Anthropic, Open AI and the hyperscalers: The balance sheets that threaten to throw AI ‘off’ balance

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.

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.

• 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.

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

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.