Insights · Infrastructure
The Next AI Bottleneck Is Geography
AI model builders plan in weeks. Utilities plan in years. Valley Recap's essay argues that this mismatch has turned location into the binding constraint on AI infrastructure, and that the builders who win will be the ones who plan power, compute, and connectivity together.
Two industries on different clocks
The essay opens with the timing problem. More than 2,000 GW of generation and storage was sitting in U.S. interconnection queues at the end of 2025. Large power transformers take roughly 128 weeks to deliver, and their prices are up 77% since 2019. Utilities order substation equipment as much as seven years out.
Demand is moving on the opposite timescale. Infrastructure Masons predicted in 2024 that the data center industry would double in five years. According to the essay, it doubled in two and has tripled in three.
“We're talking about two different industries that are not on the same time frame.”
Dr. Lamia Youseff, Jazz Computing, speaking at the NeoCloud Summit's “Powering the NeoCloud” panel, as quoted in Valley Recap
The demand picture, in numbers
| Metric | Figure | As cited from |
|---|---|---|
| Generation and storage in U.S. interconnection queues, end of 2025 | >2,000 GW | Essay |
| Large power transformer lead time | ~128 weeks | Industrial Sage |
| Large power transformer price increase since 2019 | +77% | Industrial Sage |
| Global data center electricity use, 2024 to 2030 | 415 TWh to ~945 TWh | IEA |
| Data centers' share of U.S. electricity, 2023 to 2028 | 4.4% to 6.7–12% | U.S. DOE |
| U.S. projects blocked or delayed by local opposition, Q2 2025 | $98B | Datacenter Watch |
| Middle East data center capacity, 2025 to 2030 | ~1 GW to 3.3 GW | PwC |
| Latin America colocation inventory growth, one year | +20%, 42% of pipeline pre-leased | JLL |
| Capital committed to sovereign AI compute in 2026 | >$100B | Deloitte |
| Sovereign cloud IaaS spending this year | $80B, up 36% | Gartner |
The 2030 global figure of roughly 945 TWh is about equal to Japan's current total electricity consumption. The Datacenter Watch number for a single quarter exceeds every earlier quarter since 2023 combined.
Builders are moving, and so is the map
Faced with these constraints, the essay describes three responses. Some developers now secure generation before they buy land. Others go behind the meter entirely. A growing number are looking outside the U.S., to Latin America, the Nordics, and the Middle East, where power and permits can be lined up faster.
Buyers have adjusted too. Most AI companies now consider capacity outside the U.S. The questions that matter, according to the essay, are whether 25 MW is available and whether it can scale. Speed beats location.
Sovereignty and the network
Two further forces push in the same direction. Sovereign AI is becoming a procurement requirement: Deloitte expects more than $100 billion committed to sovereign compute in 2026, and notes that roughly 90% of AI compute today is run by U.S. and Chinese companies, a share it expects to erode as other nations build their own.
Meanwhile, inference is pulling compute toward users. The internet was built around population centers, which are often far from power. Physical AI such as robotics needs latency near 10 ms, while much of today's internet runs at 50 ms or more. The essay cites Spectrum, which has enabled edge compute at more than 1,000 network sites, placing GPUs within 10 ms of 500 million connected devices.
The essay's outlook
Over the next 24 months, the author expects a harder push for sovereign AI, utilities rethinking how they plan, rapid growth outside the U.S., and a new wave of internet exchange points and edge sites. The conclusion is that the winners will be those who plan power, compute, and connectivity as one system rather than three.
Our view
Why this matters at the intersection of compute and energy
This essay describes the problem AICE Capital Partners exists to solve. Three points stand out.
- The clock mismatch is the whole opportunity. A seven-year utility planning cycle against a weeks-long AI build cycle is not a gap anyone closes by moving faster inside one industry. It closes when the same team controls generation, site, and cluster and can sequence them together.
- "Secure power first, then land" is now the right order. The essay notes developers are already doing this. We think it should be the default for every project, because a powered site is financeable and an unpowered one is a liability with a coupon attached.
- Geography is a design variable, not a given. Behind-the-meter generation, international sites, and edge placement near users are all ways of choosing where the constraint binds. Integrated capital that understands both energy and compute can underwrite options that single-layer players cannot.
Power, compute, and connectivity planned together is exactly the stack we work across. Talk to us if you are building, powering, or funding AI capacity.