Source Report 5

AI Lab Compute Commitments: Cloud and Chip Deals

Full research prompt

Estimate the roughly $1 trillion in AI lab compute commitments (OpenAI, Anthropic, xAI, Meta, others) by counterparty (Oracle, Microsoft, AWS, CoreWeave, Google, Broadcom, AMD, Nvidia). Estimate how much is take-or-pay or non-cancellable versus conditional or contingent on funding, using 10-K/10-Q remaining performance obligations, lease disclosures, and credit-rating agency analysis. Estimate how much of the reported revenue of Nvidia, Oracle, CoreWeave, AMD, and Broadcom depends on a few AI labs, and show your assumptions and ranges explicitly. Include the Meta equity raise with Nvidia and AMD participation.

From Circular Financing in AI: Who Pays Whom Between Nvidia, OpenAI and Oracle

Jon Sinclair using Luminix AI
Jon Sinclair using Luminix AI Strategic Research

OpenAI and Anthropic alone account for well over $2T in headline multi-year compute and infrastructure commitments (primarily 2025–2035 or to 2036), far exceeding the query’s “roughly $1T” framing, which likely reflects an earlier or narrower subset of signed deals.[1][2][3] OpenAI’s commitments center on a mix of hyperscaler cloud capacity, custom silicon development/leases (Broadcom), GPU purchases, and specialized cloud (Oracle, CoreWeave). Anthropic’s are detailed in its amended S-1 (~$895–901B in purchase commitments and leases as of late September/early November 2026 disclosures). Meta adds another ~$100B+ across multiple providers, with xAI/SpaceXAI appearing in both as a provider and buyer. These figures are headline/contract values (minima, “up to” caps, or reported totals); actual cash outlays depend on delivery schedules, acceptance, and lab revenue growth to fund them.[4][5]

Breakdown by counterparty (approximate aggregates across major labs, with ranges reflecting “up to”/minima and partial overlaps):
- Microsoft (Azure): ~$295B+ (OpenAI incremental $250B Azure services commitment from Oct 2025 restructuring/expansion; Anthropic at least $44.9–49.2B minimum through ~2033).[1][2]
- Oracle: ~$300B (primarily OpenAI multi-year capacity, e.g., Stargate-related or Abilene-scale deals starting ~2027; Oracle’s own RPO reached $664B by Aug 31, 2026, with OpenAI cited as a significant but not sole portion).[1][6]
- AWS/Amazon: ~$250B+ (OpenAI expanded to ~$138B total including $100B addition; Anthropic $110B minimum through 2036).[7][2]
- Google Cloud: ~$121B+ minimum from Anthropic (through 2033, take-or-pay elements) plus additional TPU estimates of $125–150B (2028–2031, outside core table).[2][8]
- CoreWeave: ~$60B+ (OpenAI cumulative ~$22.4B across expansions; Meta ~$35B total including ~$21B expansion; Anthropic multi-year undisclosed but material).[9][10]
- Broadcom: ~$500B+ (OpenAI $350B custom silicon/10 GW deployment; Anthropic $161.2B TPU leases).[1][2]
- AMD: ~$110B+ (OpenAI $90B; Anthropic >$20B Helios/custom; Meta multi-GW partnership with warrants).[1][2]
- Nvidia: Direct chip commitments harder to isolate (~$100B cited for OpenAI in one model) but substantial via guarantees/backstops (~$105B+ for OpenAI Ohio campus leases) and neocloud capacity agreements; Nvidia is more a chip supplier with equity stakes and residual guarantees than a pure compute renter.[1][11]

Roughly 40–60%+ of the largest named commitments (especially Anthropic’s hyperscaler and Broadcom deals, totaling hundreds of billions) are structured as take-or-pay or non-cancellable (payable “regardless of actual usage,” with true-up payments if spend falls short; limited cancellation only for provider breach or specific undelivered tranches).[2][12] The balance is more conditional: subject to delivery/acceptance of capacity, counterparty financing milestones, or terminable on notice (e.g., Anthropic’s SpaceXAI up to $120.5B cancellable on 90 days’ notice after early 2027; portions of Lambda/Nscale deals billed only on delivery; some OpenAI/Oracle ramps tied to construction). Hyperscaler 10-Q lease disclosures (Microsoft $364.6B leases not yet started; Oracle $288B data center leases; Amazon $164.9B; Meta ~$356B) and credit agency analysis of off-balance-sheet obligations provide supporting context for the buildout scale, though labs’ own filings (Anthropic S-1) give the most granular counterparty and cancellability detail.[5][4] Many deals include equity investments or financing facilities from providers back to labs (circular element), which can make commitments more binding or provide funding support.

