A single AI data center campus now draws power comparable to a mid-size American city, one visible symptom of the AI data center capacity bottlenecks now driving 2030 planning decisions. Buffalo, New York peaks at roughly 500 MW; several AI facilities in development are approaching or exceeding 1 GW of demand.
Quick take
Power availability and grid interconnection timelines, not chip supply alone, now set the pace at which AI data centers can be delivered.
NEMA’s Grid Reliability Study projects U.S. data center electricity consumption growing 300% over the next decade and reaching 38% of net U.S. electricity demand growth through 2037.
Four supply chain choke points, including foundry capacity, HBM memory, power, and optical interconnect components, have to scale together, because the tightest one gates growth in the rest.
Why AI Data Center Capacity Bottlenecks Are No Longer About Chips Alone
NEMA, ASHRAE, and Pacific Northwest National Laboratory responded by launching the AI Data Center Energy Performance Framework on June 10, covering planning, siting, liquid cooling, microgrids, and grid-interactive design for facilities that behave more like municipal loads than buildings.
The trigger was practical, not theoretical. Electrical equipment was already being installed in configurations that existing building codes never anticipated, according to NEMA senior vice president Patrick Hughes, who says standards development moves too slowly to match current construction pace.
The scale is intentional, not speculative excess. Thompson Research Group’s Kathryn Thompson compares the buildout to the U.S. interstate highway system in the late 1950s, arguing that early-stage capital intensity does not by itself indicate a bubble.
Semiconductor industry analysis frames the buildout as four linked choke points: advanced-node foundry and packaging capacity, tight HBM memory supply, power availability in aging grids, and optical interconnect components needed to move data between racks.
Consultant Geoff Tate’s framing is blunt: these constraints interact because the worst one gates demand for all the others, and a data center cannot run on three out of four bottleneck technologies.
AI Data Center Capacity Bottlenecks by 2030: Four Constraints That Move Together
Memory economics show why this matters for anyone underwriting a project. One Blackwell-class GPU costs roughly $6,400 to build, with HBM accounting for 45% of that cost, and sells for $30,000–$40,000; that margin is captured almost entirely at the design layer rather than by the operators renting out racks.
Capital committed to this stack is accelerating faster than the physical constraints are easing. Six buyers are on track to deploy over $760B of capex in 2026, up from $410B in 2025, while DRAM contract prices have risen ~820% year-over-year into 2026 on HBM allocation pressure.
That capital is not uniformly funded. CoreWeave’s capex-to-revenue ratio exceeds 250% and is financed substantially with debt, a materially different risk profile than hyperscaler capex funded out of 37–46% operating margins on an already-profitable core business.
On the operating side, data center operators are prioritizing token-per-watt performance over raw throughput, according to Endura Technologies chief product officer Davood Yazdani, because legacy power management approaches cannot absorb current AI power density.
Grid Interconnection and Site Selection Now Decide the Delivery Timeline
Capital tied to digital infrastructure could reach as high as $10 trillion by 2030, but Aon’s risk analysis is explicit that delivery, not fundraising, is the binding constraint in most markets.
Aon projects global data center power demand growing 165% by 2030, with data centers reaching roughly ~11% of total U.S. power demand by the same year, a share large enough that grid planners now treat hyperscale campuses as their own load class.
Grid access and interconnection timelines, often measured in years, increasingly determine which sites are viable at all. Aon’s Mark Potter frames the risk directly: developers pursuing behind-the-meter power still need to know when grid access becomes available within their project timeline, because backup capacity is not a permanent substitute.
Utilities are recalibrating around this demand in real time. NextEra’s Florida subsidiary FPL raised its data center capacity forecast to 8GW by 2032, a concrete signal that regulated utilities now treat AI load growth as a planning baseline rather than a one-off spike.
The four constraint categories rarely move at the same speed, which is why a single capacity number is a poor decision input. The comparison below lines up what the current evidence actually supports for each one.
| Constraint | Key signal from current evidence |
|---|---|
| Power & grid interconnection | Data center electricity consumption projected to grow 300% over the next decade, reaching 38% of net U.S. demand growth through 2037 |
| Chip / HBM memory supply | DRAM contract prices up ~820% year-over-year into 2026 on HBM allocation to AI accelerators |
| Facility scale & cooling design | AI facilities now approach or exceed 1 GW of demand, versus roughly 500 MW peak demand for a city like Buffalo |
| Workforce & permitting | Power, water, and workforce risk are treated as design decisions shaping bankability and insurability, not downstream operational issues |
What to Verify Before Committing Capital to AI Data Center Capacity
Choose to commit early when a site already has confirmed grid interconnection capacity, an executed power agreement, or a utility-backed load forecast like FPL’s; the evidence suggests power deliverability, not construction, is what actually gates timelines in those cases.
Avoid signing on projected capacity alone when interconnection status is undocumented or when a site depends on behind-the-meter generation without a confirmed grid fallback date, since that gap is exactly what Aon flags as an unresolved operational risk.
What the evidence establishes: power and grid interconnection lead times, chip and HBM allocation, and workforce availability are now treated as day-one design decisions rather than downstream operational problems. What the available documentation does not establish is how quickly the new NEMA framework changes actual permitting or code-approval timelines, since it launched in June 2026 with no adoption track record yet.
Workforce and supply-chain sequencing compound the power question rather than sitting apart from it. Turner Construction describes success now depending on delivering power, cooling, workforce, and supply-chain capacity together, not sequentially, on a single Southeast campus project.
Before committing capital, verify the confirmed interconnection date on the utility’s own queue, not a developer’s projected timeline, and price the cost of a behind-the-meter fallback if that date slips. If either number is unavailable, the capacity commitment is not ready to underwrite.