What is off-balance-sheet financing in AI data centers?

It is any structure where a company is exposed to a data center project but does not report the project's assets and debt as its own. Joint ventures, special purpose vehicles, sale-leasebacks and securitizations all do this. Accounting rules decide who consolidates a project based mainly on who controls it, not on who is exposed to it, so the exposure is usually disclosed in the footnotes instead of the headline debt figure.

How much of Meta's Hyperion data center sits on Meta's balance sheet?

Meta holds a 20 percent equity-method interest and does not consolidate the venture, because it says it lacks the power to direct the activities that most affect the venture's performance. Blue Owl-managed funds hold 80 percent. Meta's FY2025 Form 10-K states its maximum exposure to loss related to the venture was 45.95 billion dollars as of December 31, 2025, alongside a 12.31 billion dollar initial leaseback commitment and residual value guarantees with an aggregate threshold of about 28 billion dollars.

What is Nvidia's $105 billion SB Energy guarantee, and how is it different from Meta's Hyperion guarantee?

It is a disclosed, capped, default-triggered guarantee, not equity and not a residual-value guarantee. Nvidia's Form 10-Q for the quarter ended July 26, 2026 states that in August 2026 it entered guarantees capped at a total of 105 billion dollars to provide credit support on land, power and shell buildout for an OpenAI-affiliate tenant's leases at SB Energy's PORTS Technology Campus in Ohio, covering about 4.25 gigawatts of IT load with an option for roughly 3.8 additional gigawatts. Nvidia's payment obligations trigger on tenant default, not on the leased assets simply being worth less than expected, and the guarantee is limited to defined lease and power payments rather than the full cost of the site. That is the structural difference from Meta's Hyperion residual value guarantees above: a residual value guarantee pays out if the asset's market value falls short at lease end, while Nvidia's SB Energy guarantee pays out only if the tenant actually defaults. This is a separate, newer and larger disclosure than the Nvidia-OpenAI 100 billion dollar letter of intent discussed elsewhere on this site: no SEC filing, this one included, has ever recorded that letter of intent or its reported conversion to a 30 billion dollar equity stake.

Are electricity customers already paying for AI data centers?

In the PJM region, yes, and the figure is documented. PJM's Independent Market Monitor calculates that data center load increased costs to other PJM customers by 23,100,955,341 dollars across the last three annual capacity auctions. That is a measurement of costs already allocated, not a forecast. Elsewhere the picture is mixed: several states have adopted large-load tariffs with minimum-take provisions specifically designed to keep those costs off other customers.

Is AI financing the same as telecom vendor financing in 2000?

Partly. The form rhymes: a supplier extends capital that enables purchases of its own product. Lucent disclosed about 6.7 billion dollars of customer credit commitments at September 30, 2000, roughly 20 percent of that year's revenue, and Nortel labelled about 4.1 billion dollars of customer financing an off balance sheet instrument in its own footnotes. Two differences hold up: telecom's borrowers were largely sub-investment-grade carriers, and telecom's fiber sat unused for years whereas GPU capacity is largely contracted before it is energized. One common claim does not hold up, that hyperscalers simply fund their own building. Meta financed a 27 billion dollar campus with 80 percent outside ownership in the same year it spent 69.69 billion dollars on property and equipment. What has genuinely improved since 2000 is disclosure, but unevenly: Nvidia's filings name 10 billion dollars for Anthropic and 5 billion for Intel while giving no figure for its OpenAI letter of intent.

Who holds the risk if AI demand does not arrive?

It depends on the structure, and that is the point. In project-financed joint ventures the private credit funds and bondholders hold the first loss, with the sponsor exposed through leases and residual value guarantees up to a disclosed ceiling. In consolidated asset-backed lending, such as CoreWeave's, the borrower holds it directly. Where generation built for a large load enters a utility's rate base and that load does not arrive, the cost would fall to other ratepayers unless a minimum-take tariff prevents it. This site is educational research, not financial advice.

The Mechanism

Shadow Debt and the Public Balance Sheet

The financing structures that keep data center debt off the sponsor's books, and the power costs that land on the public, are the same story told from two ends. Where does the real cost sit, and who holds the risk if demand does not arrive.

Updated August 31, 2026

The AI Money Loop traced the private loops: chipmakers investing in their own customers, credits recycled into revenue. This page asks a narrower question: once a data center gets built, where does the real cost sit, and who holds it if demand never arrives. The loops connect to the physical grid on Outside Shocks; this page sits between them.

