeploying.com/gpus/nvidia-h100 |
| Hyperscaler on-demand list prices for GB300-class capacity | AWS p6-b300 $17.80/GPU·hr ($142.42/node); OCI $15.00/GPU·hr | cloud-gpus.com + getdeploying, Aug 18, 2026 | https://www.cloud-gpus.com/ |
| Published reserved spread for B300 | 12-mo $5.30–7.94; 24–36-mo $4.15–4.84; 60-mo low $3.98 | getdeploying B300 index (reserved columns), Aug 18, 2026 | https://getdeploying.com/gpus/nvidia-b300 |
| Marketplace check: 34 public B300 listings, two with any term under 12 months, earliest availability Oct 8 (one 8-node lot) | our survey of a large B2B cluster marketplace, Aug 18, 2026 | TheAI Cloud research; methodology notes on request | — |
| "On-demand capacity sold out across all GPU types"; new Blackwell deployments quoting mid-2026+; capacity online through Aug–Sep 2026 already booked | Apr 2, 2026 | SemiAnalysis, "The Great GPU Shortage" | https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity |
| Universe of GPU clouds behind these listings | 209 providers tracked | SemiAnalysis ClusterMAX 2.0, Nov 6, 2025 | https://newsletter.semianalysis.com/p/clustermax-20-the-industry-standard |
Units & headline numbers
| Claim | Figure | Source | Link |
|---|---|---|---|
| A Blackwell package = 2 dies; NVIDIA counts dies as GPUs | "Blackwell is really two GPUs in one Blackwell chip… each GPU die is a GPU" (Huang, GTC, Mar 2025) | Rev.com keynote transcript | https://www.rev.com/transcripts/gtc-keynote-with-nvidia-ceo-jensen-huang |
| "6 million Blackwells shipped" (Oct 2025) — counted in dies ≈ 3.0M packages | 6M dies over the prior four quarters | TrendForce News coverage of GTC DC, Oct 29, 2025 | https://www.trendforce.com/news/2025/10/29/news-nvidia-dismisses-ai-bubble-concerns-reportedly-projects-500b-in-gpu-sales-from-blackwell-and-rubin/ |
| The "3.6M unit backlog" meme is a mislabeled slide about top-4 CSP orders for 2025 | 3.6M Blackwell ordered by AWS/Azure/GCP/OCI (vs 1.3M Hopper in 2024) — GTC slide, Mar 2025; "backlog" wording appears only in an unsourced syndicated piece (Dec 29, 2025) | io-fund (orders) + FinancialContent/TokenRing (the mislabel) | https://io-fund.com/semiconductors/data-center/nvidia-ceo-predicts-ai-spending-will-increase-300-percent-in-3-years |
What has shipped, 2024 → 2026
| Claim | Figure | Source | Link |
|---|---|---|---|
| Blackwell announced | Mar 18, 2024 (GTC) | NVIDIA newsroom | https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing |
| 2024 packaging rework delayed volume by months | CoWoS-L warpage, bridge-die redesign | SemiAnalysis, Aug 4, 2024 | https://newsletter.semianalysis.com/p/nvidias-blackwell-reworked-shipment |
| Morgan Stanley cut its Q4'24 Blackwell estimate mid-stream | 450k chips (Sep note) → ~250–270k implied (Oct note: "750–800k in Q1'25, almost 3× 4Q24") | Tom's Hardware via Yahoo, Sep 25, 2024; EconoTimes, Oct 9, 2024 | https://finance.yahoo.com/news/nvidia-expected-produce-450-000-145414205.html |
| First material Blackwell revenue | $11.0B in fiscal Q4 FY25 (Nov'24–Jan'25) — "the fastest product ramp in our company's history" (Kress) | NVIDIA Q4 FY25 press release, Feb 26, 2025 | https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2025 |
| GB200 NVL72 racks became the volume vehicle | up to 80% of GB-series deployments in 2025 | TrendForce, Dec 17, 2024 | https://www.trendforce.com/presscenter/news/20241217-12412.html |
