AI expansion is concentrating hidden obligations among counterparties while conventional balance sheets understate who ultimately bears the risk.

OpenAI's confidential IPO filing shows zero debt and only $46 million in quarterly capital expenditure, making it look like a lean software business. It is not. The same filing discloses $665 billion in future procurement commitments for chips, energy, and data centres, the vast majority sitting off the balance sheet. The IPO has been pushed to 2027. Revenue growth cannot yet digest what has already been contracted. The trap was built before the company was ready to go public.

UK and US pension funds, retail investors in tech ETFs, and SoftBank shareholders are already exposed. SoftBank holds a $65 billion stake in OpenAI and took a $40 billion bridge loan to fund it, with repayment due March 2027. OpenAI burned $3.7 billion in a single quarter against $5.7 billion in revenue. Its own CFO reportedly questioned whether the company was financially ready to list. If the IPO slips past SoftBank's debt deadline, the collateral chain tightens on every retail fund holding SoftBank stock.

The Law of the Trap operates here with precision. OpenAI presents a clean cover page and buries $665 billion in footnotes. Investors see asset-light. The filing says otherwise. Most people will not read the footnotes. That is the design. The gap between what the headline says and what the footnote confirms is not an accident. It is the product. When the market finally reads the small print, the price discovery will not be gentle.

S&P cut Oracle's long-term credit rating to BBB-, one notch above speculative grade, on 9 July 2026, explicitly naming OpenAI as a key credit risk. OpenAI accounts for roughly half of Oracle's $638 billion in remaining performance obligations. Oracle spent $55 billion in capital expenditure in fiscal year 2026, turning free cash flow negative by $24 billion. S&P forecasts that deficit widening to $42 billion in 2027. Oracle's five-year credit default swap spread hit its highest level in 18 years. The market is pricing the risk the balance sheet is not showing.

Oracle stock has fallen 33% in 2026 and trades near 52-week lows. Oracle raised $43 billion in debt plus $5 billion in equity during fiscal 2026, with a further $40 billion planned for 2027. Its total debt stands at $167 billion. If OpenAI faces financing difficulties, Oracle would be left holding data centre leases it cannot exit. For UK and US investors in Oracle bonds, investment-grade protection has already thinned to its minimum. One further downgrade moves Oracle debt into junk classification, triggering forced selling by investment-grade-only funds. That repricing lands on pension holders.

The Law of the Addict: Oracle committed to a spending programme so large that stopping would crater the revenue it has already promised. Its AI cloud revenue is 27% of total revenue today, projected at 60% by 2028. The company has laid off 21,000 people to fund infrastructure. That is not expansion. That is substitution. When an entity cannot stop spending without collapsing the rationale for the spending, it is not investing. It is dependent. The market is beginning to charge the premium that dependency deserves.

Meta and Blue Owl Capital closed the largest private credit transaction in history in October 2025: a $27 billion joint venture for the Hyperion data centre in Richland Parish, Louisiana. Meta owns 20% and retains full operational control. Blue Owl owns 80% and holds the debt. PIMCO anchored $18 billion of the bond issuance. BlackRock took $3 billion. Meta is now running the same structure again for a new data centre in El Paso, Texas, with Morgan Stanley and JPMorgan leading a package expected to exceed $13 billion. Each campus Meta controls. None of it touches its balance sheet.

The Hyperion campus will draw up to 5 gigawatts of electricity when fully operational, the equivalent power consumption of 4 million American households. That grid pressure lands on local ratepayers in Louisiana and Texas, not on Meta's quarterly report. The bond issued by the SPV carried a yield of 6.58% at issuance, closer to high-yield territory than the A+ rating suggested. PIMCO, BlackRock, and pension funds holding this paper bear the risk if Meta's AI revenue does not materialise. Meta has already provided a capped residual value guarantee covering 16 years of operations, meaning its liability exists, it simply does not appear on the balance sheet.

The Law of the Landlord: Meta uses the asset without holding the liability. The community hosts the infrastructure, absorbs the grid strain, and has no claim on the returns. Bondholders and insurers carry the risk of a demand shortfall. The accounting keeps Meta's leverage ratios clean while the actual obligations sit inside a vehicle most retail investors will never examine. When Ernst and Young, Meta's own auditor, flags a red flag on this structure in the annual report, the signal is not ambiguous.