Figures below are current to late July 2026, the site's broader snapshot date. disclosed means a filing or official document, reported means journalism, estimate means analyst modeling or arithmetic from disclosed figures. Where two sourced numbers disagree, the figure is marked contested.

The clearest case: one campus, one filing

Meta's "Hyperion" venture with Blue Owl Capital is the cleanest structure to study: the whole chain sits in one audited filing. Meta contributed $4.30 billion of held-for-sale assets, net of liabilities [disclosed]] ([Meta 10-K (opens in a new window)). In return it received a one-time distribution its 10-K puts at $2.55 billion [disclosed]] (same 10-K), though Meta's October 2025 release called the same payment "approximately $3 billion" disclosed ([Meta/Blue Owl release (opens in a new window)). contested Use the audited, filed figure.

Meta owns 20% of the venture; Blue Owl-managed funds own the other 80% [disclosed]] (same release). Total development cost is about **$27 billion**, with Blue Owl's funds contributing about **$7 billion** in cash at signing disclosed (same release). Meta uses the equity method and does not consolidate: it lacks power to direct the venture's most significant activities, so the [variable interest entity stays off its books disclosed (same 10-K).

What Meta will pay: an initial sale-leaseback commitment of $12.31 billion, starting 2029. What it has guaranteed: residual value guarantees with an aggregate threshold of about $28 billion, decreasing over time, across the first 16 years [disclosed]] (same 10-K; [Meta IR release (opens in a new window)). The number tying it together: Meta's maximum exposure to loss on the venture was $45.95 billion as of December 31, 2025 disclosed (same 10-K). Everything else here is measured against how much less visible it is.

The largest disclosed guarantee: Nvidia's $105 billion at SB Energy

Meta's Hyperion guarantee above is the clearest structure to study, not the largest exposure found in this research. In August 2026 Nvidia entered guarantees capped at a total of $105 billion, providing credit support on land, power, and shell-buildout lease payments for roughly 4.25 gigawatts of IT load at SB Energy's PORTS Technology Campus in Ohio, on behalf of a tenant that is an OpenAI affiliate [disclosed]] ([Nvidia, Form 10-Q, Q2 FY2027 (opens in a new window)). Nvidia also holds an option, exercisable at its own discretion, to extend credit support to roughly 3.8 additional gigawatts as the site scales disclosed (same filing).

This is not the same instrument as Meta's guarantee, and the difference is worth stating precisely rather than glossing over. Meta's Hyperion residual value guarantee has an aggregate threshold of about $28 billion that decreases as the 16-year lease term runs, and it pays out based on what the underlying assets are worth when the venture returns them. Nvidia's SB Energy guarantee is a credit-support guarantee, not a residual-value guarantee and not equity: its payment obligations trigger if the tenant defaults on its lease and power payments, not on any asset's declining resale value, and the guaranteed amount increases as construction phases complete rather than decreasing over time. At nearly four times Meta's $28 billion guarantee ceiling, and well ahead of Meta's own $45.95 billion maximum loss exposure on the whole Hyperion venture, this is now the largest single disclosed AI-financing exposure found in this research.

What the structure does, and does not, do

None of this is hidden. It is disclosed, audited, and legal, but "does not consolidate" carries real weight.

A variable interest entity, often built as a special purpose vehicle, is a structure whose economics depend on outside parties more than its own equity. Whoever can direct its most important activities consolidates it; everyone else reports only their stake and guarantees. Meta says it lacks that power, so it books a 20% equity stake rather than the whole project's assets and debt, a project finance judgement, not an accounting trick: Blue Owl's funds hold 80% and direct it.

Stated conditionally: if the venture cannot cover its debt and lease payments, and Meta's guarantees are called, this structure would move a slice of that risk back onto Meta, up to the $45.95 billion ceiling, while keeping the debt off Meta's books until then. A sale-leaseback works the same way in miniature: hand over assets, get cash, pay rent on capacity you no longer own. This is off-balance-sheet financing done in the open, filed with regulators. The echo below is opacity, not illegality.

The rest of the market

Hyperion is unusually visible. Most of the market is not.