| NVL72 racks shipped in CY2025 | ~27.3k (Dec 2025 estimate) to ~29k (Jul 2026 retrospective) ≈ 2.0M packages | Morgan Stanley via Longbridge (Dec 8, 2025) and BigGo (Jul 12, 2026) | https://longbridge.com/en/news/269012729 |
| Cumulative Blackwell shipped through Dec 31, 2025 | ~3.0–3.2M packages: ~2.0M in NVL72 racks + ~1.2M as HGX boards / DGX | amcompute synthesis (Huang 3.0M through Oct + MS 28k racks + Epoch AI median 3.4M), Apr 20, 2026 | https://www.amcompute.com/blog/blackwell-gpu-market-size-and-fragmentation |
| Independent chip-sales model backing the cumulative figure | median 3.4M | Epoch AI, "AI Chip Sales" dataset, updated Aug 11, 2026 | https://epoch.ai/data/ai-chip-sales |
| The inverted mid-2025 narrative: JPM saw Blackwell falling to 1.8M in 2026 | 5.2M (CY2025) / 1.8M (CY2026) forecast, units unspecified | JP Morgan note via TweakTown, Jun 30, 2025 | https://www.tweaktown.com/news/106116/nvidia-expected-to-ship-5-2m-blackwell-gpus-in-2025-1-8m-2026-and-7m-rubin/index.html |
| TrendForce revised Blackwell's 2026 share up as Rubin slipped | 61% → 71% of NVIDIA high-end shipments; Rubin 29% → 22%; high-end market +26% YoY | TrendForce press release, Apr 8, 2026 | https://www.trendforce.com/presscenter/news/20260408-13003.html |
| Morgan Stanley 2026 rack projection and mid-year run rate | 70–80k NVL72 racks (~2.6× 2025); ~8,000 racks/month in June 2026 | Morgan Stanley via BigGo, Jul 12, 2026 | https://finance.biggo.com/news/4123fb83-9dc9-4048-a8c2-b5401b81bc69 |
| Data-center revenue slope confirming the ramp | DC $62.3B (Q4 FY26, Nov–Jan) → $75.2B (Q1 FY27, Feb–Apr, +92% YoY) → $89.0B (Q2 FY27, May–Jul, +117% YoY); Q3 guided to $108B ±2% | NVIDIA press releases, Feb 25 / May 20 / Aug 26, 2026 | https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027 |
| "Ramp of our Blackwell 300 products… fastest product ramp in our company's history" | Q1 FY27 | CFO Commentary + FactSet transcript, May 20, 2026 | https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q127/Q1FY27-CFO-Commentary.pdf |
| 2026 tracking toward 5–6M+ packages | derived: 70–80k racks × 72 ≈ 5.0–5.8M packages in NVL72 alone, before HGX | our arithmetic on the Morgan Stanley projection | — |
| Monthly rack tape, 2026 | ~6,500 racks in Feb (+6% MoM; Hon Hai ~2.4k, Wistron 1.4–1.5k, Quanta 1.3–1.4k) → ~8,000 in June (+5% MoM) → ~8k plateau in July (est. from ODM monthly revenues); H1 total ~40–43k racks (derived) | Morgan Stanley notes via UDN (Mar 11, 2026) and BigGo/Yahoo TW (Jul 9–12, 2026) | https://udn.com/news/story/7253/9373911 |
| Hon Hai guidance for H2 | 2026 rack shipments to "more than double" vs 2025; Q3 "high double-digit" growth QoQ; cloud+networking >50% of revenue for the first time (51% in Q2) | Foxconn Q2 2026 earnings release/call, Aug 12, 2026 | https://www.foxconn.com/en-us/press-center/press-releases/latest-news/2095 |
| Independent cumulative count through spring 2026 | Epoch AI model: 4,763,288 packages through Mar 31, 2026 (B200-class 2.27M + B300-class 2.49M); 5,165,329 through ~Apr 26 (90% CI 4.4–6.06M) | Epoch AI "AI Chip Sales" dataset CSV (updated Aug 11, 2026; Blackwell rows generated Jun 3, 2026) | https://epoch.ai/data/ai-chip-sales |
| Cumulative Blackwell as of mid-August 2026 | ~7–7.5M packages (derived: Epoch base through April + Morgan Stanley rack run-rate + HGX tail); ~6M of them inside NVL72 racks (~82–84k racks cumulative). No one publishes a direct August figure | our arithmetic on the two rows above | — |
The B300 slice
| Claim | Figure | Source | Link |
|---|---|---|---|
