Autoresearch: Has Nvidia announced any named US or European take-or-pay GPU backstop beyond SharonAI and Firmus as of 2026-09-19?

Generated by /autoresearch on 2026-09-19. Synthesized across 3 rounds from 13 web pages, anchored by the Grokipedia Nvidia entry. See Provenance. Treat as raw material — review before promoting into a project or thread. Context: vault/projects/stock-market

Summary

As of 2026-09-19, this pass did not find a Nvidia-named United States or European GPU rental take-or-pay / credit-support partner beyond the two APAC names already on the July 1, 2026 program post: Sharon AI (Australia) and Firmus (Batam, Indonesia). The latest issuer filing — NVIDIA's Form 10-Q for the quarter ended July 26, 2026 — discloses $36 billion of typically six-year "AI cloud agreements" and $108.5 billion of maximum guarantee exposure ($3.5 billion land/power/shell for AI clouds plus a $105 billion SB Energy PORTS-Pike residual-value / LPS guarantee for OpenAI leases in Ohio). The 10-Q does not use the words "take-or-pay," does not name Sharon AI, Firmus, or GMI, and does not disclose an 18% revenue take-rate. The ~18% figure is an explicit SemiAnalysis modeled scenario (40% share of rent above a floor), not issuer disclosure. SemiAnalysis's $530 billion "gross OBS" total is their sum of several separately presented 10-Q tables, not a single Nvidia-labeled guarantee line.

Findings

Named GPU credit-support partners remain APAC: Sharon AI and Firmus

NVIDIA's July 1, 2026 business-model post says AI clouds will procure NVIDIA infrastructure under a "revenue-sharing and credit-support model," with NVIDIA earning product revenue plus "a share of the cloud revenue on the supported capacity." It names Sharon AI and Firmus as "among the first companies" on that model: Sharon AI deploying up to 40,000 Grace Blackwell GB300 GPUs, and Firmus building a Batam, Indonesia DSX campus expected to scale to 360 MW and up to 170,000 NVIDIA GPUs (NVIDIA blog, 2026-07-01). The post does not name a U.S. or European partner and does not publish a take-rate percentage.

SharonAI Holdings Inc. (NASDAQ: SHAZ) disclosed a six-year Master Cloud Services Agreement and Order Form 1 dated June 8, 2026 with a contract value of up to $4.88 billion, covering 72 MW in Australia and up to 40,000 GB300 GPUs (SharonAI Form 8-K, 2026-06-12; Sharon AI press release, 2026-06-12). The 8-K uses "revenue share" and "credit-support" language, notes that capacity reserved for NVIDIA may be sold to third parties, and does not label the $4.88 billion as a Nvidia debt guarantee. The $4.88 billion is SharonAI's disclosed ceiling, not a Nvidia 10-Q line item.

Firmus's own June 2026 newsroom post announces a partnership running through 2034, anchored by a 360 MW NVIDIA DSX campus in Batam, covering up to 170,000 accelerators through 2027–2028, under the same "revenue-sharing and credit-support" wording. Firmus says that, based on customer commitments, it expects US$25–30 billion from committed offtake agreements in the first six years (Firmus newsroom). That $25–30 billion is Firmus's expected customer offtake, not a stated Nvidia floor. A Straits Times reprint (Bloomberg-origin) restates the same offtake range and DayOne construction role (Straits Times, 2026-06-29).

Data Center Dynamics, reporting the July 1 post, likewise treats Firmus and Sharon AI as first adopters of a rent-back-at-a-fixed-rate backstop and notes earlier 2025 CoreWeave / Lambda capacity deals as a different, prior pattern (DCD, 2026-07-02). Those 2025 deals are not presented here as 2026 named AICP take-or-pay announcements.

NVIDIA's September 9, 2026 Australia newsroom release lists Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC and AirTrunk for a land/power/shell DSX buildout of up to 2 GW by 2027. It does not apply the July 1 "credit-support" / revenue-share wording to the newly listed names and does not announce a U.S. or European GPU take-or-pay (NVIDIA Newsroom, 2026-09-09).