CoreWeave closed an $8.5 billion delayed draw term loan (DDTL 4.0) in a named, non-recourse subsidiary, CCAC VIII, in March 2026, but its special-purpose subsidiaries are consolidated into its own GAAP statements, so this is secured debt on CoreWeave's own balance sheet, unlike Hyperion's structure [disclosed]] ([CoreWeave 8-K (opens in a new window)). Total principal debt has kept climbing since: $25.149 billion at March 31, 2026, rising to $35.551 billion at June 30, 2026, a $10.4 billion increase in one quarter [disclosed]] ([CoreWeave, Form 10-Q, Q2 2026 (opens in a new window)). CoreWeave no longer names a single 67%-concentration customer the way its FY2025 10-K did; for the first half of 2026 its two largest anonymized customers were 40% and 23% of revenue [disclosed]]. Its SEC-filed Q2 2026 earnings release confirms **1.5 gigawatts of active power and 3.7 gigawatts contracted**; a separately reported "more than 5 gigawatts by 2030" target and a competing "8 gigawatts by 2030" figure both trace only to earnings-call commentary, not to any 10-Q or SEC-filed press release, and neither should be treated as filed disclosedreported. CoreWeave also closed a newer, structurally distinct facility in August 2026: a **$2.6 billion** delayed-draw term loan borrowed not by CoreWeave itself but by a named special-purpose subsidiary, CoreWeave Financing DDTL V-V, LLC, secured by that subsidiary's own assets and maturing September 2031, see [circular financing for the full structure.

Three disclosure gaps sit upstream, and they are the most useful thing on this page.

Oracle's remaining performance obligations grew $85 billion in one quarter, from $553 billion to $638 billion, ended May 2026 [disclosed]] ([Oracle (opens in a new window)). Oracle does not disclose customer-level concentration in that backlog. A widely repeated claim that one customer is about half of it traces to a rating-agency action reported at second hand, not to an Oracle filing, so it is left out here. Nvidia has the mirror gap: one unnamed customer was 23% of revenue, a second 16%, for the quarter ended July 27, 2025 [disclosed]] ([Nvidia 10-Q (opens in a new window)).

The third gap is the sharpest. In September 2025 Nvidia and OpenAI announced that Nvidia "intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed," under a letter of intent [disclosed]] ([Nvidia (opens in a new window)). That figure has never appeared in an Nvidia filing. Nvidia's Q3 FY2026 10-Q names $10 billion for Anthropic and $5 billion for Intel in the same passage, and for OpenAI says only that it "entered into a letter of intent" with no amount [disclosed]] ([Nvidia 10-Q (opens in a new window)). Its FY2026 10-K, filed February 2026, still says: "We are finalizing an investment and partnership agreement with OpenAI. There is no assurance that we will enter into an investment and partnership agreement with OpenAI or that a transaction will be completed" [disclosed]] ([Nvidia 10-K (opens in a new window)). In February 2026 Nvidia's chief executive said publicly that it "was never a commitment" [reported]]. Nvidia's newest filing, the Form 10-Q for the quarter ended 26 Jul 2026 (filed 26 Aug 2026), still records no such figure disclosed ([Nvidia, Form 10-Q, Q2 FY2027 (opens in a new window)); its only new OpenAI-linked disclosure is the $105 billion SB Energy guarantee above, a different and much larger exposure that did not exist when the $100 billion figure was first floated disclosed. Smaller figures circulate in reporting and are not used here. contested

So the largest number in the AI financing story is also the least precisely disclosed one, and it is less precisely disclosed than the same company's much smaller investments. That is worth more than any estimate of its size.

Underneath the named deals is a market. Developers routinely fund completed, leased campuses through securitization: Aligned Data Centers issued $1.35 billion of securitized notes through a wholly owned subsidiary, part of $1.725 billion of structured debt [disclosed]] ([Aligned (opens in a new window)), and Vantage Data Centers placed a EUR 640 million asset-backed deal against four leased German sites [disclosed]] ([Vantage (opens in a new window)). KBRA counts $48.69 billion of US data center ABS and CMBS across 88 transactions from 2018 through May 2025 [estimate]] ([KBRA (opens in a new window)). Smaller operators borrow directly from private credit: Nscale signed a $1.4 billion delayed draw term loan from funds managed by PIMCO, Blue Owl and LuminArx [disclosed]] ([Nscale (opens in a new window)).

How much more is coming is a forecast, not a fact. Morgan Stanley Research estimates a $1.5 trillion global data center financing gap through 2028, of which roughly $800 billion would come from private and asset-based credit [estimate]] ([Morgan Stanley (opens in a new window)). Treat it as a projection of what lenders expect to fund, not a measurement of what has been funded.