| Blackwell Ultra specs | 288 GB HBM3e per GPU, up to 1,400W, dual-reticle | NVIDIA developer blog, Aug 22, 2025 | https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/ |
| First commercial GB300 NVL72 deployment | CoreWeave, Jul 3, 2025 | CoreWeave press release | https://www.coreweave.com/news/coreweave-becomes-first-cloud-provider-to-deploy-nvidia-gb300-nvl72-platform |
| GB300 surpassed GB200 in Blackwell revenue within one quarter | "GB300 crossed over GB200 and contributed roughly two-thirds of the total Blackwell revenue" (Kress, Q3 FY26) | FactSet corrected transcript, Nov 19, 2025 (p. 4) | https://s201.q4cdn.com/141608511/files/doc_financials/2026/q3/NVDA-Q3-2026-Earnings-Call-19-November-2025-5_00-PM-ET.pdf |
| "Leading architecture across all customer categories" | written CFO commentary, same quarter | NVIDIA CFO Commentary Q3 FY26, Nov 19, 2025 | https://s201.q4cdn.com/141608511/files/doc_financials/2026/Q326/Q3FY26-CFO-Commentary.pdf |
| Cumulative fleet skewed old at end-2025 | ~75% B200 / 25% B300 | amcompute, Apr 20, 2026 | https://www.amcompute.com/blog/blackwell-gpu-market-size-and-fragmentation |
| 2026 production runs 70–80% GB300 | GB300 = 70–80% of global AI rack shipments in 2026; ~55k of the year's 70–80k racks expected to be GB300 (Foxconn's chairman: 50–60k) | Taipei Times (Jan 20, 2026); Morgan Stanley/consensus via Longbridge (Dec 21, 2025) | https://www.taipeitimes.com/News/biz/archives/2026/01/20/2003850873 |
| B300 out-shipped B200 3:1 in Q1 2026 | Epoch model: ~993k B300-class vs ~347k B200-class packages in Q1 CY2026 | Epoch AI dataset CSV | https://epoch.ai/data/ai-chip-sales |
| B300 is the majority of the installed base by August 2026 | ~3.5–4.5M packages ≈ 50–60% of the cumulative fleet (derived from the rack mix and Epoch trajectory) | our arithmetic | — |
Demand: capex & anchor deals
| Claim | Figure | Source | Link |
|---|---|---|---|
| Big-4 capex, 2025 actual | $413B total: Amazon $131.8B; Microsoft ≈$118B (derived: sum of four calendar quarters); Alphabet $91.4B; Meta $72.2B | Amazon Q4'25 release (cash-flow statement); Microsoft quarterly disclosures via DCD/SAMexpert; Alphabet Q4'25 call; Meta Q4'25 call | https://s2.q4cdn.com/299287126/files/doc_earnings/2025/q4/earnings-result/AMZN-Q4-2025-Earnings-Release.pdf |
| Amazon 2026 guidance, raised in July | ~$220B cash capex ("the higher cost of memory pushing this number up from our prior estimate of about $200 billion" — Jassy) | Q2 2026 earnings call transcript, Jul 30, 2026 | https://www.investing.com/news/transcripts/earnings-call-transcript-amazon-tops-q2-2026-estimates-as-aws-growth-accelerates-93CH-4826442 |
| Alphabet 2026 guidance, raised twice | $175–185B (Feb) → $180–190B (Apr) → $195–205B (Jul) | Q4'25 / Q1'26 / Q2'26 earnings calls | https://www.investing.com/news/transcripts/earnings-call-transcript-alphabet-beats-q2-2026-estimates-shares-fall-on-capex-surge-93CH-4807140 |
| Meta 2026 guidance, raised twice then narrowed up | $115–135B (Jan) → $125–145B (Apr) → $130–145B (Jul) | Q4'25 transcript PDF; Q1'26; Q2'26 calls | https://s21.q4cdn.com/399680738/files/doc_financials/2025/q4/META-Q4-2025-Earnings-Call-Transcript.pdf |
| Microsoft 2026 | ~$190B CY2026 stated in April; ~$175B after finance→operating lease reclassification (not a demand cut — Hood, Q4 FY26 call) | Yahoo Finance (Apr 30, 2026); Motley Fool Q4 FY26 transcript (call Jul 29, 2026) | https://www.fool.com/earnings/call-transcripts/2026/08/07/microsoft-msft-q4-2026-earnings-call-transcript/ |
