The smart money isn't betting on AI models anymore — it's betting on the concrete and copper that keeps them running.
The Summary
- Barry Sternlicht, CEO of Starwood Capital Group, outlined his firm's AI investment strategy at the Milken Institute Global Conference, focusing heavily on data center infrastructure rather than AI companies themselves
- Real estate capital is flooding into the physical layer of the AI economy as power demands from training and inference operations reshape commercial property markets
- The shift signals that institutional investors see infrastructure as the safer, longer-duration play in the AI buildout than model development or application layers
The Signal
Starwood Capital Group manages over $115 billion in assets. When a firm that size pivots toward data centers, it's not chasing hype. It's following power consumption projections and lease agreements that stretch 10-15 years. Sternlicht's comments at Milken reflect what's become clear to anyone watching capital flows: the AI gold rush is creating more reliable returns for the people selling shovels than for the miners themselves.
The math is straightforward. A single large language model training run can consume as much electricity as 1,000 American homes use in a year. Inference at scale requires perpetual power. Data centers don't pivot to new business models every 18 months. They sit there, humming, generating steady cash flow while the AI companies using them burn through funding rounds and fight over market share.
"Real estate investors are realizing that owning the building is better than owning the tenant."
What makes this play interesting is the compression happening across the stack. Five years ago, tech infrastructure meant office space in San Francisco or Seattle. Now it means purpose-built facilities near power substations in secondary markets. The geography of the AI economy is being written by electrical grid capacity and cooling costs, not by where founders want to live. Starwood's strategy acknowledges this.
Key infrastructure advantages:
- Data centers have 10-15 year lease terms versus 3-5 years for traditional commercial real estate
- Power infrastructure creates natural moats that take years for competitors to replicate
- AI compute demand is inelastic in the short term — companies can't just decide to use less
The second-order effect: this capital shift is accelerating the buildout of physical AI infrastructure faster than the applications layer can absorb it. We're pre-building the Fourth Web's foundation while still figuring out what to put on top of it. That's not necessarily inefficient. It means when the applications that actually matter arrive, the rails are already laid.
The Implication
If you're building in AI, understand that your infrastructure partners are making longer-term bets than you are. The data center hosting your training runs will outlive your current business model. That's not pessimistic, it's structural. Plan accordingly. The companies that will still be standing in 2030 are the ones that understand infrastructure as a strategic asset, not a line item.
For institutional capital, the pattern is clear: own the layer that can't be disrupted away. Models will get cheaper. Applications will consolidate. But the buildings full of GPUs still need to be somewhere, and someone has to own them.