The next phase of AI investing runs through power, data centers, cooling, fiber and the grid
By Darshan Honale | June 2026
I went into the California LP Summit to talk about real assets. The more I prepared for the panel, the more obvious the connection became: AI is turning parts of the physical economy into technology infrastructure.
We still talk about AI mostly through models, applications and software companies. That is understandable because the software changes are visible. But every inference request eventually lands on something physical: a chip in a server, inside a data center, connected by fiber, cooled by equipment, drawing power through a grid that may already be constrained.
The software story is sitting on top of a real-assets story.
Power is becoming a strategic input
EPRI’s 2026 scenarios estimate that U.S. data centers could consume roughly 9% to 17% of U.S. electricity by 2030, up from about 4% to 5% today. The range is wide because the buildout is uncertain, but the direction is not.
For investors, that changes how I think about AI infrastructure. The bottleneck may not always be GPU availability. In some markets it may be interconnection, transmission, substations, land, cooling or simply the ability to get enough power on the timeline a developer needs.
Texas is particularly interesting because it combines large power markets, land, industrial expertise and growing data-center development. It also exposes the tradeoff directly: fast load growth creates opportunity, but it also creates grid, permitting and pricing questions.
Real assets are becoming growth assets
Real assets are often sold to investors as inflation protection, cash yield and portfolio diversification. Those characteristics still matter.
But some infrastructure now has a different role. It is enabling growth in compute-intensive industries.
A substation, transmission line, fiber route, powered shell, cooling system or gas-generation asset can sit directly in the critical path of AI deployment. That makes parts of infrastructure look less like a defensive allocation and more like a strategic capacity bet.
The distinction matters because growth expectations can change underwriting behavior. Investors can overpay for an asset simply because it has “AI” in the story. The infrastructure still has to work economically if demand ramps more slowly, customers renegotiate, hardware gets more efficient, or the site loses its competitive power advantage.
The stack is broader than the data center
I find it useful to think of the physical AI stack in layers.
At the bottom is energy: generation, fuel, storage and grid capacity. Then transmission and interconnection. Then the data-center site, building and cooling environment. Then compute, networking and fiber. Only after that do we get to the models and applications most people think about when they hear AI.
Each layer has different economics, regulation and duration. A software product can change in six months. A transmission project may take years. A data-center lease can lock in assumptions for a long time. A power purchase agreement has another risk profile entirely.
That timing mismatch is one reason the infrastructure side of AI deserves separate underwriting.
Related: Why the harness is not the moat in vertical AI.
Where the value may accrue
I am not convinced the biggest long-term return in AI will come from any single layer of this stack.
Some software companies will create enormous value. So will some model providers and chip companies. But when demand grows faster than the physical system can respond, the bottleneck itself becomes valuable.
That could be power in one market, land with interconnection rights in another, cooling technology in a third, or fiber connectivity in a fourth.
The investor job is to distinguish a durable constraint from a temporary shortage.
The diligence questions change
If an asset is being underwritten partly on AI demand, I would want to understand a few things clearly: who is the customer, how firm is the load commitment, who pays for interconnection and upgrades, what happens if the project is delayed, how much water or cooling capacity is required, how exposed the asset is to hardware-efficiency gains, and whether the economics still work without the most aggressive demand case.
Those are not software diligence questions. They are infrastructure questions.
That is why I think the AI conversation is expanding. We are moving from asking which model wins to asking what the entire system needs in order to scale.
For real-asset investors, that is not a side story. It may be one of the more important demand shifts of the decade.
Sources and notes
EPRI, Powering Intelligence 2026: https://powering-intelligence.epri.com/executive-summary.html – U.S. data-center electricity scenarios and AI-driven load growth.
U.S. Department of Energy, Clean Energy Resources to Meet Data Center Electricity Demand: https://www.energy.gov/oe/clean-energy-resources-meet-data-center-electricity-demand – Data centers as a significant source of near-term U.S. electricity-demand growth.