| Combined 2026 guidance as of August | ~$720–760B (derived: sum of current ranges) | our arithmetic on the four rows above | — |
| NVIDIA's top five customers ≈ half of DC revenue | "little over 50%" (Kress) | FactSet transcript Q4 FY26, Feb 25, 2026 | https://s201.q4cdn.com/141608511/files/doc_financials/2026/q4/NVDA-Q4-2026-Earnings-Call-25-February-2026-5_00-PM-ET.pdf |
| OpenAI–NVIDIA | ≥10 GW of NVIDIA systems; up to $100B invested progressively | NVIDIA newsroom, Sep 22, 2025 | https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems |
| OpenAI–AWS | $38B / 7 years; "hundreds of thousands" of GB200/GB300 | OpenAI announcement, Nov 3, 2025 | https://openai.com/index/aws-and-openai-partnership/ |
| Anthropic–Microsoft–NVIDIA | $30B of Azure compute; up to 1 GW on Grace Blackwell / Vera Rubin | Microsoft blog, Nov 18, 2025 | https://blogs.microsoft.com/blog/2025/11/18/microsoft-nvidia-and-anthropic-announce-strategic-partnerships/ |
| DOE Solstice | 100,000 Blackwell GPUs | NVIDIA newsroom, Oct 28, 2025 | https://nvidianews.nvidia.com/news/nvidia-oracle-us-department-of-energy-ai-supercomputer-scientific-discovery |
| Amazon on 2027 capacity | "The lion's share of capacity in 2027 is largely reserved, and we have quite a bit of capacity that's already been reserved for 2028" (Jassy) | Q2 2026 earnings call, Jul 30, 2026 | https://www.investing.com/news/transcripts/earnings-call-transcript-amazon-tops-q2-2026-estimates-as-aws-growth-accelerates-93CH-4826442 |
Manufacturing ceilings
| Claim | Figure | Source | Link |
|---|---|---|---|
| TSMC CoWoS capacity | 70–80k wafers/month (2025) → 130–140k targeted by end-2026 | TrendForce, Dec 8, 2025 + Aug 5, 2026 | https://www.trendforce.com/news/2025/12/08/news-tsmcs-cowos-l-s-reportedly-fully-booked-osat-partners-step-up-with-ases-cowop-in-focus/ |
| Entire CoWoS lineup fully booked | CoWoS-L and CoWoS-S booked by NVIDIA, Google, Amazon, MediaTek | TrendForce, Dec 8, 2025 | https://www.trendforce.com/news/2025/12/08/news-tsmcs-cowos-l-s-reportedly-fully-booked-osat-partners-step-up-with-ases-cowop-in-focus/ |
| NVIDIA holds roughly half of CoWoS | ~half of 2025 capacity (TrendForce PR); 800–850k wafers reserved for 2026, >50% (Aug 2026 reports) | TrendForce, May 30, 2024 + Aug 5, 2026 | https://www.trendforce.com/presscenter/news/20240530-12154.html |
| TSMC CEO on the gap | "The capacity is very tight… probably this year, next year, we have to work extremely hard to narrow the gap" (C.C. Wei, Jan 2026); "our packaging capacity is so tight that now it's limiting my customers' growth" (Jul 2026) | TSMC Q4'25 transcript PDF; Motley Fool Q2'26 transcript | https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-01/51d09df96cd89ac19d65af39032b038dc2896a24/TSMC%204Q25%20Transcript.pdf |
| HBM sold out through 2026 | SK hynix: full 2026 DRAM+NAND demand secured (Oct 29, 2025); Micron: "Memory production for calendar 2026 is sold out, including HBM4" (Dec 18, 2025) | SK hynix newsroom; Micron via Blocks & Files | https://news.skhynix.com/en/sk-hynix-announces-3q25-financial-results/ |
Rubin
| Claim | Figure | Source | Link |
|---|---|---|---|
| Full production declared at CES | "NVIDIA Rubin is in full production, and Rubin-based products will be available from partners the second half of 2026" | NVIDIA newsroom, Jan 5, 2026 | https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer |
| First VR200 samples shipped | week of Feb 25, 2026 (Kress) | Motley Fool Q4 FY26 transcript | https://www.fool.com/earnings/call-transcripts/2026/02/25/nvidia-nvda-q4-2026-earnings-call-transcript/ |