One instrument cuts both ways. AMD's warrant to OpenAI for up to 160,000,000 shares at $0.01, vesting against 1 GW to 6 GW deployment milestones, still had zero shares vested and zero financial impact as of 27 June 2026, two quarters past AMD's fiscal 2025 year end [disclosed]] ([AMD, Form 10-Q, Q2 FY2026 (opens in a new window)). AMD issued an identical, separate warrant to Meta in February 2026, also 160 million shares at $0.01, tied to a 1-6 GW AMD GPU deployment agreement with Meta, which this page has not previously covered [disclosed]]. Read one way, that is leverage waiting to transfer value from shareholders. Read the other, proof nothing has moved: zero shares vested, zero accounting impact. Both readings are argued in full on [the Case Against page.

The other balance sheet: what the public carries

The structures above sit on private balance sheets, but the cost is not always private.

The PJM Independent Market Monitor calculates that data center load added $23.1 billion ($23,100,955,341 exactly) to costs paid by other PJM customers across the last three annual capacity auctions [disclosed]] ([PJM IMM (opens in a new window)). This is a fact about costs already allocated, not a projection: in the 2027/2028 auction alone, data center load drove a $6.5 billion increase, about 40% of a $16.4 billion total rise disclosed (same report). It is the strongest evidence here that someone besides the owner is already paying.

States also give up revenue to attract the load. Virginia's data center tax exemption produced a reported aggregate benefit of about $1.9 billion in FY2025, up from about $1.3 billion in FY2024, against $54.8 billion of qualifying investment [disclosed]] ([Virginia Dept. of Taxation (opens in a new window)). The same report projects a revenue-to-cost ratio of 1.7, the state's own analysis saying it comes out ahead over time disclosed (same report). contested Present both figures. Virginia's 2026 budget separately enacted a new, temporary data center electricity consumption tax of $0.011/kWh, 1 July 2026 through 30 June 2028, capped at $600 million a year, a real complication to a "no compensating charge" framing, though this pass did not independently confirm the rate and cap against the enrolled statutory text, relying instead on multiple consistent law-firm client alerts reported.

Georgia's own regulator gives two different numbers for the same 19 December 2025 vote. The Commission's same-day media advisory says it approved 9,885 MW of new generation, without stating a data-center share [disclosed]] ([Georgia PSC media advisory (opens in a new window)). Its March 2026 fact sheet says 9,985 MW, about 80% of it expected to serve data centers [disclosed]] ([Georgia PSC fact sheet (opens in a new window)). Both documents are the PSC's own releases about the same vote; the underlying signed stipulated order was located but not independently read this pass, so the 100 MW discrepancy is reported here rather than resolved. Both documents agree Georgia Power backstops the costs through 2031 if load does not materialize, with base rates frozen through 2028. Stated conditionally: if that generation joins the utility's rate base and load falls short, this structure would shift the gap onto other ratepayers once the backstop and freeze expire.

One thing we looked for and could not find: a municipal bond official statement whose stated use of proceeds names a specific data center project. Two research passes, including searches of the MSRB's EMMA disclosure system and of individual public power issuers, turned up none. There is a structural reason to expect that, rather than a gap in the search. Tax-exempt municipal bonds carry private business use limits, so a privately owned and operated data center generally cannot be named as the purpose of a tax-exempt issue [reported]] ([Whiteford (opens in a new window)). Issuers instead fund publicly owned transmission, water and road capacity in general terms. The practical consequence is that this channel would be difficult to see even if it were large, so this page makes no claim about its size in either direction.

The case against alarm

Regulators in several states have already moved to keep a stranded data center's cost off other ratepayers, and the record is more protective than it might seem.

AEP Ohio's data center tariff, approved by the Ohio commission in case 24-508-EL-ATA on July 9, 2025, is stricter than the headline suggests. Minimum billing demand runs to 85% of contracted capacity, over a ramp of up to 4 years plus a further 8 years. A customer without an A-minus or A3 rating must post "a guarantee or collateral at the time of signing the contract equal to 50% of the total minimum charges for the full term," and leaving after year five costs the minimum charges for a further 36 months [disclosed]] ([AEP Ohio tariff (opens in a new window)). Amazon, Google, Microsoft and Meta opposed the settlement, which is a reasonable sign the terms have teeth. Virginia went the same way, and is now approved by final order: SCC Case No. PUR-2025-00058, issued 25 November 2025, creates a new GS-5 rate class effective 1 January 2027 requiring a minimum 85% of contracted transmission and distribution demand [disclosed]] ([Virginia SCC news release (opens in a new window)). The order's own text describes a further 60% figure as a second minimum-take threshold on generation demand; the SCC's later February 2026 fact sheet describes what appears to be the same 60% as a collateral requirement, a guarantee covering up to 60% of a customer's minimum charges, rather than a demand threshold [disclosed]]. These may describe the same provision two ways, or two different provisions; the signed order text was not independently parsed this pass, so both framings stand. Georgia's PLL-18 tariff sets minimum monthly billing at whichever is greatest of the contract minimum, **50%** of contract capacity, or 500 kW, over 5 years disclosed ([Georgia Power tariff (opens in a new window)), and a 2025 Georgia rule sends any 100 MW-plus contract to the PSC 30 days before signing, revenue required to reduce, not raise, other bills [disclosed]] ([Georgia PSC rule (opens in a new window)).

