Photo: Tima Miroshnichenko / Pexels
Infrastructure bottlenecks in power, cooling, and transformers could leave a significant chunk of next year's AI capacity unpowered
Elon Musk posted a blunt forecast on X on August 29, 2026: roughly 15 gigawatts of AI computing capacity scheduled to come online in 2027 will likely sit dark, because the physical infrastructure needed to power it simply will not be ready in time.
The warning lands at a specific moment. The AI industry spent the past two years treating compute as the scarce resource, throwing billions at Nvidia and AMD chips, securing GPU allocations years in advance, and racing to announce flagship clusters. Musk is now arguing the real constraint is everything that comes after the chip: the transformers, the wiring, the liquid cooling loops, and the high-capacity networking fabric that turn a pile of silicon into a functioning data center.
Why the grid cannot keep up with the GPU #
The core problem is timing. High-voltage transformers, the large custom units that step utility power down to data center voltages, currently carry lead times of 48 to 60 months. That is four to five years from order to delivery, on hardware that every new AI facility needs before it can draw a single watt.
Permitting adds another layer of delay. Grid interconnection queues in North America are notoriously backlogged, and large industrial loads face scrutiny from utilities and regulators that can stretch timelines by additional years.
Analysts tracking US AI infrastructure have estimated that domestic data center power demand could reach 66 GW by 2027. If 15 GW of associated compute cannot be energized on schedule, the shortfall represents a real drag on deployment timelines across the industry.
SpaceX’s position inside the bottleneck #
Musk’s warning is not purely abstract. SpaceX, which went public earlier in 2026 and disclosed its compute ambitions during its debut earnings call in early August, is itself one of the companies most exposed to these constraints.
The company closed Q2 2026 with 1.4 GW of compute capacity, up from 1 GW in Q1 and from just 400 megawatts a year prior. SpaceX has set a target of exceeding 2 GW by year-end 2026, and it has outlined a goal of 20 GW in power-and-cooling infrastructure by the end of 2027. Musk himself framed the realistic expectation closer to 15 GW after accounting for the delays he is now publicly flagging.
The Colossus cluster in Memphis, operated under the xAI umbrella, is a central piece of that buildout, and the company is planning to deploy Nvidia’s next-generation Vera Rubin chips as they become available. SpaceX is also reportedly in discussions about significant offtake agreements with major cloud customers, including Microsoft.
What stranded compute means for the broader AI race #
For the companies that have already paid for chips and signed contracts expecting them to generate revenue in 2027, idle capacity is not a theoretical problem. Depreciation on AI hardware runs fast, so every month a cluster sits unpowered is a month of lost economic life on expensive equipment.
For investors, the most important near-term signal is not which company has the most chips on order. It is which company has locked in power purchase agreements, secured transformer deliveries, and cleared permitting hurdles far enough in advance to actually flip the switch when the hardware arrives.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our