Oracle's Nvidia GB200s for Abilene

Oracle fills the Abilene campus with Nvidia GB200 systems it runs for OpenAI: 'more than 450,000' GPUs by Ellison's count, about 400,000 by a press report; neither discloses what Oracle pays.

A deal between Oracle, Nvidia, OpenAI

Why it matters

Chips absorb much of Stargate's spending and are its least disclosed cost. No Oracle or Nvidia filing ties the other to these chips, and the only reported price is the Financial Times' $40B, which one estimate halves. Oracle's delivery record includes 131,000 GPUs handed over in the three months to August 2026.

Tap a figure for the line in its source.

Oracle gives

  • payment for the GB200 systems; no party discloses it — the FT reported ~$40B, which one research group's arithmetic halves → Nvidia
  • GPUs delivered to the customer at Abilene in Jun–Aug 2026 alone, 1.9x the quarter before → OpenAI

Nvidia gives Oracle

  • GB200 systems for the eight Abilene buildings: 'more than 450,000' GPUs in all by Ellison's count, about 400,000 by the FT's, a figure OpenAI repeats

Mar 7, 2025 reported by Bloomberg (64,000 GB200s by end-2026)updated Nov 20, 2026six of eight buildings delivered disclosed by the parties

What happened

The Abilene campus uses Nvidia's GB200 systems to serve OpenAI on Oracle's cloud. OpenAI says it trained GPT-5.5 there. Larry Ellison puts the finished campus's count at 'more than 450,000'. A May 2025 Financial Times report put it at about 400,000, which OpenAI later repeated. Oracle says it handed over 131,000 of them at Abilene in the three months to August 2026 alone. Neither Oracle nor Nvidia has filed what Oracle pays for the chips. The only reported price, about $40 billion, comes from the same Financial Times report. One research group estimates the same chips cost nearer $20 billion.

Sources 14 documents — 4 company releases · 3 transcripts · 3 press reports · 1 SEC filing · 1 other document · 1 research note · 1 social post

Research trail Found 384 documents and read 356 of them while checking this deal and related ones.

  1. company releaseopenai.com
    States platform
  2. SEC · SEC filingsec.gov
    States oracle filing
  3. Motley Fool · transcriptDec 10, 2025fool.com
    States delivered dec 2025
  4. Stock Analysis · transcriptstockanalysis.com
    States delivered Q1 fy27
  5. company releaseoracle.com
    Cited by the record; no single figure is quoted from it.
  6. company releasecdn.openai.com
    Cited by the record; no single figure is quoted from it.
  7. transcriptMar 10, 2025fool.com
    Cited by the record; no single figure is quoted from it.
  8. company releasecrusoe.ai
    Cited by the record; no single figure is quoted from it.
  9. press reportdatacenterknowledge.com
    Cited by the record; no single figure is quoted from it.
  10. press reportspyglass.org
    Cited by the record; no single figure is quoted from it.
  11. press report (archived copy)datacenterdynamics.com
    Cited by the record; no single figure is quoted from it.
  12. other documentsamaritan-research.org
    Cited by the record; no single figure is quoted from it.
  13. research noteepoch.ai
    Cited by the record; no single figure is quoted from it.
  14. social postpublish.x.com
    Cited by the record; no single figure is quoted from it.

The record's own summary of its sourcing: OpenAI post (Apr 29 2026) + Korea blueprint (Oct 2025); Oracle e-book (Oct 2025), Q3 FY25 / Q2 FY26 / Q1 FY27 calls and FY2026 10-K; Crusoe release (Mar 2025); Bloomberg via Data Center Knowledge, the FT via Spyglass and DCD, Samaritan Research and Epoch AI for the contested cells (read 2026-10) · as of Sep 2026

The evidence 4 sourced figures · 2 disputes

The record's headline Oracle's Nvidia GB200s for Abilene: 'more than 450,000' GPUs by Ellison's count, ~400,000 by the FT's, which OpenAI repeats; 131,000 delivered in Jun–Aug 2026 alone — no filing by either company ties the other to these chips or gives a price