Revenue dependence on a handful of AI labs is high for specialized players and material (but diversified) for larger ones; single-lab concentration is a noted risk in filings and analysis. Oracle’s $664B RPO (as of Aug 31, 2026) has OpenAI as a substantial portion (estimates around half in some analyses); CoreWeave’s ~$132.5B backlog has OpenAI (~$22B contracts) and Meta as key anchors (OpenAI/Meta together a large share of its named deals).[13][7] Broadcom and AMD have significant exposure via custom/GPU deals with OpenAI/Anthropic/Meta (one analysis puts Nvidia at ~39% of frontier lab contracted compute share, Broadcom ~24%, AMD ~15%, with labs driving much of the growth).[14] Nvidia’s data center revenue (dominant portion of its totals) benefits broadly from AI buildout but is less concentrated on any single lab; it also holds guarantees and backstops tied to labs/neoclouds. Assumptions here include that “dependence” reflects contract/backlog attribution rather than recognized revenue to date (many deals ramp later); ranges are wide because not all providers break out lab-specific revenue, and actual realization depends on lab funding and usage.[15]

Meta’s ~$97.5B equity raise (upsized Oct 2026 offering, priced at $625/share with strong demand) included direct participation from Nvidia ($3B) and AMD ($1B), part of broader supplier interest up to $5B aggregate (not all binding); this deepens chip partnerships (e.g., expanded Nvidia Vera Rubin/Muse collaboration, AMD Helios testing and up to 6 GW deals) while funding Meta’s own multi-GW AI infrastructure.[16][16][17] This fits the pattern of circular financing where chip providers invest in (or guarantee for) customers who then commit to their hardware/cloud.

Key assumptions and ranges: Figures aggregate public announcements, S-1/10-Q disclosures, and analyst models (e.g., Tunguz for OpenAI vendor split); many are “up to” or estimated totals not fully reconciled in filings, so actual minima are lower and maxima higher. Delivery is phased (often 2027+ ramps), with ~$200B+ of Anthropic commitments noted as subject to termination/conditions in earlier disclosures (updated totals reflect evolution). Revenue dependence estimates are directional (high single-lab or few-lab concentration for Oracle/CoreWeave/Broadcom elements) and could shift with diversification or cancellations. Additional research on latest 10-Qs or specific credit analyses would refine conditional portions. Overall, the ecosystem shows strong commitments but material execution/funding risk concentrated among a few labs and providers.


Recent Findings Supplement (November 2026)

Meta's October 2026 equity raise includes direct participation from chip suppliers Nvidia and AMD, providing fresh capital for its AI buildout while deepening hardware partnerships.[1][2]

Meta priced $25 billion in Class A common stock (at $625/share, upsized from $22.5B) and $12.5 billion in mandatory convertible preferred depositary shares on October 27, 2026, with a planned $60 billion at-the-market (ATM) program for non-voting Class C shares starting in Q1 2027 (total potential ~$97.5 billion). Nvidia committed $3 billion and AMD $1 billion (Broadcom had indicated non-binding interest as part of a combined $5 billion ceiling, described as non-binding). Proceeds support general corporate purposes, including AI infrastructure amid 2026 capex guidance of $135–145 billion. The deals coincide with expanded deployments: Meta intends to broaden Nvidia Vera Rubin GPUs/CPUs, Spectrum-X networking, and BlueField-4; AMD Helios racks are installed and under testing for broader rollout.[3][4]

This structure mixes immediate and deferred equity (forward sales settle by Nov 19, 2027) with supplier equity stakes that align incentives on hardware roadmaps. For competitors, it highlights how AI labs can tap public markets and supplier balance sheets rather than pure debt or internal cash, potentially pressuring pure-play neoclouds or smaller labs without similar access.

Anthropic's amended S-1 filings (October–November 2026) provide the most granular recent disclosure of lab-side commitments, totaling ~$900.9 billion in purchase commitments and leases (substantially all compute/infrastructure) as of late September/early November updates, against an ~$81 billion revenue run-rate.[5][5]

Key counterparty breakdown (minimums or stated figures; some "up to"):

  • Broadcom (TPU equipment leases): ~$161.2 billion (5-year terms, non-cancellable except lessor default; rent starts on "ready for service").
  • Google Cloud: At least ~$121.3 billion minimum (take-or-pay/"regardless of actual usage," true-up for shortfalls; Apr 2026–Jul 2033).
  • SpaceXAI (xAI-related): Up to $120.5 billion (delivery-based; cancellable by either party on 90 days' notice after initial periods ending ~Jan/Mar 2027).
  • Amazon (AWS): At least $110 billion minimum (take-or-pay, true-up; May 2026–Apr 2036).
  • Microsoft (Azure): $47.2–49.2 billion (mostly non-cancellable except uncured breach or undelivered tranches; Nov 2026–mid-2033).
  • Nscale: Up to ~$44.6 billion (fees accrue on delivery/acceptance; into 2033; some financing conditions).
  • Lambda: ~$41.3 billion (~$34.9 billion terminable; into 2034; delivery-based).