| SEC filing language | "We expect our Rubin platform to start shipping in the second half of fiscal year 2027" (= Aug 2026 – Jan 2027) | NVIDIA 10-Q, filed May 20, 2026 | https://www.sec.gov/Archives/edgar/data/1045810/000104581026000052/nvda-20260426.htm |
| Production starts in fiscal Q3; supply constrained | "We will start in Q3" (Kress); "My sense is that we'll be supply constrained throughout the entire life of Vera Rubin" (Huang) | Q1 FY27 earnings call, May 20, 2026 | https://stockanalysis.com/stocks/nvda/transcripts/568907-q1-2027/ |
| 2026 volume consensus | ~250–350k packages; Kuo's 5–7k racks in H2'26 imply up to ~500k | NextBigFuture (Jul 21, 2026); Kuo via Investing.com (Jan 6, 2026) | https://www.nextbigfuture.com/2026/07/hbm4-and-fab-limits-prevent-1000-vera-rubin-racks-per-day-in-2026-or-2027.html |
| Rubin's 2026 share cut as HBM4/CX9/cooling slipped | 29% → 22% | TrendForce, Apr 8, 2026 | https://www.trendforce.com/presscenter/news/20260408-13003.html |
| NVIDIA raised the HBM4 spec mid-stream | >11 Gbps/pin required in 3Q25; mass production pushed to end-1Q26 | TrendForce, Jan 8, 2026 | https://www.trendforce.com/presscenter/news/20260108-12869.html |
| VR200 NVL72 rack power | 190 kW (Max-Q) – 230 kW (Max-P); GPU TDP up to 2,300W vs 1,000–1,400W Blackwell | Kuo via Investing.com (Jan 6, 2026); SemiAnalysis (Feb 25, 2026) | https://newsletter.semianalysis.com/p/vera-rubin-extreme-co-design-an-evolution |
| 100% liquid-cooled, including HGX form factor | no air-cooled variant | SemiAnalysis (Feb 25, 2026); ModulEdge (Jun 7, 2026) | https://www.moduledge.com/blog/nvidia-vera-rubin |
| Greenfield data centers for 200kW+ racks | 18–36 months from design freeze to energization | ModulEdge, Jun 7, 2026 | https://www.moduledge.com/blog/nvidia-vera-rubin |
| Independent clouds quoted mid-2027 | first VR waves go to hyperscalers/large NCPs; e.g. Civo locked 2,016 VR GPUs for Q1 2027 | Civo blog, Jun 2, 2026 | https://www.civo.com/blog/blackwell-sold-out-in-weeks-heres-what-rubin-demand-will-look-like |
Q2 FY27 update (reported Aug 26, 2026)
| Claim | Figure | Source | Link |
|---|---|---|---|
| Q2 FY27 results | Revenue $96.2B (+106% YoY, +18% QoQ); data center $89.0B (+117% YoY), "driven by the ramp of our Blackwell Ultra infrastructure"; gross margin 75.0%; consensus was $91.85B | NVIDIA press release + CFO Commentary, Aug 26, 2026 | https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q227/Q2FY27-CFO-Commentary.pdf |
| First hyperscaler-vs-neocloud revenue split | Hyperscale $48.7B (+102% YoY) vs "AI Clouds, Industrial & Enterprise" $40.3B (+138% YoY); ACIE "includes our neocloud, industrial, and enterprise customers" (Kress) | CFO Commentary + earnings call transcript, Aug 26, 2026 | https://stockanalysis.com/stocks/nvda/transcripts/676136-q2-2027/ |
| Vera Rubin production shipments commenced | "We commenced production shipments of Vera Rubin earlier this month" — with purchase orders "from every major hyperscaler, AI cloud, and system OEM" | earnings call transcript (Kress), Aug 26, 2026 | https://stockanalysis.com/stocks/nvda/transcripts/676136-q2-2027/ |
| Rubin guided to ~20% of data-center revenue in fiscal Q3 (Aug–Oct 2026) | "We see Vera Rubin accounting for about 20% of data center revenue in Q3" | earnings call transcript (Kress) | https://stockanalysis.com/stocks/nvda/transcripts/676136-q2-2027/ |
