Map the major off-balance-sheet and project-finance structures used for AI data centers…

The market has split AI data-center finance into two layers that do not match the accounting. Project debt and equity sit in bankruptcy-remote vehicles that own the shell, the power, or the GPUs. The hyperscaler (or a vendor standing behind a weaker tenant) keeps the economic risks that actually make the paper investment-grade: construction completion, rent, and residual value. Rating agencies largely price the tenant. GAAP often does not consolidate the debt, and often does not even book the full lease or the guarantee.

The template: Meta–Blue Owl Hyperion

Hyperion is the cleanest public example of how form is separated from substance. Meta and Blue Owl funds formed a joint venture to own the Richland Parish, Louisiana campus. Blue Owl holds 80 percent and the formal governance; Meta holds 20 percent, contributed land and construction-in-progress, and took a roughly $3 billion cash distribution at close. A bankruptcy-remote issuer, Beignet Investor LLC, sold $27.294 billion of 6.581 percent senior secured notes due May 30, 2049, priced at about Treasuries plus 225 basis points and rated A+ by S&P (single agency). PIMCO anchored roughly $18 billion; BlackRock took more than $3 billion. The notes are fully amortizing 144A project bonds that fund construction from day one, with proceeds parked and drawn as the campus is built. Rent is structured to start around June 1, 2029.
The financing channel works by splitting ownership from the risks that actually make the paper investable. A bankruptcy-remote vehicle owns the shell, the power infrastructure, or the GPUs and borrows against them. The hyperscaler—or a vendor standing behind a weaker tenant—keeps construction completion, rent, and residual value. Rating agencies largely price the tenant. US GAAP often consolidates neither the debt nor the full economic commitment.

Meta–Blue Owl is the template, and the accounting is the product

Hyperion in Richland Parish, Louisiana is the deal the rest of the market is copying because it shows how to keep a multi-decade, single-tenant campus off the operator’s balance sheet without giving up the credit that lenders require.

Funds managed by Blue Owl own 80 percent of the joint venture and the formal governance. Meta owns 20 percent, contributed land and construction-in-progress, and took a roughly $3 billion cash distribution at close. A bankruptcy-remote issuer, Beignet Investor LLC, sold $27.294 billion of 6.581 percent senior secured notes due May 30, 2049, priced at about Treasuries plus 225 basis points and rated A+ by S&P. PIMCO anchored roughly $18 billion; BlackRock took more than $3 billion. The notes are fully amortizing 144A project bonds that fund construction from day one. Rent is structured to begin around June 2029. [1] [2] [3] [4]

Three contractual features do the work:

• Lease slicing. Occupancy is written as four-year non-cancellable blocks with renewal options that can run total occupancy to about 20 years. The economic life of the campus is closer to 25–30 years. The short initial term is what Meta uses to argue the lease is an operating lease, not a finance lease, under ASC 842: renewal is asserted to be “not reasonably certain,” and the present value of four years of rent plus the residual-value guarantee is asserted to be less than “substantially all” of fair value. [1] [5]
• Residual value guarantee. For the first 16 years, if Meta does not renew, terminates early, or defaults, it owes a guaranteed minimum value sized so the related bonds can be repaid from sale proceeds or a Meta cheque. The threshold starts near $28 billion and declines. The last four years of the 2049 bond are covered by a termination fee equal to the outstanding balance, not the RVG. Meta’s parent also guarantees rent. [1] [6] [4]
• VIE control test. Under ASC 810, Meta argues it is not the primary beneficiary because Blue Owl holds the equity majority and the power to direct the activities that most significantly affect economic performance. Meta therefore does not consolidate the campus or the $27.3 billion of notes. It books a minority equity investment and, once leases commence, a short operating-lease liability. As of December 31, 2025, Meta disclosed maximum exposure to loss on the venture of about $45.95 billion—equity carrying value, lease commitments, future funding, and the RVG threshold—while recording no RVG liability because a payout was judged “not probable.” Ernst & Young flagged the VIE determination as a critical audit matter. [7] [8] [9]

S&P treats the paper as Meta credit in project-finance clothing: A+, one notch below Meta’s AA−, because of structural separation (no direct recourse to Meta assets, legal distance between the JV and the bond issuer). During construction, S&P’s view is that contractual protections shift cost-overrun and delay risk to Meta. In operation, a debt-service coverage ratio of about 1.12x plus the RVG supports the same A+. S&P has not added the RVG to Meta’s adjusted debt, citing headroom between the guarantee and a third-party appraisal. It has said it could change that view. [4] [10]