Wisconsin went further: its PSC announced its decision in an April 2026 press release that explicitly disclaimed itself as preliminary; its Final Decision was in fact signed and served 21 May 2026, per a July 2026 court filing that quotes the signed date directly, though the Final Decision document itself was not located on the PSC's own filing system this pass [reported]]. The substance holds either way: the large-load threshold fell from 500 MW to **100 MW** and the minimum contract term extended to **15 years**, to protect other customers. Commissioner Kristy Nieto: "Existing Wisconsin customers should not pay a single cent to subsidize the service of data centers." The Western Virginia Water Authority's Google agreement, executed October 14, 2025, has Google paying for its own facility's water upgrades, ratepayers explicitly protected disclosed ([Western Virginia Water Authority (opens in a new window)).

Local governments are also collecting, not just paying. Loudoun County, Virginia collects over $100 million a year from data centers, 38% of general fund revenue, from roughly 4% of its commercial parcels [disclosed]] ([Loudoun County (opens in a new window)). ERCOT's large-load process requires new loads to post financial security within 180 days of study approval and energize within 365 days, customer money at risk first [disclosed]] ([ERCOT (opens in a new window)). The newest tariffs read like they were written with the last boom's mistakes in mind.

Does the telecom parallel hold?

The closest historical parallel is the telecom vendor financing wave of 1999 to 2001, covered in full on the telecom and fiber case study: a supplier extends credit that lets a customer buy its own product, then books the sale as revenue.

Lucent had commitments to extend credit of about $6.7 billion as of September 30, 2000, plus guarantee commitments of about $1.4 billion, of which $770 million was outstanding [disclosed]] ([Lucent 10-K405 (opens in a new window)); against FY2000 revenue of $33,813 million, that was about 20% of a year's revenue, our arithmetic from those two figures [estimate]]. Nortel called its customer financing an "off balance sheet instrument" in its own footnotes: about **$4,100 million** of commitments at end-2000, up from **$2,400 million** a year earlier disclosed ([Nortel 10-K405 (opens in a new window)), against FY2000 revenue of $27,948 million, roughly 14.6% [estimate]]. A year later Nortel's commitments had fallen to **$1,611 million**, only because revenue collapsed to **$17,511 million** after customer bankruptcies forced financing provisions disclosed ([Nortel FY2001 10-K (opens in a new window)). Lucent's episode later became an SEC matter: improper recognition of $1.148 billion of revenue, a $470 million pre-tax overstatement in FY2000, and a $25 million fine [disclosed]] ([SEC (opens in a new window)).

Does it hold today? Partly. It is worth being precise about which parts, because the easy answer is wrong in an interesting way.

The easy answer is that telecom's borrowers were shaky start-ups while today's buyers are hyperscalers spending their own cash, so the parallel fails. Two of those three claims survive. Telecom's borrowers really were mostly sub-investment-grade carriers, and Nortel's own FY2001 filing names customer bankruptcies as the cause of its provisions. Fiber really did sit dark for years, while GPU capacity is largely contracted before it is energized, which makes today's risk one of utilization and obsolescence rather than an empty asset, argued in full on The Demand Question.

The third claim does not survive. "Hyperscalers fund it themselves" is exactly what the Hyperion structure is designed not to do. A $27 billion campus was built with 80% outside ownership, an initial $12.31 billion lease obligation and up to $28 billion of guarantees, by a company that spent $69.69 billion on property and equipment that year and could have paid for it outright [disclosed]] ([Meta 10-K (opens in a new window)). The choice to finance it elsewhere is the story, not an exception to it.