Size
GPU count 400,000–450,000+ (the FT's figure, which OpenAI repeats, vs Ellison's) · price undisclosed (FT: ~$40B)
When
Mar 2025 (first reported: 64,000 by end-2026) → May 2025 (FT: ~$40B for ~400,000) → Oct 2025 ('more than 450,000', Ellison) → Jun–Aug 2026 (131,000 delivered)
Status
six of eight buildings delivered (Sep 2026); 131,000 GPUs handed over in Oracle's Q1 FY2027
Structure
the GPU systems run on Oracle Cloud Infrastructure inside buildings Oracle leases from the Crusoe–Blue Owl venture (sg-crusoe-jv), and Oracle sells their computing to OpenAI (sg-oai-orcl). That Oracle buys them from Nvidia is the FT's report ('lease the computing power to OpenAI'); no purchase agreement, price or delivery schedule between the two is disclosed
Programme
Stargate
Parties
Oracle operator of the GPUs on its cloud; their buyer, by the FT's report · Nvidia maker of the GB200 systems · OpenAI the customer the GPUs are delivered to — it trains and serves its models on them
Site
Stargate Abilene
Record
disclosed / primary source · as of Sep 2026

Flows 2 as the record states them

  • Oracle → Nvidiachip-purchasesundisclosed — ~$40B for ~400,000 GB200s by the FT's May 2025 report, ~$20B by Samaritan Research's per-GPU arithmetic; no Oracle or Nvidia filing ties the other to these chips
  • Nvidia → OraclegpusGB200 systems for the eight Abilene buildings: 'more than 450,000' GPUs (Ellison) vs ~400,000 (OpenAI, citing the FT)

Resolved figures 4 each with the verbatim line from its document

  • platform
    OpenAI's own description of its flagship: the Abilene site 'operates on Oracle Cloud Infrastructure and runs NVIDIA GB200 systems', and trained GPT-5.5 (Apr 29 2026). OpenAI buys the computing from Oracle (sg-oai-orcl); neither company says who owns the hardware — the FT reported that Oracle buys it
    “Our latest and smartest model yet, GPT‑5.5, was trained at our flagship Stargate site in Abilene, Texas. The site operates on Oracle Cloud Infrastructure and runs NVIDIA GB200 systems.”
    OpenAI, 'Building the compute infrastructure for the Intelligence Age' (Apr 29 2026) · openai.com
  • delivered-dec-2025more than 96,000
    more than 96,000 GB200s delivered at Abilene by Dec 10 2025, by Oracle's own account on its Q2 FY2026 call — a running total, in Motley Fool's transcript (the only one on file)
    “Our super cluster in Abilene, Texas is on track more than 96,000 NVIDIA Grace Blackwell GB 200 delivered.”
    Oracle Q2 FY2026 earnings call (Motley Fool transcript, Dec 10 2025) · fool.com
  • delivered-q1-fy27131,000
    131,000 GPUs delivered at Abilene in Oracle's Q1 FY2027 (Jun–Aug 2026), '1.9 times the volume delivered in Q4' — so roughly 69,000 in Mar–May 2026, by our division. Deliveries in a quarter, not an installed total: there is no figure on file for Dec 2025–Feb 2026, so the quarters cannot be summed into a campus count. The same call: six of the eight buildings, 618 MW, delivered to the customer (sg-oai-orcl)
    “We delivered 131,000 GPUs there in Q1, 1.9 times the volume delivered in Q4.”
    Oracle Q1 FY2027 earnings call (StockAnalysis transcript, Sep 10 2026) · stockanalysis.com
  • oracle-filing
    what Oracle's filings say about its GPU supply: the FY2026 10-K, that it 'at times' accepts 'less favorable terms with suppliers to minimize supply constraints', naming no supplier; its furnished Dec 2025 earnings release, Ellison saying Oracle will keep buying NVIDIA's latest GPUs, with no site, volume or price. The 10-K never names Nvidia, and Nvidia's Q2 FY2027 10-Q never names Oracle (in the text on file), so neither side files this purchase, its price or its volume. Oracle's filings also name NVIDIA as one of its cloud customers, in its Mar and Dec 2025 earnings releases and its Feb 2026 funding plan (all three cited on sg-oai-orcl: oracle-filings, rpo-q2-fy26, funding-2026)
    “For example, industry supply capacity for AI accelerators, including graphics processing units, as well as memory devices, is competitive, and we at times have to accept less favorable terms with suppliers to minimize supply constraints.”
    Oracle FY2026 10-K (risk factors) · sec.gov

