Nvidia wants to build AI factories at gigawatt scale. Digital Realty wants to be the landlord. The two companies are deepening a collaboration centered on Nvidia’s AI Factory Research Center, housed at Digital Realty’s Northern Virginia campus. Together they are developing the Omniverse DSX blueprint, a framework for building scalable, gigawatt-class AI compute facilities designed for customers who are not Amazon, Google, or Microsoft.
The $500 billion question
On August 10, 2026, Nvidia announced memoranda of understanding with six major asset management firms, including Apollo and BlackRock, targeting more than $500 billion in third-party capital for AI infrastructure buildout.
The focus is explicitly on non-hyperscale environments, the tier of compute infrastructure below the mega-campuses run by cloud giants. These facilities serve enterprises, governments, and research institutions that need serious GPU capacity but are not building their own sovereign clouds.
Digital Realty is coming to the table with its own capital formation already underway. In March 2026, the company closed a $3.25 billion US hyperscale data center fund, which it says is designed to support more than $10 billion in new AI infrastructure investment.
Nvidia disclosed approximately $530 billion in gross off-balance-sheet guarantees as of the second quarter of fiscal year 2027. That figure reflects a growing web of commitments tied to AI and cloud-related projects, commitments that do not sit on Nvidia’s formal balance sheet but represent real financial exposure if things go sideways.
Why funding an AI factory is harder than it looks
GPUs are not like commercial real estate. A building holds its value over decades. An H100 GPU, which was commanding premium prices not long ago, has seen its market value drop sharply as newer chips arrive. Lenders who understand office towers and distribution centers are far less comfortable underwriting a data center whose primary collateral is a rack of chips that could be superseded in 18 months.
Nvidia is reportedly considering residual-value support mechanisms of up to 25% on specific projects, essentially telling lenders that if the hardware loses value faster than expected, it will absorb some of that downside. It is a reasonable tool, but it also means Nvidia is effectively backstopping the debt on its own products, which adds another layer to those already substantial off-balance-sheet figures.
Whether the six asset managers who signed MOUs in August can convert those letters of intent into actual capital deployment will be the near-term test. MOUs are a starting point, not a closing. The structural work of creating loan documentation, valuation frameworks, and secondary markets for AI hardware as collateral is still ahead.