No confirmed named US/EU GPU take-or-pay; GMI is not issuer-named as one

GMI Cloud's July 22, 2026 Mountain View release says it has a "strategic collaboration with NVIDIA," is "among the earliest" on "this new compute partnership model," committed $500 million of its own CapEx, and has "nine-figure contracts with a leading U.S. frontier AI enterprise." It does not state a Nvidia take-or-pay floor, a dollar backstop, or a named U.S./European campus under Nvidia credit support (GMI Cloud, 2026-07-22). SemiAnalysis later estimates a GMI AICP obligation of about $2.2 billion and lists GMI as a third announced AICP deal beside Firmus ($21.1 billion, SemiAnalysis) and SharonAI ($4.2 billion, SemiAnalysis) (SemiAnalysis, 2026-09-11). Those dollar splits are analyst allocations of the $36 billion 10-Q line, not Nvidia naming.

This pass did not retrieve an official NVIDIA page that names a European cloud (Nscale, Nebius, Fluidstack, Scaleway, or similar) as an AICP / credit-support / take-or-pay GPU partner. Official program language on July 1 still stops at Sharon AI and Firmus (NVIDIA blog, 2026-07-01).

The U.S. named Nvidia support is PORTS-Pike LPS / residual-value, not a GPU take-or-pay

On August 17, 2026 NVIDIA announced credit support on land, power and shell at SB Energy's PORTS-Pike campus in Pike County, Ohio, to exclusively host NVIDIA compute, with OpenAI as tenant on a 20-year lease. The newsroom summary is LPS credit support for an initial 4.25 IT-GW (option on remaining capacity), plus a $1.5 billion equity investment in SB Energy — not a named GPU rental floor for a neocloud (NVIDIA Newsroom, 2026-08-17).

The Q2 FY27 10-Q and the accompanying CFO commentary match that structure: guarantees capped at $105 billion, effective in phases as sites become ready for service (first expected fiscal 2029), triggered on certain tenant defaults, limited to defined portions of lease and power payments, declining as OpenAI pays, plus an option for about 3.8 additional GW (NVIDIA 10-Q, quarter ended 2026-07-26; CFO Commentary exhibit). That is a U.S. named guarantee, but it is not a SharonAI/Firmus-style GPU take-or-pay.

What the latest NVDA 10-Q actually discloses (issuer numbers, issuer words)

Latest periodic report used here: Form 10-Q for the quarterly period ended July 26, 2026 (nvda-20260726.htm), with matching tables in the August 26, 2026 CFO commentary (q2fy27cfocommentary.htm).

Ordinary future commitments as of July 26, 2026 (billions):

Supply and capacity rose from $119 billion last quarter to $279 billion, "primarily related to the procurement of memory" (CFO Commentary).

Additional commitments:

Issuer wording on AI cloud agreements (not "take-or-pay"): AI clouds procure NVIDIA data-center products and NVIDIA "commit[s] to cloud service agreements, which the AI clouds can unilaterally stop providing to us and sell to third-party customers at more advantageous rates." Commitments are "typically six years," totaled $36 billion as of July 26, 2026, and "decrease as capacity is used by third-party customers or by us" for R&D. "If certain criteria are met, we will participate in revenue share" from third-party customers — no rate disclosed (10-Q MD&A / Note 10). Risk-factor language adds that if partners do not sell committed capacity, NVIDIA has agreed to purchase it (10-Q).

Guarantees (maximum gross exposure):

A grep of the fetched 10-Q HTML found no "take-or-pay" / "take or pay" string and no 18% revenue-share figure. The only nearby "18%" in that file is sequential revenue growth, not a take-rate (10-Q).

The same 10-Q also records August 2026 MOUs with large capital providers "designed to mobilize more than $500 billion of third-party capital," with NVIDIA optionally providing "limited residual-value support for a portion of specific projects" — a prospective program, not a named US/EU GPU take-or-pay (10-Q).

SemiAnalysis $530B / $36B / 18%: secondary model, not issuer fact

SemiAnalysis (Sep 11, 2026) writes that the 2Q F1/27 10-Q disclosed "$530B of gross off-balance sheet guarantees across six different line items," up from $184 billion, driven by supply commitments ($119B → $279B), guarantees/LPS ($3.5B → $108.5B on PORTS-Pike), and two new lines: $36 billion AI-cloud agreements and $20 billion of datacenter leases NVIDIA expects to reassign (SemiAnalysis, 2026-09-11). Adding the 10-Q tables as presented ($366B + $56B + $108.5B) equals $530.5 billion. NVIDIA itself does not publish a single "$530 billion guarantees" caption; SemiAnalysis is summing separately labeled commitments and guarantees (including supply/memory purchase obligations that are not guarantees).

