If AI is too complex to comprehend, so are the jargons that surround the business models of AI companies. Not keeping track of them could mean being left out or staying behind the curve when it comes to what is apparently the most disruptive and transformative force the world has seen since the industrial revolution. It probably is the most disruptive force the investing world has seen as well. Keeping this in mind, at bl.portfolio, we have been demystifying the AI world via a series of articles in this Big Story section namely, ‘Too big to fail’ (September 21, 2025), ‘When bits and bytes take a large bite’ (May 31, 2026), ‘Welcome to the world of AI-nomics’ (July 12, 2026) and ‘AI’s dotcom déjà vu’ (August 2, 2026).

Today, we deep dive into a few key aspects that have been making ripples recently, when it comes to the AI theme. We are talking about off balance sheet debt and variable interest entities.

These have become too big to ignore that quite a few experts have voiced their concerns on them. Big Short investor Michael Burry, who is known to have short positions on AI stocks such as Nvidia, Micron Technology and Nebius, recently warned in his newsletter that the off balance sheet liabilities are in ‘hypergrowth mode’ and added, ‘When the music’s over, these off-balance sheet commitments become real liabilities very quickly’. Another Big Short investor Steve Eisman, had expressed that the off balance sheet debt and the existence of variable interest entities or VIEs reminded him of the Enron scandal.

So, what is up with these jargons? We unpack here.

Off balance sheet debt

Off balance sheet debt or off balance sheet obligations ( sometimes called hidden debt too) is simple, if one understands the substance of these transactions and why they are disclosed off the balance sheet.

In so far as hyperscalers are concerned, off balance sheet obligations appear in two forms:

* Leases (largely for data centres) that are yet to commence

* Obligations to purchase network infrastructure (including GPUs, CPUs, memory chips, servers and other network gear) and long-term agreements to procure energy to run the data centres.

Why off balance sheet?

Financial statements are prepared by following certain rules. These rules are called accounting standards. Companies need to statutorily comply with them and doing so enables an investor to compare the financials of two companies.

For assets and liabilities to be recognised on the balance sheet, they need to satisfy a few conditions — technically called recognition criteria. For example, to book an item as inventory, it should be a tangible asset and is either held for sale in the ordinary course of business or consumed during production of other goods or in the process of rendering a service.

As far as leases go for certain long-term assets, the standards stipulate that under certain conditions, the lessee must simultaneously capitalise a lease asset (called right of use asset) and the present value of all future lease payments as a liability, on either sides of the balance sheet respectively. This is called initial recognition. How the lease asset and lease liability are accounted for subsequent to initial recognition is more complex and need not be focused now. The standards also prescribe the timing of initial recognition on a day termed as the ‘commencement date’. This is the day on which the lessor makes the leased asset available for use by the lessee. Hence, lessees must wait until the commencement date to record the lease asset and lease liability.

In case of hyperscalers, they are the lessees of data centres, which take years to build. Since these leases are contracted for already, companies disclose these material transactions in the notes to accounts (off the balance sheet), until the data centres are delivered for their use by the developer/ lessor.

Similarly, obligations to purchase network/ compute infrastructure are also disclosed in the notes, till they meet the recognition criteria meant for fixed assets. You can check out Amazon’s off balance sheet obligations in page 15 here: https://tinyurl.com/muhnka9t

How the leases are structured

As an example, we shall look at the structure of Project Jupiter, which is part of the $500-billion Stargate project. At the core are four large data centres, co-developed by STACK Infrastructure and BorderPlex Digital Assets. STACK is ultimately controlled by Blue Owl Capital, a private credit firm. The project is being developed in Doña Ana County of New Mexico State in the US.

Once built, the developers transfer the project assets temporarily to the county for 30 years and in turn, the county leases them back. This is to make use of a tax-efficient model to avoid paying high property taxes (property title is with the county itself for 30 years). PILOTs (payments in lieu of taxes) replace property taxes and act as consideration for the 30-year lease.

Once up and running, the data centre is contracted to be leased to Oracle. Oracle pays Project Jupiter out of the revenue it generates from OpenAI’s use of compute. The deal between OpenAI and Oracle is a $300-billion multi-year one.

Why is the street alarmed?

The off balance sheet liabilities of major hyperscalers are given in the graphic. Together, they add up to $2.9 trillion in undiscounted future payments (notwithstanding time value of money) — for leases, network gear and energy. Against this gigantic figure, the current cash balance (net of debt) of these generally cash-rich companies sits at a mere $126 billion. Their combined TTM operating cash flows work out to $724 billion, meaning unless cash flows grow/ sustain from potential AI-related revenue, the hyperscalers may not be able to pay for these obligations. Some of them have even turned to debt, for the first time in years and debt issuances of these companies in 2025-26 has touched about $380 billion.

