Methodology

What Counts as a Deal

A deal enters the ledger when all four hold:

1. At least one party is publicly traded, a hyperscaler subsidiary, or a private company at obvious scale (OpenAI, Anthropic, xAI class).
2. The subject is AI compute or AI data center capacity: GPU cloud contracts, IT-load leases for AI campuses, chip or server purchase commitments, AI data center build contracts, financing backstops, or equity investments explicitly tied to compute supply.
3. The headline value is disclosed or computable at $100M or more, or the capacity is 50 MW or more when no dollar value is disclosed.
4. Two independent sources exist, and at least one is primary (a filing, press release or shareholder letter) or tier-1 (Reuters, Bloomberg, CNBC, WSJ, FT). Aggregator sites never establish a row on their own.

Excluded: generic enterprise cloud contracts with no AI or capacity specifics, internal capex programs with no counterparty, and land purchases with no IT-load commitment.

Status Grades

• Binding: a signed, definitive agreement per the parties themselves.
• LOI / MOU: a letter of intent or memorandum, announced by the parties but not definitive. The Nvidia-OpenAI $100B partnership is the canonical example.
• Framework: announced umbrella agreements whose commercial terms are not yet contracted.
• Reported: credible tier-1 reporting of talks or unsigned terms, denied by nobody, confirmed by nobody.
• Terminated / Superseded: dead rows stay visible, they never vanish.

Only binding deals enter the headline totals. Everything else is displayed, labeled, and excluded.

Why the Totals Refuse to Double Count

• Expansion chains. When a second lease lands on the same campus with the same counterparty, it is a separate row linked to its parent. Each row is its own incremental contract, so summing rows is safe, and "campus total" restatements are never stored as deals.
• Snapshot rows. A company restating "$36B of total contracted backlog" or "more than $60B of orders received" is disclosing a snapshot that overlaps its own discrete deals. The same applies to program envelopes like Stargate's announced $500B, whose discrete contracts earn their own rows as they sign. Snapshot rows (basis: total backlog, order intake or program envelope) appear in the table, labeled, and are excluded from every summed figure.
• Renewal options. "Up to $50.2B with all renewal options" is upside, not a contract. Option value lives in its own field and its own column, never the headline.
• Circular deals. When the chip supplier finances its own customer, both the financing and the purchase can be real, and adding them still overstates net external demand. Circular rows carry a visible ↻ flag; the methodology position is flag, never net.
• MW before dollars. Contract dollars get restated, re-based and re-reported as prices move; megawatts and terms almost never do. Every dollar figure carries the date it was stated. Capacity leads the aggregates.

Verification

Every figure passes four checks before a row is published:

1. Scale: does the value fit the announcing company's size and prior deals?
2. Arithmetic: value, megawatts and term must multiply out to a sane price per MW-year; totals must equal their parts.
3. Consistency: the row must not contradict figures already published on this site or in the ledger without a logged revision explaining why.
4. Sourcing: two independent sources, one primary or tier-1, recorded on the row itself.

A validation suite enforces the mechanical half of these rules in CI; a row that fails cannot be published at all. When a published figure later proves wrong, the row is corrected and the change is logged in the changelog: corrections stay visible, they are never silent.

The Candidate Queue

Deals we believe exist but have not yet verified to the standard above. Listing them here is deliberate: it tells you what the ledger is missing, and it stops an unverified figure from borrowing the credibility of the verified ones.

• xAI Colossus expansion and any third-party compute or financing deals at ledger scale.

  Holding because: Not yet researched in-session; xAI discloses irregularly, mostly via posts rather than filings.

• Stargate site-level build contracts (Abilene/Crusoe, additional sites) that may deserve rows distinct from any top-level OpenAI-Oracle contract.

  Holding because: Not yet researched in-session; high double-counting risk against the Oracle contract, needs careful expansion_of/basis mapping.

• Iron Mountain signed 75 MW of new data center leases subsequent to Q2 2026 quarter-end (51 MW Mumbai, 25 MW London) with unnamed 'major global hyperscalers', on top of 110 MW YTD leasing through July.

  Holding because: No counterparty named and no per-lease dollar figure disclosed as of the 2026-08-05 Q2 earnings materials; below the ledger's specificity bar even though MW clears the 50 MW threshold. Needs the actual lease agreement or a named tenant.

• Hyperscale Data (NYSE American: GPUS) signed a 10-year MSA (plus two 5-year extension options) with an unnamed California-based neocloud provider for 20 MW of critical AI compute capacity at its Michigan data center campus, expandable to 52 MW; 'in excess of $1.2B' at Maximum Term, up to ~$3B with all options and expansions exercised. Announced 2026-06-24 via PRNewswire; restated in an 8-K filed 2026-08-04.

  Holding because: Meets the $100M/50MW inclusion bar on paper, but GPUS is a thinly-covered micro-cap (formerly Ault Capital-affiliated) with no tier-1 pickup found, an unnamed counterparty, and a headline value built almost entirely from unexercised renewal/expansion options (the disclosed base 20MW/no-options figure was not separately broken out). Needs the base-term-only dollar figure isolated from the Maximum Term figure, and ideally one more independent source, before it can be written as company_stated with max_with_options_usd used correctly.

• Nvidia and SK Group announced a partnership on 2026-07-24 that Nvidia values at 'more than $500 billion over multiple years', covering AI data centers and a long-term SK Hynix HBM4 memory supply agreement for Nvidia's Vera Rubin chips.