Moody’s is less willing to stop at the footnotes. In February 2026 it said short initial terms plus RVGs can leave reported lease liabilities well below plausible cash outflows, and that it stands ready to make a quantitative debt adjustment for the likely renewal period, the likely RVG payout, or both. [10] [9]

A second Meta vehicle followed the same pattern at a wider price. Reporting in mid-2026 describes Sopaipilla Investor notes of about $12.5 billion for an El Paso campus, issued around a 7.5 percent yield (roughly Treasuries plus 290 basis points) versus Hyperion’s 6.58 percent. Same 16-year backstop logic; the market charged more for the second use. [11] [12]

What this means for a competitor. The one-notch concession is cheap only while appraisal headroom and a AA− tenant exist. The structure fails the moment the tenant’s rating, the residual appraisal, or the “not reasonably certain” renewal judgment is challenged. Equity sponsors are paid to hold governance and first-loss equity, not to underwrite AI demand.

Stargate is an Oracle lease wrapped around an OpenAI contract

Stargate sites do not put OpenAI on the project debt. Developers and their lenders own the buildings. Oracle signs the long lease. OpenAI is the economic user of the compute Oracle sells. Lenders are underwriting Oracle; Oracle is underwriting OpenAI. That double layer is why construction loans can look like project finance and still trade as a hyperscaler credit—until Oracle’s own spread widens.

Site	Developer / owner vehicle	Reported project debt	Tenant chain	Notable risk allocation
Abilene, Texas (flagship, ~1.2 GW)	Crusoe with Blue Owl Real Assets and Primary Digital Infrastructure	Phase construction loans of about $2.3 billion then $7.1 billion, JPMorgan-led; FT counted roughly $13 billion invested by Blue Owl and JPMorgan, including about $10 billion of debt, into the SPV	Oracle lease → OpenAI	Build-to-suit on Lancium’s campus; equity in the JV, banks in construction debt
Shackelford County, Texas and Port Washington, Wisconsin	Vantage Data Centers	About $38 billion in two senior secured facilities (~$23.3 billion Texas, ~$14.8 billion Wisconsin); four-year maturity plus a two-year extension option; about SOFR plus 250 basis points; interest-only in construction, then amortizing	Oracle for OpenAI	Bank syndicate (JPMorgan, MUFG, Wells Fargo, BNP, Goldman, SMBC, Société Générale); syndication stretched by single-tenant concentration limits
Doña Ana County, New Mexico (Project Jupiter)	STACK Infrastructure (Blue Owl)	About $18 billion of bank construction loans; roughly $3 billion of Blue Owl equity underneath	Oracle; cannot terminate the lease	Power delivery is Oracle’s job. A force-majeure notice can defer rent, but Oracle still owes carry costs covering debt interest and a Blue Owl equity return for up to about three years. Loans quoted around 89–91 cents
Saline Township, Michigan (“The Barn,” >1 GW)	Related Digital; Blackstone-affiliated equity	About $16 billion package; roughly $14 billion of bonds at a 7.5 percent coupon, maturity around 2045, PIMCO anchoring about $10 billion; Blackstone equity about $2 billion	Oracle for OpenAI	Blue Owl declined an earlier version. Banks retreated; a bond fund set long-dated terms matched to the lease. Power customer is an Oracle subsidiary

[13] [14] [15] [16] [17] [18] [19] [20] [21] [22]

Oracle’s own filings show how large the lease layer has become relative to on-balance-sheet debt. As of May 31, 2026, Oracle reported $260 billion of additional lease commitments, substantially all data-center related, expected to commence between fiscal 2027 and 2029 for terms of 15 to 19 years, not yet on the balance sheet, including one lease with a lessor-borrowing guarantee of up to $3.3 billion. By August 31, 2026, the uncommenced lease figure cited from the 10-Q was $288 billion, against about $125 billion of notes and borrowings already on the books. [23] [24]