What has genuinely improved is disclosure, and it has improved unevenly. Meta's arrangement is visible down to a single maximum-loss figure in an audited filing, which is more than anyone could say of Lucent's loan book at the time. But Nvidia's filings name $10 billion for Anthropic and $5 billion for Intel while giving no figure at all for the largest commitment in the sector. Disclosure quality is best where the deals are documented real estate and worst where they are the biggest.

The honest formulation: the 2026 structures are better collateralized than the 2000 ones, and the credit risk sits with lenders who chose it rather than with vendors who backed into it. What they share with 2000 is that the numbers that matter are contingent ones, disclosed in footnotes rather than in headline debt, and contingent numbers move fastest at exactly the moment everything else is moving too. A sector that has solved the 2000 problem would disclose its largest commitments at least as precisely as its smallest. This one does the opposite.

Hidden leverage resurfacing at the worst time is a recurring pattern (see Circular Financing, Explained). Enron kept about $27 billion, nearly half its assets, off its balance sheet by October 2000 [reported]] ([Senate PSI staff report (opens in a new window)). Citigroup's consolidation of its structured investment vehicles added $59 billion to assets at end-2007 [disclosed]] ([Citigroup 10-K (opens in a new window)), see the housing case study. The 1920s had an earlier version, trusts sold at large premiums over what they held, see the 1929 case study, economic history, not a primary document. The shared lesson: these structures are legal and routine, until demand falls and the contingent number stops being hypothetical.

What would tell us

A municipal bond statement naming a data center project surfaces.
Meta's next 10-K shows loss exposure well beyond $45.95 billion, or the guarantee schedule is extended, not stepped down.
Oracle, or any hyperscaler, starts disclosing customer-level backlog concentration.
AMD's OpenAI warrant shares start vesting against the 1 GW to 6 GW milestones.
The PJM Monitor's cost figure keeps climbing, rather than leveling off; tracked on The Warning Lights.
More states adopt Wisconsin- and Georgia-style minimum-take rules, the same falsifiable test used on Conclusions.

Updates and corrections

Every change to a stated fact on this page is logged here: what changed, what it said before, what it says now, and the source for the new figure. Typography and grammar fixes are not logged.

Date	What changed	Old	New	Source
2026-08-30	Nvidia-OpenAI $100B LOI, newest-filing check	contested flag stood, without confirmation from Nvidia's most recent filing	confirmed absent from Nvidia's 26 Aug 2026 10-Q as well; contested flag retained	Nvidia Corporation, Form 10-Q, quarter ended 26 Jul 2026 (opens in a new window)
2026-08-30	CoreWeave total debt and revenue concentration	$25.149 billion debt (31 Mar 2026); 67% concentrated in Microsoft	$35.551 billion debt (30 Jun 2026, +$10.4B in one quarter); two largest anonymized customers 40% and 23% (Microsoft no longer named)	CoreWeave, Inc., Form 10-Q, Q2 2026 (opens in a new window)
2026-08-30	Georgia generation approval, both PSC figures published	9,985 MW stated as the only figure	both 9,885 MW (Dec 2025 advisory) and 9,985 MW (Mar 2026 fact sheet) published side by side, discrepancy noted, neither picked	Georgia PSC media advisory, 19 Dec 2025 (opens in a new window)
2026-08-30	Virginia and Wisconsin ratepayer-protection status	Virginia's rate class described as "proposed"; Wisconsin's decision dated to an April 2026 preliminary release	Virginia's rate class is approved by final order (25 Nov 2025, effective 1 Jan 2027); Wisconsin's Final Decision was signed 21 May 2026, not just previewed in April	Virginia State Corporation Commission, news release, 25 Nov 2025 (opens in a new window)
2026-08-30	Vantage Data Centers ABS site count	against six leased German sites	against four leased German sites	Vantage Data Centers, press release, 9 Jun 2025 (opens in a new window)
2026-08-30	AMD warrant vesting status and Meta warrant addition	zero vested at AMD's FY2025 year end; no mention of a Meta warrant	still zero vested as of 27 Jun 2026; identical Meta warrant (Feb 2026) added	Advanced Micro Devices, Inc., Form 10-Q, Q2 FY2026 (opens in a new window)
2026-08-30	Virginia data-center taxation, new complementary tax added	exemption figures only, no mention of the 2026 electricity consumption tax	same exemption figures, plus the new $0.011/kWh electricity tax (2026 budget), source-caveated	Virginia Dept. of Taxation / VEDP, Biennial Data Center Report (RD40) (opens in a new window)