It is to be noted that a lot of the obligations are non-cancellable or on take-or-pay basis, and this seems to be the reason behind Oracle issuing force majeure (expressing inability to discharge obligations due to circumstances out of a party’s control) to STACK Infrastructure in late September. Contractually, Oracle is still liable to pay the developer, even if the first phase of the project is not completed by the H1 2027 deadline, per Financial Times. The setback seems to have stemmed from permit delays, amid local opposition. The two parties have since clarified that the project is still on.

On the income side, the companies are banking on massive revenue from compute deals. These are called remaining performance obligations (RPOs) or simply, revenue backlog, and they total to $2.5 trillion. However, about 40 per cent is expected from OpenAI and Anthropic alone — a serious concentration risk.

In 2025, Anthropic is reported to have generated a revenue of $4.6 billion, lost $8 billion on operating level and a net loss of a staggering $42 billion. These are as per a confidential prospectus accessed by Reuters. In the same year, OpenAI had earned $13 billion in revenue, incurred an operating loss of $21 billion and a net loss of $39 billion. These figures are as reported by Ed Zitron, an independent journalist and an outspoken critic of AI companies. At this rate, the two AI labs would like their IPOs to happen soon and at desired valuations, helping them sustain the cash burn. While Anthropic may go public by November-end, OpenAI has postponed its IPO to 2027.

Steve Eisman had an interesting argument about why Anthropic may be pushing for a November IPO. An IPO now would mean the latest revenue figures in the prospectus would likely be that of H1 2026, which also marks the end of tokenmaxxing — a trend where employees were incentivised based on the number of tokens consumed, as a proxy for productivity, rather than actual outcomes.

Token prices for users are declining as well. Measured by the Ramp AI index, Anthropic’s token price has sharply declined from $2.56 per million tokens in January to $0.84 now. OpenAI’s has fallen from $0.95 to $0.64. In this backdrop, per the Financial Times, OpenAI, as late as Thursday, has clarified that its annualised revenue run rate is only $50 billion against the Axios-reported $70 billion. In reaction, Oracle’s shares fell 5.5 per cent, yield on its 10-year bond climbed to 7.6 per cent (coupon was 5.7 per cent on issue in February 2026), CDS spread thereon has widened 1.76x year to date and is at all-time highs.

Variable interest entities

Before moving on to the concept of VIEs, let’s quickly brush up the basics of consolidation. Assume two companies P and Q. P consolidates Q if it has controlling interest in Q. The usual way to check if P has a controlling interest is to check if it has a majority stake (more than 50 per cent voting power) in Q. There could be other ways by which an entity can control another, even without a majority stake. For example, P has the power to appoint a majority of the members on the board of Q, through an agreement. This can be example of controlling interest as well.

To keep it very simple, when a company consolidates its controlling interest in another, irrespective of its voting share in the latter (51 per cent stake, 90 per cent stake, etc.), 100 per cent of the latter’s assets, liabilities, revenue and expenses are added to those of the former. There are also consolidation adjustments like elimination of inter-company balances, accounting for minority interest and so on. But a discussion on them is not warranted now.

In case there is no controlling interest, P accounts for its investment in Q using equity method or at the fair value of its stake, as given in the chart. Do note the difference here. Unlike above, where 100 per cent of Q’s assets, liabilities, etc. become P’s on consolidation, under equity method or when recording at fair value, Q’s assets, liabilities need not be added. The investment in Q merely reflects the value of P’s partial stake in Q.

Companies can, at times, be creative and structure deals such that the other company’s liabilities (Q’s) are kept off their books. Optically, there may be no controlling interest, but not so in substance. The accounting standards (US GAAP) provide for such cases and lay down the rules when a company must consolidate its variable interest in an entity.

Consolidation criteria

Without getting into too much detail, readers can understand variable interest as follows. A variable interest is any stake in an entity whose value goes up or down as the entity’s value goes up or down. Under US GAAP, advancing loans and financial guarantees are also considered variable interests, besides outright equity stake.