Contested figures 6 not reconciled — every position side by side, with its source and tier

gpus-at-full-build4 positions, side by side
'more than 450,000 NVIDIA GB200 GPUs and over 1.2 billion watts in power capacity' — Larry Ellison's count, in Oracle's own AI World e-book (Oct 2025) Oracle e-book, 'AI is Changing the Game'
'approximately 400,000 of Nvidia's next-generation GPUs' on 'over 1.2 GW of power' — OpenAI's Korea economic blueprint (Oct 2025), footnoted to the FT and to OpenAI's own launch post, which gives no GPU count — so, by its own citations, the FT's figure repeated. Crusoe's design figure points the same way only after correction: each building is 'designed to operate up to 50,000 NVIDIA GB200 NVL72s' — an NVL72 is a 72-GPU rack, so read literally that is 3.6 million GPUs a building; read as 50,000 GB200s, eight buildings make 400,000. The Information's Anissa Gardizy (Mar 2026): '~400k Nvidia Blackwell GPUs' OpenAI blueprint (Oct 2025); Crusoe release (Mar 18 2025); The Information reporter on X
403,200 physical GPUs — 100,800 B200 in the first two buildings and 302,400 B300 in the other six — which Epoch counts in service from Nov 1 2026 (as fetched Nov 2 2026; 201,600 then and 403,200 projected for Q4 2026 on its explorer page as read Oct 7 2026), a step past the six buildings Oracle has confirmed. Epoch's table marks only the first two buildings' chips as company disclosure, citing Crusoe's impact report, which gives the design figure quoted with the FT's count — 'up to', a ceiling, not a count installed. The other six buildings' chip and count are its own estimate (the same design with slightly larger chillers, and the timing of that chip's sales ramp) — a satellite-and-permit model, not a disclosureEpoch AI, data_center_chip_quantities.csv, "OpenAI Stargate Abilene" — 2025-12-24: B200, 100800 units (Company disclosure) — 2026-11-01: B300, 302400 units (Estimate) Epoch AI Data Centers hub (data_center_chip_quantities.csv, snapshot 2026-11-19) + satellite explorer (read Oct 2026); Crusoe Impact Report 2024, as Epoch cites it · epoch.ai
the early plan, since overtaken: 64,000 GB200s by the end of 2026, 16,000 of them by summer 2025 (Bloomberg, Mar 7 2025, one person familiar); Ellison on Oracle's Mar 10 2025 call: 'We are in the process of building a gigantic 64,000 GPU liquid-cooled NVIDIA GB200 cluster for AI training' — a site he did not name Bloomberg via Data Center Knowledge; Oracle Q3 FY2025 call (Motley Fool transcript)
purchase-value2 positions, side by side
about $40B: the FT (May 24 2025, Tabby Kinder and George Hammond, as quoted verbatim by Spyglass that day) — 'Oracle will spend about $40bn on Nvidia's high-performance computer chips to power OpenAI's new giant US data centre' and 'would purchase about 400,000 of Nvidia's GB200 chips … and lease the computing power to OpenAI, according to several people familiar with the matter'. DCD relays it as 'in talks to spend as much as $40bn'; OpenAI's own Korea blueprint footnotes its ~400,000 figure to the same story ('Oracle to buy $40bn of Nvidia chips for OpenAI's new US data centre') and states no price itself. One report, repeated — confirmed by neither Oracle nor Nvidia FT via Spyglass (May 24 2025) and DCD (Wayback); OpenAI blueprint bibliography
about $20B: Samaritan Research finds the $40B 'inconsistent with the reported chip count, given that GB200 racks (including networking) typically cost <$50,000 per GPU for a total cost of approximately $20 billion' — an estimate from a per-GPU price, not a report Samaritan Research, AI chip owners methodology

Scope note what this record leaves to other records

Not this record: the buildings and their owners (sg-crusoe-jv); Oracle's contract with OpenAI, its backlog, capex and the campus's delivered megawatts (sg-oai-orcl); Epoch's estimates of the site's compute and capital cost today (the abilene entity); Oracle's company-wide deliveries ('850 megawatts of AI capacity containing more than 300,000 GPUs' in Q1 FY27, all sites) and its AMD MI355 capacity, which no source places at Abilene. Not counted: Samaritan's estimate that Oracle bought 7–8% of Nvidia's AI chips in 2025 (a model of Oracle-wide purchases, not this campus). Not adopted: 'Oracle is spending $40 billion on purchasing NVIDIA's high-end GPUs' (the Public Enterprise report 'Bubble or Nothing', Nov 2025 — a reported plan restated as a fact). Dropped from the previous version: the datacenterfrontier.com page cited for the $40B, which does not state it, and a thesis about TSMC packaging that no source on this record supports.

More between these companies

  1. Jul 2026 updated Nov 21
  2. Sep 22, 2025 updated Nov 18
  3. Jul 2, 2025 updated Nov 21
  4. May 22, 2025 updated Nov 21
  5. Jan 21, 2025 updated Nov 21

Everything between Oracle and OpenAI → · Everything between Nvidia and OpenAI → · Everything between Oracle and Nvidia →