Musk Warns of AI Power Gap: 15 Gigawatts Left Unpowered by 2027

At a G20 tech ministers' meeting in Chapel Hill, North Carolina, Elon Musk put a specific number on a problem that has been circulating through markets for a week: roughly 15 gigawatts of AI computing capacity planned for 2027 may not actually come online that year. The cause isn't a shortage of chips or grid power, but infrastructure that physically can't keep pace with the rate at which Nvidia and its rivals are producing chips.

Musk's post on X, published on August 29, 2026 and then repeated and expanded on at the G20 tech meeting on September 1, triggered a wave of analysis among investors hunting for companies poised to benefit from the bottleneck. Musk, who attended the meeting alongside Mark Zuckerberg, delivered a shared message to G20 leaders: energy, not silicon, is now the main brake on AI development.

What Doesn't Add Up

Musk broke the problem down into numbers. AI chip production is growing at 40-50 percent a year, while available power capacity outside China is increasing by only 10-20 percent a year. With that kind of gap, the two curves inevitably diverge, and the surplus computing capacity simply sits idle because there's no power to run it.

The thing growing faster will eventually overwhelm the thing growing slower - Elon Musk

Crucially, the problem isn't power generation itself. Even once a power plant or solar farm is up and running, that electricity still has to reach a specific server hall through transformers, switchgear, medium-voltage wiring, liquid cooling systems, massive chillers and the complex networking that ties thousands of processors together. Each of these components now has its own production backlog, often longer than the time it takes to build the data center itself.

The Shortfall by the Numbers

US data center power demand is projected to rise from 31 gigawatts in 2025 to 66 gigawatts in 2027, more than doubling in two years. Global AI chip output is expected to reach 45-50 gigawatts of computing capacity, but only 28-40 gigawatts of that may actually be usable. North America, where most large data center projects are concentrated, is set to receive chips totaling 33-35 gigawatts, of which only 17-23 gigawatts can likely be deployed.

Morgan Stanley reached a similar conclusion in an independent analysis, putting the power gap at 38 gigawatts by 2028. The bank notes that historically only 50-60 percent of capacity announced for a given year actually reaches the grid on schedule, with the rest slipping into subsequent years due to permitting delays, grid interconnection queues and equipment shortages.

Who Benefits From the Bottleneck

The shift in attention from chipmakers to power infrastructure suppliers already has a specific stock market address. Analysts point to GE Vernova as the main beneficiary, the company holds a $176 billion order backlog, and orders tied specifically to data centers have topped $5 billion, more than double the 2025 figure. Also mentioned are Bloom Energy, Constellation Energy and Vistra, companies that supply both power generation and the on-site equipment to distribute it.

For Nvidia itself, the bottleneck doesn't mean falling demand for its chips, but a real risk that some of the units ordered will sit in warehouses for months or years, unable to be connected to power. That's changing the calculus for investors, who until now focused almost exclusively on chipmakers and now have to factor in the time and cost of building the entire power chain around them.

What It Means for Poland and Europe

Musk's warning fits into a broader, global problem that also affects Poland. National grid operator PSE (Polskie Sieci Elektroenergetyczne, Poland's transmission system operator) estimates that data centers could consume as much as a tenth of the country's electricity by 2040, and Poland and the Czech Republic have already signed a memorandum on a joint AI gigafactory, which will add further pressure on the transmission grid. If even the wealthiest markets, like the United States, can't keep up with building transformers and grid connections, European data center projects could face similar delays, but with a smaller scale of investment available to soften the problem.

For Polish companies planning investments in AI computing capacity, there's one practical takeaway: a chip supply contract is only half the project. The other half, the grid connection and power distribution equipment, can today take longer to secure than buying the chips themselves.