Default recourse on these SPV loans is to the project—land, building, and in some structures the chips—not to the developer’s corporate parent. It is not, in form, recourse to Oracle either. In substance, date-certain rent, non-termination clauses, and carry-cost obligations put delay and tenant risk back on Oracle. Jupiter is the stress test: permitting slippage on a gas pipeline did not release Oracle from funding the capital structure. [18] [15]

What this means for a competitor. A neocloud or developer cannot replicate Stargate by copying the SPV. The rating and the syndication depend on an investment-grade lessee who will pay even if power is late. Weaker tenants clear only with a hyperscaler or chip-vendor backstop, and they clear at much wider spreads. A September 2026 survey of single-campus bonds showed hyperscaler or investment-grade tenant coupons roughly in a 5.70–7.875 percent band, Google-guaranteed Fluidstack leases around 6.19–7.75 percent, and CoreWeave-leased paper around 7.00–9.875 percent. [25]

xAI’s SPV finances the chips, not the building

xAI’s public structure is a hardware lease, which matters because GPU residual value depreciates on a three-to-five-year product cycle, not a 25-year shell life.

The disclosed transaction is Valor Compute Infrastructure, managed by Valor Equity Partners. Apollo-managed funds led a $3.5 billion capital solution for a $5.4 billion purchase of compute, including Nvidia GB200 GPUs, leased to an xAI subsidiary on a triple-net basis to support training for Grok. Nvidia invested in the vehicle as an anchor limited partner. Apollo described the deal as downside-protected and asset-based. A later fund document describes a related VCI vehicle leasing GPUs to SpaceX on a triple-net basis, with a SpaceX guarantee after SpaceX’s acquisition of xAI, and a private investment-grade rating on $3.5 billion of debt at a 7.375 percent coupon maturing in 2031. [26] [27]

Earlier reporting, before that close, described a larger raise of about $20 billion for Colossus compute—roughly $7.5 billion of equity and as much as $12.5 billion of debt—in an SPV that would buy Nvidia processors and rent them to xAI for about five years, with xAI’s commitment limited to rent. Those figures were pre-close and should not be stacked on top of the $5.4 billion Apollo transaction. [28] [29]

Risk sits differently than at Hyperion. Lenders are secured by the GPUs and a five-year triple-net lease, not by a 16-year residual-value guarantee sized to repay 24-year bonds. Equity (Valor, Nvidia, and other LPs) eats obsolescence and re-lease risk after the term. Nvidia’s equity check also creates a round-trip: the chip vendor funds the buyer of its own GPUs, which supports shipments but concentrates residual risk in a vehicle whose only customer is one Musk company. [30]

CoreWeave has used a related pattern at smaller scale. In March 2025 it said an SPV created to fulfill an up-to-$11.9 billion OpenAI contract would “incur indebtedness to finance its obligations,” and later borrowed further against OpenAI contracts. CoreWeave’s own credit is below investment grade, so project lenders are underwriting the offtake contract and the hardware, not a AA tenant. [29] [31]

Vendor guarantees are the other off-balance-sheet channel

Where the tenant is not itself investment-grade, the chip vendor or cloud incumbent writes the credit support and accounts for it as a derivative or a contingent guarantee, not as debt.

• Alphabet. As of June 30, 2026, Alphabet reported about $43.8 billion of credit-derivative notional tied to data-center backstops, up from $6.5 billion at the end of September 2025, plus $7.6 billion of maximum potential payments under financial guarantees and an estimated $24.1 billion of additional future backstops still subject to final terms. The fair value recorded on the balance sheet was about $815 million. The mechanism is a promise to cover lease payments if a tenant such as Fluidstack defaults, which lets the landlord’s lenders price closer to Google than to the tenant. Alphabet also had about $85.2 billion of yet-to-commence data-center leases at the end of June 2026. [32] [33] [32]
• Broadcom. Maximum potential liability under one AI-rack lease backstop for Anthropic is about $29 billion. The exposure is not a full cheque: if the customer defaults, Broadcom owes the gap between 85 percent of outstanding backstopped lease amounts and whatever the racks fetch. No payments had been made at the last filing cited, and fair value was described as not material. [34] [32]
• Nvidia. The largest single disclosed cap is guarantees of about $105 billion (about $108.5 billion maximum gross exposure in one primary-source compilation) supporting an OpenAI campus that SB Energy is developing in Ohio, roughly 4.25 GW. The guarantee phases in as leases commence, runs against 20-year OpenAI leases, and declines as rent is paid. Nvidia has also described limited residual-value support—up to 25 percent of an opportunity, project by project—as a complement to independent underwriting, not a full wrap. [34] [34] [32]

These are not homogeneous liabilities. A filing compilation of primary-source maximums put Nvidia, Broadcom, Meta’s RVG threshold, and Alphabet’s guarantees and derivative notionals at roughly $217 billion currently disclosed, or about $241 billion including Alphabet’s not-yet-final backstops. That is a ceiling stack, not debt outstanding, and not the same number as SPV principal. [32]

How big the channel is depends on which number you mean

Three published aggregates measure different things. They should not be added.