SemiAnalysis interprets the $36 billion line as "take-or-pay floors under Neocloud capacity" at about $2.35/hr/GPU versus a $4.50–$4.60 five-year market, and allocates Firmus $21.1 billion, SharonAI $4.2 billion, and GMI ~$2.2 billion (estimated). It also says NVIDIA "reportedly paused AICP" and that SemiAnalysis still models the line closing F1/27 at $51 billion (SemiAnalysis, 2026-09-11). Those allocations, the pause, and the $51 billion close are not in the 10-Q text fetched here. Note the SharonAI $4.2 billion SemiAnalysis figure versus SharonAI's own up-to-$4.88 billion 8-K ceiling (SharonAI 8-K).

The ~18% take-rate is from SemiAnalysis's July 6, 2026 backstop explainer, in an illustrative table: "In the case where Nvidia earns 40% revenue share" of rent above a year-1 floor of $3.68/hr against a $6.75/hr 1-year rental, "the Nvidia take rate works out to be about 18% on average" over six years (SemiAnalysis, 2026-07-06). SemiAnalysis says terms are negotiated per neocloud. That is a model output under an assumed 40% above-floor share, not a number NVIDIA filed.

Encyclopedic anchor (non-load-bearing for contested claims)

The Grokipedia Nvidia page is a company overview (founding, CUDA, FY2025 income, market-cap milestones). It does not discuss take-or-pay, AICP, Sharon AI, Firmus, or 10-Q OBS lines (Grokipedia: Nvidia). Contested financing claims above rest on SEC and company primaries, with SemiAnalysis used only as secondary.

Contradictions and open questions

• Issuer structure vs "take-or-pay" label. SemiAnalysis and much of the secondary tape call the $36 billion line a six-year take-or-pay floor. The 10-Q describes a NVIDIA commitment to buy cloud service that the partner can unilaterally replace with a better third-party sale, with the commitment declining as that happens. Whether that is economically equivalent to classic take-or-pay depends on unused-capacity behavior; the issuer did not use the phrase.
• SharonAI $4.88B vs SemiAnalysis $4.2B. Both cannot be treated as the same disclosed floor without an allocation note. $4.88B is SharonAI's contract-value ceiling; $4.2B is SemiAnalysis's assigned Nvidia obligation.
• Firmus $25–30B vs SemiAnalysis $21.1B. Firmus's number is expected customer offtake; SemiAnalysis's $21.1B is an estimated Nvidia obligation. Different objects.
• GMI. SemiAnalysis treats GMI as a third named AICP deal (~$2.2B). GMI's own release does not publish a backstop amount or geography of a Nvidia-floored cluster. Unresolved whether GMI is a U.S. named GPU backstop in any primary document this pass did not fetch (e.g. a later 8-K or financing memo).
• AICP pause. SemiAnalysis (Sep 11) says NVIDIA "reportedly paused" further AICP deals. This pass did not fetch a primary NVIDIA confirmation or denial (WSJ/Reuters/Bloomberg were not used; Reuters/Bloomberg are hard-blocked or instructed-skip).
• Undisclosed partners inside the $36B. If SemiAnalysis's three-deal stack is ~$27.5B, a remainder of the $36B is unnamed. That remainder could include U.S. or European counterparties that have not been publicly named. Absence of a name is not proof that no such contract exists — only that this pass found no official naming.
• $530B framing. Summing $366B + $56B + $108.5B ≈ $530.5B is arithmetic on issuer tables; calling the whole stack "guarantees" overstates the guarantee subset ($108.5B).
• Korea sovereign AI / other demand color was out of scope and not re-researched; it is not a named backstop in the sources fetched here.

Provenance

Rounds run: 3 (full)

Sub-questions by round:

Round 1 (broad survey):

1. Has NVIDIA announced any named US or European GPU take-or-pay / credit-support partner beyond Sharon AI and Firmus as of 2026-09-19?
2. What does the latest NVDA 10-Q / CFO commentary say about off-balance-sheet commitments, LPS, and guarantees?
3. What does SemiAnalysis's September 2026 note claim about $530B OBS, $36B AI-cloud, and the 18% take-rate?
4. What is the official wording on the Sharon AI and Firmus deals?

