Apollo's near-10% stake in Mubadala Capital makes Abu Dhabi the cheapest funder of AI data centers

The structure matters more than the percentage. Apollo's release describes a minority equity stake plus capital commitments across Mubadala Capital's strategies, and the $5 billion figure covers Apollo together with other strategic investors at initial close. Nothing in the announcement ties that money to a specific data center, a specific loan or a specific technology. Apollo bought a share of a distribution and origination platform, not a portfolio of assets — and a platform with more than $755 billion of assets behind it can be pointed at any credit strategy Mubadala chooses to launch.

Abu Dhabi used to buy the asset manager. ADIA took a 10% holding in Apollo before its 2011 initial public offering and later sold it down. In January 2023, Alpha Dhabi and Mubadala formed a partnership to co-invest in global credit. On Nov. 11, 2024, Apollo and Mubadala extended a partnership first formed in 2022 and put $2.5 billion into a joint venture to co-invest in global private credit. Now the asset manager buys into the sovereign's own manager, and the logic is not sentiment. For a manager, LP capital is the input that cannot be manufactured; a shareholder who arrives with $385 billion of sovereign assets and a Gulf distribution network converts a fundraising constraint into an origination advantage.

The timing was not coincidental. SpaceX reached a binding commitment for an approximately $40 billion master finance lease for GPUs and related equipment with an entity managed by Apollo, delivering NVIDIA's next-generation Vera Rubin chips for SpaceX's terrestrial compute buildout. Financing parties were led by Apollo, Goldman Sachs acted as sole structuring agent, and anchor commitments came from BlackRock-managed funds, Goldman Sachs Asset Management funds and KKR-advised capital accounts. The lease is expected to close before year-end 2026, with lease schedules beginning on delivery in the first half of 2027. A lease of this size needs equity behind the debt, and equity for asset-backed structures is the one input a private-credit manager cannot conjure from its own funds.

Apollo can fund its side of the commitment without strain: $26.8 billion of cash and cash equivalents, $419.1 billion of investments and $13.7 billion of total debt at June 30, 2026 make a $5 billion commitment an allocation rather than a stretch. That is exactly why it works as a lever on other people's money. The caveat belongs in the same paragraph: Apollo's fee engine and its GAAP bottom line are not the same thing. The company reported a $1.91 billion GAAP net loss in the first quarter of 2026, with $1.69 billion of income tax expense, in the same year it reported record fee-related earnings of $785 million in the second quarter. Apollo can fund its side of the commitment without strain — that balance-sheet capacity is what makes the Mubadala seat strategic rather than financial.

FactSet's July 2026 analysis put aggregate hyperscaler debt at roughly $700 billion, with Amazon placing a $25 billion bond and Oracle planning about $20 billion of debt in fiscal 2027 as part of a $40 billion debt-and-equity package. On Oct. 19, 2026, S&P Global warned that the largest hyperscalers likely carry more AI-linked debt than is publicly known. The buildout is no longer self-funded, and the fourth pillar of capital has just been added. Oracle is the cleanest case study: remaining performance obligations reached $664 billion at the end of its first fiscal 2027 quarter, up $209 billion year over year, yet the company sold $20 billion of stock through its at-the-market program in the same quarter and spent $28.5 billion on capital expenditure, against $8.5 billion a year earlier, while carrying $155.9 billion of total debt and $119.6 billion of net debt. The buildout is no longer self-funded — a backlog is a promise, and capital expenditure is cash.

On Oct. 16, 2026, the repricing arrived. CoreWeave's bonds yielded 15% to 16%, its credit-default swaps traded near 1,200 basis points, and LSEG put its debt at about $50 billion. SMBC's James Martin told the same roundup that investors are demanding more compensation and that lenders should require wider spreads, lower leverage, stronger covenants and tighter customer-concentration limits. In the same week, Carlyle published an estimate that private credit could be called on to finance roughly $1 trillion of AI compute, more than half the industry's current assets under management, and warned that rapid lending creates concentration risk of the kind that surfaced in software lending. Note what is happening: the public market is repricing the asset, while the private market that financed the same assets private books did not reprice with the bonds.

The financed buildout converts into orders that are paid in cash, and that is where the cycle shows up with the least ambiguity. Vertiv reported second-quarter 2026 net sales of $3,274 million, up 24% year over year on 18% organic growth, with operating profit up 44%, and raised full-year guidance to $14.0 billion to $14.2 billion of sales and 30% to 32% organic growth. Its backlog stood at $15.0 billion at the end of 2025, up 109%, on a book-to-bill of about 2.9x. Micron's fiscal fourth quarter ended Aug. 28, 2026 delivered $54.2 billion of revenue at an 86.7% gross margin, taking fiscal 2026 revenue to $133.2 billion, up 256%, with $33.2 billion of adjusted free cash flow in the quarter. SK hynix posted second-quarter revenue of 79.3 trillion won and operating profit of 60.5 trillion won, up 557% year over year at a 76% margin, and has finalised long-term agreements with around 10 customers. Suppliers of power, memory and cooling get paid in cash, not in promises, which is why the equipment cycle is the cleanest read on how much financed capacity is actually being built.

Advantest's first quarter of fiscal 2026 grew revenue 39.3% year over year on tester demand for AI and high-performance computing chips, and management said inference-AI test demand would significantly exceed its April 2026 projections, with its fiscal 2026 plan pointing to roughly ¥1.7 trillion of sales. Test equipment is bought out of fab and system budgets rather than out of the operator's credit line, so it carries far less borrower risk than the GPUs it validates. That asymmetry is why upstream is the cleaner expression of the same thesis.

The strategic logic only closes with demand. Mubadala's AI-infrastructure build-out runs through G42 and its data-center arm Khazna, which holds more than 70% of the UAE market, alongside Yondr Group, in which Mubadala invested in 2024 when the developer held 878MW of contracted and reserved capacity with 58MW live. In November 2025, Microsoft and G42 announced a 200MW expansion of Khazna capacity as part of an investment of more than $15 billion. This is sovereign-mandated, politically supported demand for compute: the least price-sensitive demand in the world, and the reason a sovereign balance sheet wants a seat in the financing rather than the assets.

Blackstone is running the same race from the other side of the Atlantic. Its digital infrastructure trust held $275 billion of data center and digital infrastructure investments as of June 30, 2026, it raised $1.75 billion for a data center real estate investment trust in a U.S. initial public offering in May 2026, and it has said it plans a public company to buy up ready-made AI data centers. The competition is no longer for the limited partners, not the loans: once sovereign capital buys into managers rather than projects, the scarce asset becomes the allocator's cheque, and the cheque is what funds the origination pipeline.