• SPV principal already raised. A Financial Times analysis in December 2025 found more than $120 billion of debt and equity supplied by PIMCO, BlackRock, Apollo, Blue Owl, JPMorgan and others into SPVs for Meta, xAI, Oracle, and CoreWeave. The named pieces—Hyperion’s ~$30 billion, Abilene’s ~$13 billion, Vantage’s $38 billion, Jupiter’s $18 billion, xAI’s chip SPV, CoreWeave’s OpenAI vehicle—are the core of that figure. Later 2026 closes (Michigan’s $16 billion, Meta’s reported El Paso deal) sit outside that December snapshot. [29] [15]
• Contingent guarantees. A September 2026 Financial Times report, as relayed by secondary write-ups, put Big Tech AI-infrastructure exposure kept off balance sheets through guarantees and SPVs at roughly $300 billion. That scope includes residual-value guarantees and vendor backstops, not just SPV debt. [35] [34]
• Signed leases not yet on the balance sheet. Moody’s, in February 2026, put undiscounted future data-center lease commitments of Amazon, Meta, Alphabet, Microsoft, and Oracle at $969 billion at the end of 2025, of which more than two-thirds—about $662 billion—had not commenced and therefore was not yet a balance-sheet liability. Moody’s said that uncommenced amount was about 113 percent of the five companies’ combined adjusted debt. Many of those leases are the offtake that supports the SPVs above; some are ordinary colocation. They are obligations of the tenant, not a measure of project-finance principal. [10] [36] [37]

Forward estimates are larger still and should be labeled as forecasts. Morgan Stanley said in October 2025 that roughly $1.5 trillion of external financing would be needed across the AI ecosystem, of which as much as $800 billion could come from private credit in asset-tied deals, including SPVs, by 2028, against a $2.9 trillion global datacenter spend estimate through 2028. [38] [39]

Who actually bears each risk

Construction risk is contractually pushed to the tenant or a guaranteed-maximum-price contractor wherever the rating depends on it. On Beignet, S&P’s explicit premise is that Meta absorbs overruns and delays so the project is not exposed to ordinary construction risk. On Jupiter, Oracle remains liable for carry costs if the site is not ready, including if power slips. Equity sponsors (Blue Owl, Vantage, Crusoe, Related, Blackstone) bear completion risk only to the extent the lease does not start paying and the tenant’s delay protections do not cover them. Banks that cannot syndicate—Jupiter loans stuck with underwriters, quoted in the high 80s to low 90s—are already wearing that execution risk at a mark. [4] [18] [24]

Tenant credit risk is the rating. Clifford Chance’s 2026 review of the data-center bond wave describes three buckets: the bond tracks the tenant when an investment-grade hyperscaler takes a date-certain lease through maturity and stands behind rent with an RVG or absolute guaranty; the bond is notched down for structural gaps; the bond is high-yield when hyperscaler support is deferred or absent. Every agency treats tenant credit as the ceiling. In Stargate, that tenant is Oracle, not OpenAI. If OpenAI cannot pay Oracle, the project lenders still look to Oracle’s lease—until Oracle itself is stressed, which is what the Jupiter marks and the Michigan bank retreat already price. [40] [41]

Residual-value risk is the risk the structures were built to move, and it has not fully moved. On Hyperion, bondholders are protected for 16 years by Meta’s declining guarantee and thereafter by a termination fee; Blue Owl’s equity is the cushion above that floor and the owner of whatever value remains if Meta stays. S&P currently assigns the RVG zero incremental corporate debt because it believes stressed asset value still covers the notes. Meta records nothing because a payout is “not probable.” Those two judgments can diverge from cash reality at the same time: the guarantee is most likely to be called in the scenario where the appraisal headroom disappears. On GPU SPVs, there is often no equivalent 16-year floor. Lenders have a five-year lease and hardware collateral; equity and, where offered, a capped Nvidia residual-support slice take obsolescence. On Alphabet and Broadcom backstops, the vendor takes the gap between contractual coverage (or 85 percent of remaining rent) and recovery value, and books only a modeled fair value. [10] [9] [34]

