METAL LAB
Add us on GoogleMETAL

Musk says 15GW of 2027 AI compute won't switch on even with power secured

The bigger bottleneck isn't power supply — it's transformers, wiring, cooling, and networking construction

이미지: METAL LAB 생성

Summary

  • Elon Musk shared a consensus estimate on X saying roughly 15 gigawatts of AI compute set to be produced by 2027 won't be operational that same year.
  • He pointed out that securing power alone isn't enough — transformers, wiring, liquid cooling, large chillers, and complex networking all still need to be built.
  • The comment comes as NVIDIA and OpenAI expand gigawatt-scale data centers in places like Ohio.

Musk: "15GW of compute won't turn on even with power secured"

Hugging Face official website

An arrow leaves a thickly filled circle representing secured power, but a broken-ring gate symbolizing wiring and cooling construction blocks the way. As a result, the arriving node is only a half-filled circle, showing that only half the compute can actually be powered up.An arrow leaves a thickly filled circle representing secured power, but a broken-ring gate symbolizing wiring and cooling construction blocks the way. As a result, the arriving node is only a half-filled circle, showing that only half the compute can actually be powered up.

Elon Musk posted a consensus estimate on his X account on August 29 (local time) suggesting that about 15 gigawatts (GW) of AI compute due to be produced by 2027 won't actually be up and running that year. He noted that this problem is harder than simply securing power, since transformers, wiring, liquid cooling systems, large chillers, and complex networking all need to be built from scratch as well. As the source for the estimate, he linked to a Grok conversation he shared.

Just how big is 15 gigawatts

The number alone doesn't mean much on its own, but comparing it to recent data center deals confirmed by METAL LAB shows the scale involved. The "Portsmouth Pike" site that NVIDIA and SB Energy are building in Portsmouth, Ohio, has an initial deployment capacity of 4.25 gigawatts, with OpenAI designated as the anchor tenant and room to expand by up to another 3.75 gigawatts. OpenAI's total NVIDIA compute commitments already add up to roughly 12 gigawatts through 2030, and expanding this site alone could push that figure to 16 gigawatts. The same site also includes plans for a $33 billion, 9.2-gigawatt natural gas power plant. But building the power plant is one project — running the wiring and cooling needed to actually deliver that electricity to the GPU racks is an entirely separate, equally massive undertaking. In other words, 15GW represents the combined scale of three or four projects of this size, and the bottleneck Musk is describing isn't the power plant itself — it's the last-mile construction work needed to actually get that power into the server room.

models mobile

Wiring and cooling are the next constraint on compute

Debate over AI infrastructure has mostly centered on GPU supply and securing generation capacity, but Musk's comment points to the next bottleneck in line. Transformers are notorious for taking years to deliver even after being ordered, and large chillers, liquid-cooling piping, and ultra-high-density networking gear are all similarly supply-constrained. Even a company that has bought up GPUs and locked in power contracts can end up with a finished data center that still can't switch its servers on if this physical construction work falls behind. Ultimately, gauging the 2027 AI compute market means looking not just at GPU shipment volumes or power contract sizes, but at how far along the transformer, wiring, and cooling construction actually is.

Comments

Related stories