The standards stipulate that when a company’s variable interest in another (VIE) satisfies both the following conditions, the company is termed the primary beneficiary of the VIE. On becoming the primary beneficiary, it must consolidate the VIE’s asset, liabilities, revenue and expenses in full.

a) Whether the company has the power to direct activities that most significantly affect the economic performance of the VIE

b) Whether the company absorbs losses or receives benefits that could potentially be significant to the VIE

Take a look at the illustration. A has a small 9-per cent stake in B that doesn’t outright warrant A consolidating B. But take a closer look at B’s balance sheet. Of the $2.1 billion of total assets that B owns, $2 billion of them have been financed by a loan that A arranged, by providing a guarantee to the bank, in case B defaults. Owing to this deal, B agrees to let A decide to whom the compute is leased. Thus, A directs B’s activities that most significantly affect its economic performance (leasing compute) and also absorbs B’s potential losses, in the event when B defaults and A repays the bank. With these conditions satisfied, A becomes the primary beneficiary of B and A is required to consolidate B.

That said, a key point to note here is that the standards require the management of companies themselves to use their judgment to determine which activities most significantly affect the economic performance of the VIE, among others.

The Enron story

Market voices have expressed concerns that these kinds of VIE structures remind them of the Enron scandal. Enron is known to have floated special purpose entities (SPEs) to keep debt off its balance sheet. Often Enron’s own C-suite were made the owners of these SPEs, with a minor low single-digit stake held by unrelated parties. Enron largely didn’t have any direct stake in the SPEs. The SPEs then borrowed from banks, backed by Enron’s guarantee. Next, with the loan funds, the SPEs bought Enron’s loss-making investments off its balance sheet, while Enron booked revenue on the sale. Thus, Enron managed to keep losses and debt off its balance sheet. Legally, it wasn’t required to consolidate the SPEs, as regulations were not as stringent then. Particularly, Enron took advantage of a rule that exempted consolidating an SPE, if independent, unrelated entities hold at least 3 per cent of the SPE’s capital.

Today, regulations and disclosure requirements are much stringent and sophisticated, and are to some extent, products of the Enron scandal. In our view, comparing today’s VIEs to Enron’s SPEs may be far-fetched. But there may be lacunas in this space, when the managements’ judgments determine whether they are the primary beneficiaries or not. The managements have an inherent incentive to opine that VIEs need not be consolidated, protecting their debt metrics from appearing alarming. The onus therefore falls on the independence of auditors.

Real life examples

Going through the 10-Qs and 10-Ks of the hyperscalers reveals a divided picture. Microsoft and CoreWeave have concluded that they are not the primary beneficiaries of any of their VIEs, thus taking the equity method or the fair value route (as applicable) to account for them. Amazon has disclosed that its variable interest/ stake in Anthropic and OpenAI do not call for a consolidation, which appears logical. On the other hand, Alphabet has determined it to be the primary beneficiary of some VIEs (consolidating them, of course), while not so for the rest.

However, the most interesting case has to be that of Meta and its variable interest in the Hyperion data center. While Meta has concluded that it is not the primary beneficiary, its auditors, EY have pointed the same as a critical audit matter (CAM), noting that they found the audit of Meta’s determination of the primary beneficiary of the VIE was especially challenging. Basically, auditors use the CAMs section to invite the attention of stakeholders to areas of audit that involve complex judgments or were challenging. However, a CAM does not amount to an audit qualification.

Let’s examine the case now. The data centre is owned 80:20 by Blue Owl and Meta (no outright majority stake). The deal is such that Hyperion will initially lease the data centre to Meta for $12 billion over a four-year period. Meta will also have the option to renew the lease for up to 20 years. This opens it up for interpretation whether Meta exercising the option would mean it has effectively directed Hyperion’s activities for the next 20 years or not.

That is not all. The deal also includes a residual value guarantee (RVG) element. RVG ensures Hyperion will be indemnified by Meta, in case Meta terminates the lease or it fails to renew the lease. The indemnity value is as per the formula given in the chart. The initial RVG threshold is set close to the development cost but will taper down with time.

Here is a simulation. Assume at the end of the second year of the lease, Meta wants to call the deal off due to a slowdown in AI usage. Let the RVG threshold then be taken as $25 billion. Hyperion assesses the fair value of its business at $10 billion. Here, Meta has to indemnify Hyperion to the tune of $15 billion. This too, makes it open for interpretation whether the RVG in a way, absorbs potential losses of the VIE (Hyperion) — the second condition for determination of primary beneficiary. Further, there is only limited information in the 10-Qs/ 10-Ks, and right now, investors have no recourse than to rely on the auditors.

In a world where AI is highly capital-intensive and is increasingly turning the cash-rich companies towards debt, investors need to warm up to these concepts. After all, only then can they not say, ‘Didn’t see it coming’ when actually it was all in clear plain sight!

Published on October 10, 2026