The practical map is therefore not “debt has left the hyperscaler.” Construction risk sits with whoever signed the completion and delay protections, usually the tenant. Tenant risk sits with the lessee and, one level down, with that lessee’s unrated customer. Residual-value risk sits with the guarantor for as long as the guarantee is in the money, with project equity for the first loss above that floor, and with lenders only if both the tenant and the guarantee fail or the collateral is worth less than the notes after enforcement. Accounting keeps the first two largely off the balance sheet until leases commence and a payout is probable. Rating agencies are already debating whether to put them back into adjusted leverage. The spread between Hyperion’s A+ at Treasuries plus 225 basis points in October 2025 and wider prints on the second Meta deal, CoreWeave-tenant bonds, and sub-par Oracle project loans is the market’s running vote on how much of that risk was actually transferred.

---

Recent Findings Supplement (October 2026)

Meta-Blue Owl Hyperion (Louisiana) remains the clearest, largest, and most documented template for off-balance-sheet AI data center project finance, with fresh 2026 confirmations on scaling, lease terms, guarantees, and accounting treatment.[1][2]

The structure uses Beignet Investor LLC (a bankruptcy-remote SPV/JV): Blue Owl funds hold an 80% interest and Meta 20%. The vehicle owns the assets and issued ~$27.3 billion in senior secured notes (due 2049, ~6.58% coupon, priced at ~T+225 bp, A+ rated by S&P—one notch below Meta’s AA-). PIMCO anchored (~$18B), with BlackRock and others participating; Morgan Stanley arranged. Meta contributed land/CIP (~$5.7–5.8B equity/in-kind) and receives a one-time distribution (~$3B); Blue Owl contributed ~$7B cash equity. Total committed capital base ~$28.8B.[3][4]

Meta leases via operating leases (4-year initial term per building, with renewal options that can extend to ~20 years total). Rent obligations begin June 1, 2029 (even during construction delays, with narrow force majeure abatements). Meta provides a payment guaranty plus a Residual Value Guarantee (RVG, initial threshold ~$28B, declining over time). If Meta does not renew/terminates and asset value falls below the threshold, it covers the shortfall. Creditors have no direct recourse to assets.[1][5]

Accounting/rating treatment: Structured as operating leases (short initial term + options help avoid finance lease classification under GAAP). Meta does not consolidate the VIE/SPV. The A+ rating reflects Meta’s credit with a modest notch for structural/execution risk. This keeps tens of billions off Meta’s balance sheet while converting capex into lease obligations.[3]

Project updates (mid-2026 onward): Scaled to 5 GW across ~3,200 acres (first 2 GW targeted online by 2030, full by ~2032–2036); total investment expectations rose significantly from the initial ~$10B announcement (now often cited in the $27B+ financing range or higher aggregate). Similar structures are viewed as replicable templates.[1][6]

Stargate-linked sites (notably Oracle/Blue Owl/Stack Project Jupiter in New Mexico) highlight execution and tenant-credit risks in construction-phase project finance.[7]

~2.45 GW campus (part of the broader OpenAI/Oracle/SoftBank Stargate initiative). Developer: Stack Infrastructure (Blue Owl portfolio). Financing includes ~$18B in bank construction loans (syndicated by ~20 banks, e.g., Santander/Jefferies involvement; quoted at 89–91 cents on the dollar pre-notice) plus Blue Owl equity (~$3B). Oracle is the tenant under a long-term lease (power responsibility on Oracle; limited termination rights).[8]

September 2026 development: Oracle sent a force majeure notice to Blue Owl/Stack citing permitting delays on a needed natural gas pipeline, seeking to defer rent start (originally targeted ~2028). This does not eliminate ultimate obligations but extends timelines and carry costs. Loans traded distressed; equities reacted negatively. Underscores construction/permitting risk borne primarily by the developer/tenant side, with heavy reliance on Oracle’s credit (recently downgraded).[7][9]

Broader Stargate financing involves multi-party JVs, partner-level debt, and vendor commitments across sites (e.g., Michigan “The Barn” campus with Related Digital/Blackstone equity and bond financing).