  Holding because: Explicitly described as letters of intent, not a completed supply contract, so status would be loi_mou (excluded from binding aggregates) at best. The $500B figure blends data-center investment and memory supply into one number; needs the two components split out (memory purchase commitment vs. data center investment are different deal_types) and each verified against a primary source (Nvidia/SK 8-K or press release) before any row, given the scale ($500B would be the single largest row in the ledger if binding).

• The National Nuclear Security Administration selected Amentum to negotiate a phased lease at the Savannah River Site (SC) for a 1 GW AI data center plus about 2 GW of dedicated on-site generation.

  Holding because: Not yet researched in-session. Government counterparty (NNSA) is an unusual fit for the ledger's inclusion criteria (publicly traded/hyperscaler/AI-lab-scale party), no dollar value found yet, and 'selected to negotiate' suggests this may not yet be a binding or even signed agreement (could be framework/reported status at best).

• Core Scientific's pre-existing CoreWeave data center lease(s), predating the new 2026-07-28 AMD deal. The AMD press release states the AMD deal 'doubled' Core Scientific's total leased AI capacity to ~1.1 GW, implying roughly 570 MW was already leased (most likely to CoreWeave, which has a well-known, long-running commercial relationship with Core Scientific dating to 2025).

  Holding because: Not yet researched in-session; needs the original CoreWeave-Core Scientific lease terms (value, MW, term, dates) pulled from primary sources, since it predates this ledger and could be a large missing row similar to the TeraWulf and Galaxy Digital gaps found this week.

• Sharon AI Holdings (SHAZ) has several other recently-announced contracts beyond the $373M deal added to the ledger this week: a 2026-04-01 five-year $1.25B AI infrastructure agreement with ESDS Software Solutions Ltd (8,200 Nvidia B300 GPUs, Australia, revenue from Q3 2026), a separate ~US$1.32B five-year cloud agreement with a 'global AI Lab', a ~$950M deal (May 2026) with a global technology company with major APAC presence, and $8.8B of total disclosed contracted revenue against 212 MW of capacity. Update 2026-08-14: the ~$4.9B gap between those four contracts ($3.89B combined) and the $8.8B total is explained by a six-year, $4.9B Nvidia strategic compute collaboration (up to 40,000 GB300 GPUs / 72MW) first announced 2026-06-12 and reiterated at Q2 2026 earnings; adding it closes the reconciliation almost exactly ($373M + $1.25B + $1.32B + $950M + $4.9B ≈ $8.79B ≈ $8.8B).

  Holding because: Same thin-coverage caveat as the $373M row already added, now including the $4.9B Nvidia collaboration: all sourcing found is SharonAI's own press releases/8-Ks plus trade press and wire aggregators (StockTitan, Benzinga, TechTimes, DataCenterDynamics, ConvergeDigest, Grafa, BusinessWire), no Reuters/Bloomberg/CNBC/WSJ/FT pickup for any of the five contracts. A Bleecker Street Research short-seller report (bleeckerstreetresearch.com/research/shaz, dated 2026-04-30, flagged 2026-08-14) specifically challenges the ESDS $1.25B contract's credibility: ESDS's own disclosed FY2025 revenue ($39.9M) and total assets ($69.5M) are far below the implied ~$250M/year payment obligation and the $140M letter-of-credit requirement, and flags a retracted 'NVIDIA is a strategic shareholder' claim (asserted in a 10-K, then walked back via 8-K two weeks later). Update 2026-08-21: an 8-K filed 2026-08-20 disclosed customer acceptance of Phase 1 under the $950M contract (escrow release, revenue still staged for Q3/Q4 2026) -- a milestone, not new corroboration, sourced only to SharonAI's own release. Two S-1/A amendments (2026-08-18, 2026-08-20) are routine secondary-resale-registration housekeeping restating the same figures, not new disclosures. More importantly, SharonAI's Q2 2026 results (reported 2026-08-06/08-13) showed actual revenue of just $1.93M against the $8.8B claimed total contract value, with the company itself pushing 'material revenue' out to Q4 2026 -- this is company-reported and doesn't independently corroborate anything, but it sharpens exactly the credibility gap Bleecker Street raised. No tier-1 outlet has picked up any of the five contracts. This argues for holding the whole cluster at needs_verification until either tier-1 pickup appears or the counterparty capacity questions are resolved.

• CoreWeave and quantitative trading firm Hudson River Trading announced a multi-year AI cloud deal on 2026-08-20, built on Nvidia Vera Rubin NVL72 compute for trading research and model development. Confirmed by both a CoreWeave press release and independent Bloomberg reporting. Notable as a new buyer category for the ledger (a quant trading firm, not an AI lab or hyperscaler).

  Holding because: No dollar value or MW/capacity figure disclosed by either party; CoreWeave's own CRO explicitly declined to give a number when asked, describing it only as 'multi-billion-dollar'. Fails the ledger's inclusion criteria (a deal needs a disclosed dollar value or a disclosed MW figure) as reported. Revisit if either company discloses a specific figure (e.g. at a future earnings call) or if a credible outlet publishes a sourced estimate.

License & How to Cite

The dataset is licensed CC BY 4.0. Use it, republish it, chart it, build products on it, including commercially. The only requirement is attribution with a link. Suggested forms:

AI Compute Deal Ledger, Regards of Wallstreet, retrieved 2026-08-23, https://www.regardsofwallstreet.com/data/ai-deals

Source: Regards of Wallstreet AI Compute Deal Ledger (CC BY 4.0)

Machine-readable exports: JSON · CSV. Corrections and missing deals: contact us.