Kevin O'Leary Is Right on AI Power: 100-GW Gap Means Data Centers Must Earn the Grid-and the Community

AI's bottleneck is shifting from chips to power delivery

The investment lesson is straightforward: in AI, the plug is becoming the product. Model wins and chip speed still matter, but Bank of America sees a U.S. supply gap of more than 100 GW over the next five years. That is the real constraint. Investors who focus only on frontier-model hype risk missing the harder question: whether a project can connect to the grid quickly enough to turn compute into revenue.

Why the bottleneck matters now

The numbers are the first clue. Data centers could add roughly 125 GW of U.S. electric load over the period, while utilities are expected to add only about 93 GW of accredited supply. PJM also expects about 70 gigawatts by 2038 of new large data-center load, and its latest auction showed power prices near record highs at about $325 per megawatt-day. Just as important, PJM says it fell short of its reliability requirement for a second consecutive auction.

The long-term AI demand case still looks strong, and utilities have been revising demand forecasts upward in each of the past three years. The risk is that the demand is real but the delivery system is not keeping pace. The key tension is no longer whether anyone wants AI compute; it is whether electricity can be delivered where and when it is needed.

AI racks are straining local grids, not just project budgets

The problem is no longer finding capital for an AI campus. It is that the local electrical neighborhood was built for a different era. AI racks now pull 30 kW to over 100 kW, compared with 5 kW to 15 kW for traditional racks. That change pushes older sites far beyond a simple efficiency upgrade. It can require new switchgear, larger transformers, stronger distribution lines, and sometimes new substations.

The pipeline looks large, but execution is the bottleneck

That is why 30% to 50% of U.S. AI data centers planned for 2026 are now cancelled or delayed. Of the roughly 16 GW of U.S. data center capacity announced for 2026, only around 5 GW is currently under construction. In plain English, the pipeline looks large, but the share actually breaking ground is much smaller.

Why first-time developers are more exposed

This is where execution matters most. Getting a data center online now requires coordination across financing, land control, power delivery, and offtake agreements. BloombergNEF notes that the majority of new projects are sponsored by first-time developers, which increases execution risk because these teams may have capital and ambition but not yet the experience needed to navigate utility queues, equipment lead times, and local politics. BNEF's own Mark Daly described the U.S. as a really tough environment to build data centers right now, citing intense competition for labor and power equipment.

The practical filter is simple: projects that can prove they understand the local power system, fund the necessary upgrades, and move through community approval will advance. For investors, the clearest signal is not another announcement but visible construction progress.

Who benefits when grid access becomes the scarce asset?

The investable point is straightforward: when the grid slows projects, capital should favor the players that can secure power first and finish construction next. That means shifting attention from headline-worthy announcements to who has a credible path to delivery.

Who is better positioned

The likely winners are not necessarily the best-known AI software names. They are the projects and businesses that can lock power, keep the project bankable, and move through permitting and grid interconnection faster than peers.

Who is more exposed

The more vulnerable stories are the ones still waiting for ideal grid timing. BNEF expects U.S. data centers' power capacity could reach 194 gigawatts by 2035, which would mean data centers account for about 20% of electricity consumption by the middle of the next decade. That demand is substantial, but interest alone does not build capacity. The same 30% to 50% of U.S. AI data centers planned for 2026 that are delayed or cancelled show how much planned demand can stall before it becomes operating supply.

The stress test

Keep the test simple. A rise in electricity prices or construction costs can affect economics, but it does not automatically break the long-term demand story. The more important test is whether a project can show direct power access, utility coordination, and real construction progress. That is a stronger indicator than a compelling slide deck.

What would validate the grid-bottleneck thesis?

Signals that support it

The clearest validation would be continued grid strain and regulatory action. PJM says it fell short of its reliability requirement for a second straight auction and is moving ahead with backstop power procurement, with results expected in early December. If those proposals are approved, that would be a practical signal that system operators see power adequacy as a live issue, not just a future narrative.

Signals that weaken it

The main risk to this thesis is that delays keep pushing project timelines outward without forcing a clean repricing of announced winners. Developers are already facing grid interconnection queues, transformer shortages, permitting gridlock, and increasing community opposition. BNEF also says data centers are unpopular across the political spectrum, and some states could pursue additional limits. In that world, demand may stay high while the actual buildout lags expectations for longer than bulls anticipate.

What to watch next

• Grid operators: whether PJM's backstop procurement plan is approved and whether other systems show similar strain.
• Build rates: whether the share of announced capacity under construction rises meaningfully.
• Developer mix: whether experienced sponsors continue to outexecute first-time developers.
• Local politics: whether community opposition and moratorium discussions spread or ease.