Crusoe/Abilene (Texas) and related sites use JV/project-level debt with hyperscaler tenants (OpenAI/Oracle historically; Microsoft expansion).[10]

Abilene campus: JV among Crusoe, Blue Owl, and Primary Digital Infrastructure (~$15B total; construction financing in the $7B+ range reported earlier). Leased initially to OpenAI/Oracle (1.2 GW phase); Microsoft confirmed for 900 MW expansion (adjacent buildings + on-site generation/BESS). Power mix: grid primary + gas backup/turbines.[11]

Recent updates (2026): Crusoe filed for additional Jayton, TX “spur” buildings (~$4.8B investment, part of broader Project Hyper/Childress-linked plans; construction start ~early 2027). Crusoe closed initial $3.9B Series F equity (~$30.9B post-money valuation, Sept 2026). Abandoned a $1.25B Boom Supersonic turbine deal (Sept 2026) while maintaining other gas generation plans.[12][13]

Vantage developments center on corporate/development-stage facilities rather than pure tenant-specific project finance SPVs.[14]

September 14, 2026: Closed a $2B 5-year revolving development credit facility (extension options) for early-stage North American projects. Initial collateral: three unnamed development assets (expandable pool). Funded by a dozen insurance/institutional investors (arrangers: Evercore, Wells Fargo). Functions as a “warehouse” for land/sitework/early construction before permanent financing. Part of >$40B total capital raised by Vantage in 2026. Separate securitizations of operating leased assets (tenant leases as collateral) carry A-(sf) ratings.[15]

This supports developer balance sheet flexibility but does not shift tenant credit or residual risks in the same SPV/lease-guarantee manner as Hyperion.

xAI structures appear more equity- and direct-build focused, with limited public detail on large SPV/project-finance leases or RVGs.[16]

xAI has pursued rapid own-build Colossus clusters (Memphis-area and others) with mobile/gas turbine power and GPU leasing to third parties (e.g., Anthropic multi-year full capacity; Google ~110k GPUs starting Oct 2026; short termination clauses in some deals). Equity raises (e.g., $20B+ Series E). Broader reporting groups xAI among users of SPVs to move data center/compute debt off-balance, but specific legal structures, leases, or guarantees are not detailed in recent sources. One related vehicle (Valor/Apollo-backed) uses triple-net leases for GPUs supplied to xAI.[17]

Broader channel size and risk allocation (new 2026 estimates and disclosures):

• Single deals like Hyperion (~$27B debt) and Jupiter (~$18B loans) are prominent. Aggregate: reports cite >$120B in data center financing debt shifted via SPVs; Big Tech guarantees/RVGs supporting ~$200–300B total AI infrastructure exposure (e.g., Meta ~$28B RVG; Nvidia $105B cap on Ohio PORTS-Pike/OpenAI campus via SB Energy; Broadcom ~$29B; Alphabet guarantees rising to $43.8B).[18][16]
• Who bears what: Construction/permitting/execution risk often on developer or tenant (e.g., Oracle power obligations; delays triggering notices). Tenant credit (hyperscaler IG or near-IG) is central to pricing and ratings (one-notch concessions common). Residual-value risk shifted to Big Tech guarantors via RVGs (covers shortfall on re-lease/sale if tenant exits and values fall due to obsolescence). Payment guarantees ensure cash flow continuity. Accounting keeps most debt/leases off sponsor balance sheets (operating classification + VIE non-consolidation); guarantees disclosed with low probability assessments in some cases.[7]

Implications for entrants/competitors: These structures lower sponsor capex intensity and leverage metrics but embed long-term tenant dependence and residual risk on Big Tech balance sheets. Execution surprises (permitting, power) can pressure secondary debt pricing and highlight the limits of credit support. New development facilities (e.g., Vantage) and equity raises provide alternative capital but at earlier, riskier stages. No major regulatory or policy shifts on lease accounting or SPV treatment noted in post-April 2026 sources.[7]

Sources are drawn from 2026-dated reporting (primarily July–October) on the specific transactions and market commentary; earlier foundational announcements are referenced only for context where needed to explain current structures.