~ $186 billion of the total is subject to termination rights or conditions (e.g., counterparty financing or delivery milestones), down slightly from prior filings even as the headline rose. Additional estimated $125–150 billion for extra TPU capacity (2028–2031 deliveries, not in the main total) and other deals (e.g., AMD hardware >$20 billion, Fluidstack) sit outside the core figure. Leases not yet commenced and post-Sept commitments contribute heavily; ~$451.6 billion expected by end-2030.[6][5]

This data (from Anthropic's IPO registration amendments) is firmer on the buyer side than most seller disclosures and shows a mix of locked-in hyperscaler take-or-pay deals versus more conditional neocloud/GPU capacity. Implications for entrants: Labs face extreme leverage to utilization/revenue growth; suppliers with firm take-or-pay exposure (e.g., Google, Amazon, Broadcom) have stronger downside protection than those reliant on delivery milestones.

OpenAI's reported multi-year Oracle commitment (~$300 billion over ~5 years, ~4.5 GW, slated to start 2027) remains a focal point of counterparty concentration risk, referenced in recent Oracle filings and analyst commentary.[7][8]

Oracle's remaining performance obligations (RPO) reached $664 billion as of Aug 31, 2026 (up sharply YoY; ~13% expected in next 12 months, ~37% in months 13–36, etc.). Analysts (including S&P references) estimate ~$300 billion (~45%) tied to OpenAI. Oracle also disclosed ~$248 billion in future lease commitments (substantially data centers/cloud capacity, 15–19 year terms, commencing through FY2028/2029). The Oracle-OpenAI deal is part of broader Stargate-related efforts but carries execution risk given OpenAI's cash burn and funding needs.[9][9]

Oracle's RPO growth outpaces near-term revenue conversion, highlighting lease-duration risk on the supplier side. For market participants, this underscores how a few labs drive outsized backlogs at infrastructure providers, with limited public detail on cancellation terms.

CoreWeave and AMD show notable revenue/backlog dependence on a handful of AI labs or hyperscalers, per recent 10-Q/10-K disclosures, while Nvidia's direct AI chip share has declined amid diversification.[10][11]

  • CoreWeave: Microsoft accounted for ~67% of FY2025 revenue (diversifying in 2026 quarters, with top customers at 30–45% ranges in recent periods but still concentrated). OpenAI contracts total up to ~$22.4 billion (multiple 2025–2026 deals). RPO/backlog ~$132.3 billion as of Sep 30, 2026 (~98% of revenue from committed contracts). Debt ~$35–50 billion.[12][13]
  • AMD: Issued warrants tied to OpenAI/Meta chip deals (up to ~10% stakes possible on milestones), triggering a $1.07 billion non-cash Q3 2026 charge as shipments ramp. Meta equity participation ($1B) and OpenAI commitments add exposure.[14]
  • Nvidia: Share of contracted AI compute infrastructure for frontier labs fell to ~39% (from 83% historical chip revenue share), with Broadcom (~24%), Google TPU (~16%), AMD (~15%), and Trainium (~6%) rising. Data center revenue remains dominant but customer concentration disclosures emphasize broader ecosystem.[11][15]
  • Broadcom: Prominent via large Anthropic TPU lease exposure.

Overall ranges and assumptions (post-May 2026 data only): Aggregated trackers (e.g., aicomputedeals.com as of mid-Nov 2026) show OpenAI deals totaling hundreds of billions to ~$1.6T headline across ~21 deals and Anthropic ~$549B across 15, with combined lab commitments often cited in the $1T+ range when including Meta/xAI/etc. (many overlapping or estimated). Firmness varies: hyperscaler deals (Google, Amazon, Microsoft, Broadcom) skew take-or-pay/non-cancellable (~$440B+ in Anthropic alone); neocloud/GPU deals more conditional on delivery/financing (~$186B+ terminable in Anthropic case). Oracle RPO concentration and CoreWeave debt amplify supplier-side risks. Meta's raise adds a circular equity element. These figures derive from S-1s, 10-Q/10-K RPO/lease notes, and press on specific contracts; actual recognized revenue depends on utilization, delivery, and funding. Ranges reflect "up to"/minimum distinctions and analyst estimates (e.g., Oracle-OpenAI split). Additional research on full 10-K details or credit agency reports would refine conditional portions.[16][17]

New data since May 2026 centers on Anthropic's detailed S-1 disclosures, Meta's supplier-backed raise, updated RPO/backlog figures, and concentration analyses—shifting focus from headline announcements to execution, firmness, and balance-sheet implications.

Get Custom Research Like This

Start Your Research