“The sort of Silicon Valley model is running into a real-world infrastructure constraint.” That line, from Kimmeridge Energy Management co-founder and managing partner Ben Dell, is the most important sentence in the AI story right now — and it has nothing to do with chips. The thesis, spoiled up front: the great AI buildout has stopped being a software problem and become a heavy-industry problem, and Kimmeridge estimates that up to half of the data centers currently proposed in the United States are at risk of delay or cancellation.

Half the pipeline may never get built

The headline figure, reported by Bloomberg, is blunt. The investment firm Kimmeridge Energy Management estimates that roughly 50% of proposed US data centers are at risk of delay or cancellation. The reasons are not exotic. They are the same reasons any large physical asset is hard to build: permitting bottlenecks, power-grid shortfalls, community resistance, and a plain shortage of the equipment that turns electricity into usable capacity.

Dell’s framing is worth sitting with. The technology industry is accustomed to scaling by writing more code and renting more cloud. Doubling capacity has historically meant provisioning, not pouring concrete. Data centers at the scale AI now demands break that habit. You cannot ship a substation over the internet, and you cannot fast-forward a multi-year interconnection queue with a better roadmap.

The bottleneck is copper, not silicon

What makes this a structural story rather than a temporary hiccup is the specific nature of the shortages. Beyond limited utility power, Bloomberg’s reporting points to constrained supplies of transformers, switchgear and batteries — the unglamorous middle layer of the electrical system that steps power up and down, protects circuits, and smooths delivery. These are long-lead-time industrial goods with their own supply chains, their own skilled-labor requirements, and their own multi-year backlogs.

Add power-grid shortfalls and permitting queues on top, and you get a compounding problem. Each constraint alone would slow a project. Stacked together, they raise the odds that a proposed site slips its timeline, gets repriced, or is quietly shelved. That is the mechanism behind the “up to half at risk” estimate: not one fatal obstacle, but several ordinary ones arriving at once.

The money is enormous — the grid doesn’t care

The tension in 2026 is the gap between the capital flowing into AI and the physical reality absorbing it. Consider the megamoney. Nvidia’s data-center revenue reached $85.7 billion in a single quarter. Anthropic signed a deal reported at roughly $45 billion with Nscale for 460 megawatts of capacity. Those numbers describe demand that is, for practical purposes, unlimited — capital chasing compute at a pace the sector has never seen.

None of it changes how long it takes to energize a transformer. The buildout the industry is counting on runs directly through the parts of the economy that move slowest: utilities, regulators, equipment manufacturers, and the communities that live next to proposed sites. When Silicon Valley’s growth model collides with that world, the model is what bends.

• Permitting: interconnection and approval queues stretch timelines that spreadsheets assume are short.
• Power: utility supply is limited, and grid upgrades are themselves multi-year projects.
• Hardware: transformers, switchgear and batteries are in short supply on top of the power constraint.
• People: community resistance and political backlash can stall or kill an otherwise fundable site.

What this means for AI in 2027

The slowdown threatens the very pipeline Big Tech needs to keep training and serving ever-larger models. When supply of physical capacity tightens against demand this intense, the predictable results are higher costs and longer timelines — exactly what Bloomberg’s reporting describes.

The strategic reframe is this: for the past two years the industry has treated GPUs as the scarce resource, the thing that gates progress. Kimmeridge’s estimate suggests the scarce resource is shifting. If up to half of proposed data centers really are at risk, then the binding constraint on AI in 2027 may not be chips at all. It may be electricity, and the copper-and-steel machinery required to deliver it. The companies that win the next phase may be the ones that treat power procurement, grid access and long-lead equipment as core strategy rather than as a facilities line item — because that, not the next chip, is where the ceiling now sits.

Written for Red Robot with AI assistance and human editing. Based on reporting by Bloomberg.