| Neocloud installed capacity, per NVIDIA | ACIE cohort expected to exit 2026 with ~8 GW installed, up from ~3 GW at end-2025; AWS "deploying an additional 2 million GPUs starting this quarter through Q2 FY2029"; OpenAI commitments ≈ 12 GW of NVIDIA compute | earnings call transcript (Kress) | https://stockanalysis.com/stocks/nvda/transcripts/676136-q2-2027/ |
| PORTS-Pike guarantee: NVIDIA's own site-density figure | up to $105B of guarantees for a 4.25 GW SB Energy campus (Ohio) leased exclusively to OpenAI for 20 years; "each generation of NVIDIA infrastructure deployed at PORTS-Pike could represent approximately 1.5 million NVIDIA GPUs, or approximately $150 billion to $200 billion in NVIDIA revenue" | CFO Commentary (Guarantees section), Aug 26, 2026 | https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q227/Q2FY27-CFO-Commentary.pdf |
| Supply commitments more than doubled on memory | $119B → $279B in one quarter, "primarily related to the procurement of memory"; Kress: "extreme pricing conditions in memory" | CFO Commentary + call transcript, Aug 26, 2026 | https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q227/Q2FY27-CFO-Commentary.pdf |
| China in Q2 FY27 | Hopper shipments to China <1% of data-center revenue; no China DC compute in the outlook | CFO Commentary, Aug 26, 2026 | https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q227/Q2FY27-CFO-Commentary.pdf |
The H100 precedent
| Claim | Figure | Source | Link |
|---|---|---|---|
| Peak and collapse | market rates went $4.70 → $8+/hr in H2 2023; auction slots $1–2 by Aug 2024 | Eugene Cheah, "$2 H100s: How the GPU Bubble Burst", Oct 11, 2024 | https://www.latent.space/p/gpu-bubble |
| Composite index across the cycle | $6.62 (2H23) → $3.91 (4Q24) → $2.80 (Mar 2026) | SemiAnalysis H100 pricing index (live) | https://gpu-index.semianalysis.com/ |
| One-year contract bottom, then rebound | $1.70/hr (Oct 2025) → $2.35 (Mar 2026), +40%; on-demand sold out again | SemiAnalysis, Apr 2, 2026 | https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity |
| Card prices, peak vs today | ~$40k (eBay, 2023) → $6–15k secondary (2026) | CNBC via Slashdot (Apr 14, 2023); CloudZero (May 20, 2026) | https://www.cloudzero.com/blog/h100-gpu-cost/ |
| "H100 prices are melting up" | Mar 28, 2026 | Latent.Space AINews | https://www.latent.space/p/ainews-h100-prices-are-melting-up |
| Current H100 median for scale | $3.37/GPU·hr on-demand median (+12% YoY) | getdeploying H100 index, Aug 18, 2026 | https://getdeploying.com/gpus/nvidia-h100 |
Methodology notes
- Medians are computed per provider (each provider's cheapest public on-demand listing), not per configuration; mixing methods moves the figure ±10%. Tracker medians shift week to week as stock statuses change.
- Tracker FAQ percentages are unstable: getdeploying's own FAQ read "+98% to $9.91" (Jul 2), "+66% to $8.31" (Aug 5), and "+57% to $7.87" (Aug 18) across snapshots. We compute all changes from the per-provider listing tables in each snapshot, not from the FAQ. The November 2025 baseline ($5.00) exists only as the tracker's own claim — no archive snapshot of the B300 page predates Mar 9, 2026.
- Our marketplace survey (34 listings / 2 short-term / earliest Oct 8) is a manual count on one large B2B cluster marketplace on Aug 18, 2026; we exclude our own listings from all medians quoted above (they change the median by $0.02).
- Derived figures (marked) are our arithmetic on cited sources; assumptions stated inline. Corrections: research@theai.com.