Conversations surrounding AI have been dominated by model launches and benchmark scores. Every week seems to bring news of another breakthrough model or application promising to transform how businesses operate.
Beneath that excitement sits a less visible story. The infrastructure required to support AI is absorbing unprecedented amounts of capital and placing further pressure on supply chains. While the industry continues to discuss AI primarily through the lens of software, the biggest obstacles to future growth sit much closer to the physical infrastructure that makes AI possible.
The sums being committed to AI infrastructure are already staggering. Last year, Microsoft announced plans to spend approximately $80 billion on AI-enabled data centers. Meta increased its capital expenditure forecast to between $64 billion and $72 billion, driven largely by AI infrastructure investments. Collectively, Alphabet, Amazon, Meta, and Microsoft are projected to spend hundreds of billions of dollars on infrastructure in the coming years.
Access to compute is becoming more expensive
For many years, cloud computing created the perception that organizations could access as much computing power as they needed, when and where they needed. Capacity could be provisioned quickly and expanded alongside demand.
However, training and serving large generative models has drastically altered the supply-demand equation.
Any organization hoping to develop or deploy advanced AI systems depends on infrastructure that requires substantial investment long before commercial returns can be realized. In many cases, access to land, power, water and computing capacity is becoming just as important as advances in the models themselves.
AI infrastructure requires a different financing model
Building a modern AI data center requires years of planning and construction, along with billions of dollars in investment before a single workload is deployed.
To date, this has relied on an arcane, convenient, lenient and fully legal structure called off-balance-sheet debt financing, with some bells and whistles. In practice, it allows the corresponding end clients such as Oracle, Meta, OpenAI and Anthropic to avoid reporting those massive debts on their balance sheets, or even disclosing much about them.
However, there is one catch. Those perks disappear the day the data center goes live, as the debt then must return to the balance sheet.
Meanwhile, many of these financing structures have found their way into private funds offered through wealth and institutional channels, and away from the original lenders.
Fintech innovation will shape the next phase of AI competition
Compute infrastructure has become more than a competitive advantage. It is reaching the stage of a critical commodity. While capital has so far scaled alongside the ever-growing demand for this infrastructure – and while there is still plenty of capital available to keep doing so – the conduits through which that capital is deployed need to become more standardized and sophisticated to ensure the trifecta of scalability, efficiency, and resilience is met.
Commercial real estate offers a useful comparison. Every year, large debt financings for office buildings, hotels, warehouses, and other commercial assets are pooled into standardized investment products, with different levels of risk and protection packaged for different investors. AI infrastructure needs an equivalent.
Just as commercial mortgage-backed securities (CMBS) and the CMBX market standardized capital conduits for commercial real estate, AI infrastructure will require its own equivalent – a market for standardized, tradable infrastructure-backed securities. This “DMBS” and “DMBX” are how the next many trillions of dollars of AI infrastructure will succeed.