Round 2 (drill-down):

1. Exact 10-Q Note / MD&A language on "AI cloud agreements" — targeting whether the issuer says take-or-pay or something narrower.
2. Is GMI a named US GPU take-or-pay in a primary document, or only a SemiAnalysis estimate? — targeting the only candidate non-APAC AICP name from round 1.
3. Does SharonAI's own SEC filing state $4.88B and on what legal basis? — targeting the $4.88B vs $4.2B gap.

Round 3 (resolve remaining uncertainty):

1. Firmus primary numbers for Batam / $25–30B — targeting whether that figure is customer offtake or a Nvidia floor.
2. Any official NVIDIA naming of a US/EU AICP partner after July 1, 2026, versus the Ohio LPS announcement — targeting the US/EU named-backstop question.
3. Confirm the 18% figure's home in the July SemiAnalysis model — targeting issuer vs model.

Anchor source (Grokipedia, fetched before round 1):

• Nvidia — 4,017 chars extracted via _lib/grokipedia.py fetch "Nvidia" --max-chars 4000 (source: api, 316 citations on the live page) — company overview only; no AICP / OBS / take-or-pay content. Used as vocabulary primer, not as a source for contested financing claims.

URLs fetched (13 successful, 1 failed):

Round 1:

• NVIDIA Unlocks AI Compute at Scale — official — names Sharon AI and Firmus as first credit-support / revenue-share partners.
• NVIDIA Expands AI Infrastructure Capacity in Partnership With Australia's Data Center Ecosystem — official — Sep 9 Australia DSX ecosystem; no new US/EU take-or-pay names.
• CFO Commentary on Second Quarter Fiscal 2027 Results — official (8-K exhibit) — $366B / $56B / $108.5B tables.
• Nvidia's Backstop Universe – Heads I Win, Tails Who Loses? — specialist blog — $530B framing, $36B AICP, Firmus/SharonAI/GMI split, pause claim.
• Nvidia acts as backstop for customer GPUs — trade press — Firmus/Sharon first adopters; prior CoreWeave/Lambda color.
• [Failed: https://last10k.com/sec-filings/nvda/0001045810-26-000075.htm] — timeout; replaced in round 2 by the SEC HTML 10-Q.

Round 2:

• NVIDIA Form 10-Q for the quarter ended July 26, 2026 — official (sec.gov HTML; PDF-from-IR skipped as off *.gov whitelist) — controlling issuer language on AI cloud agreements, LPS, SB Energy, residual-value MOUs.
• GMI Cloud Commits $500 Million… — official — partnership-model language without a named take-or-pay amount.
• SharonAI Holdings Inc. Form 8-K — official — up-to-$4.88B MCSA; revenue-share / third-party-resale mechanics.
• Sharon AI Announces Six Year Strategic Compute Collaboration with NVIDIA — official — 72 MW / 40,000 GB300 Australia.

Round 3:

• Nvidia GPU Debt Backstop Unleashes the AI Project Trinity — specialist blog — 18% take-rate as a 40%-above-floor model scenario (2026-07-06).
• Firmus to Build 170,000 GPU AI Factory Campus with NVIDIA — official — Batam 360 MW; $25–30B expected committed offtake (customer, not Nvidia floor).
• NVIDIA Guarantees SB Energy's PORTS-Pike… — official — US LPS / OpenAI tenant; not a GPU take-or-pay.
• AI start-up Firmus, Nvidia to build data centre in Batam — news (Bloomberg syndication) — corroborates Firmus offtake range; Reuters.com / Bloomberg.com not fetched (hard-blocked / instructed skip).

Not fetched (routing): reuters.com and bloomberg.com skipped per SOURCE_RELIABILITY.md / instruction. NVIDIA IR PDF s201.q4cdn.com/.../NVDA-2027-Q2-10Q-Final-including-exhibits.pdf skipped (PDF off *.gov whitelist); equivalent text taken from SEC HTML.

Tools used: WebSearch, WebFetch, grokipedia-fetch via python3 .claude/skills/_lib/grokipedia.py fetch "Nvidia" --max-chars 4000. No --include-x. Priors skipped.

Generated: 2026-09-19 20:42 UTC