Elon Musk posted on X on August 29, 2026 that roughly 15 gigawatts of AI computing capacity due online in 2027 will likely sit dark because power infrastructure will not be ready. Crypto Briefing reports the warning. The claim is not that the chips will fail to ship. It is that the gear around the chips—the grid iron, the copper, the cooling, the fabric—will not be in place when the servers are.
A dark gigawatt is a cluster that exists on a purchase order and may even exist on a raised floor, but that cannot be fed. Musk’s post puts that fate on roughly 15 gigawatts of AI computing capacity timed for 2027. Crypto Briefing is the outlet that carried the post into a reported account of what he blamed, what the transformer market looks like, and how the same bottleneck sits inside his own companies’ power-and-cooling targets.
The bottleneck is not the chip
Crypto Briefing reports he blamed transformers, wiring, liquid-cooling loops, and high-capacity networking—not chips. The list is a bill of materials for a modern AI hall. Transformers step high-voltage grid power down to something a hall can use. Wiring carries that power across a campus and into racks. Liquid-cooling loops pull heat off dense accelerators that air can no longer handle. High-capacity networking ties those accelerators into a training fabric. None of those four is a GPU. All four, on this telling, are the reason roughly 15 gigawatts of AI computing capacity due in 2027 will likely sit dark.
The insistence on not chips is the point of the post. The industry has spent years treating accelerator supply as the constraint that mattered. Musk, posting on X on August 29, 2026, is treating accelerator supply as the constraint that no longer is. If the chips arrive and the transformers, wiring, liquid-cooling loops, and high-capacity networking do not, the chips sit. That is what sit dark means for AI computing capacity: silicon without the power path, the cooling path, or the network path that makes a cluster a cluster.
Crypto Briefing’s account does not add other culprits. It does not name generation shortages or a fuel crisis. It names the long-lead electrical and mechanical kit, plus the North American process that attaches a site to the grid. The four items Musk blamed are physical. The queues and permits that Crypto Briefing pairs with them are procedural. Together they are power infrastructure that will not be ready for the 2027 wave.
Four to five years for a high-voltage transformer
High-voltage transformers now carry 48-to-60-month lead times. That range is four to five years. A transformer ordered in 2026 on a 48-month clock is not a 2027 asset. On a 60-month clock it is even later. The year Musk said roughly 15 gigawatts of AI computing capacity will likely sit dark is the year those lead times have already missed.
Lead time is not a metaphor. It is the calendar between a purchase order and delivery of a high-voltage transformer. 48-to-60-month lead times mean the transformer market has already told the 2027 buildout that it is late. Musk blamed transformers first in the list Crypto Briefing reports. The lead-time figure is why that blame has a date on it.
North American grid interconnection queues and permitting add further delay. Further, in that sentence, means delay on top of the 48-to-60-month transformer wait. Interconnection is the process of attaching a large new load to the grid. A data center that wants tens or hundreds of megawatts does not simply plug in. It enters a queue. Permitting is the separate civil and environmental clock. Crypto Briefing places both in North America, which is where the U.S. data center power demand estimate also sits.
A project can order transformers and still sit in an interconnection queue. It can clear a queue and still wait on permitting. It can have wiring, liquid-cooling loops, and high-capacity networking on site and still lack the high-voltage step-down that makes the campus live. The 48-to-60-month transformer calendar and the North American grid interconnection queues and permitting clocks are why power infrastructure will not be ready even if chips are.
Sixty-six gigawatts of demand, fifteen left dark
Analysts estimate U.S. data center power demand could reach 66 GW by 2027. That is the size of the market Musk’s warning is aimed at. Leaving 15 GW unpowered would slow industry deployment. The arithmetic is plain: roughly 15 gigawatts of AI computing capacity sitting dark is a large slice of a 66 GW U.S. data-center power figure for the same year. Crypto Briefing reports the estimate and the consequence together.
66 GW by 2027 is an estimate, attributed to analysts, not a measured load. It is the demand side of the same year in which Musk said roughly 15 gigawatts of AI computing capacity will likely sit dark. Leaving 15 GW unpowered is the industry-level version of that warning. Deployment slows not because the chips are missing but because power infrastructure—transformers, wiring, liquid-cooling loops, high-capacity networking, plus North American grid interconnection queues and permitting—cannot bring those gigawatts live.
A slowdown of that scale is a scheduling problem with a balance-sheet tail. Capacity due online in 2027 that cannot be powered is capacity that cannot train, cannot serve inference, and cannot earn. Analysts putting U.S. data center power demand at 66 GW by 2027 are describing a build that assumes power. Musk, on X, is describing a build that may not get it. Leaving 15 GW unpowered is how Crypto Briefing translates the post into an industry effect: slow industry deployment.
The two 15 figures in the report are easy to confuse and should not be. One is roughly 15 gigawatts of AI computing capacity due online in 2027 that will likely sit dark. The other, later in the same account, is Musk calling about 15 GW of SpaceX power-and-cooling more realistic than a 20 GW target. They are not the same object. They are the same unit, in the same year, making the same point: infrastructure, not silicon, sets the ceiling.
SpaceX, public, and a climb from 400 megawatts
The warning is not only about other people’s halls. SpaceX, public earlier in 2026, closed Q2 with 1.4 GW of compute—up from 1 GW in Q1 and 400 MW a year earlier. Those three prints are a ramp: 400 MW a year earlier, 1 GW in the first quarter, 1.4 GW at the close of the second. SpaceX aims to exceed 2 GW by year-end 2026.
Public earlier in 2026 is the corporate backdrop. Crypto Briefing is reporting SpaceX compute figures as public-company numbers, not as a private rumor. Q2 closed at 1.4 GW of compute. Q1 was 1 GW. A year earlier the same measure was 400 MW. The sequence is acceleration: from 400 MW to 1 GW to 1.4 GW, with a stated aim to exceed 2 GW by the end of 2026.
Compute in gigawatts is a power rating, not a chip count. It is the same unit Musk used for the roughly 15 gigawatts of AI computing capacity he said will likely sit dark in 2027, and the same unit analysts used for 66 GW of U.S. data center power demand. SpaceX’s ramp from 400 MW to 1.4 GW, with an aim to exceed 2 GW by year-end 2026, is a company-scale version of the industry’s race to energize silicon. It is also why a transformer and interconnection bottleneck is not an abstract grid story. It is on the same ledger as a public company’s compute target.
Exceeding 2 GW by year-end 2026 would be a further step up from the 1.4 GW that closed Q2, and a long way from the 400 MW print a year earlier. The aim is aggressive relative to those prints. The 48-to-60-month high-voltage transformer lead times and the North American grid interconnection queues and permitting delays are the reason an aggressive aim can still meet a dark floor. Chips can land on a SpaceX floor on a semiconductor calendar. Power infrastructure lands on a transformer and interconnection calendar. Musk is saying those calendars do not match.
Twenty gigawatts on the slide, fifteen as the realistic mark
SpaceX targets 20 GW of power-and-cooling by end-2027, with Musk calling about 15 GW more realistic. The target is power-and-cooling, not chips. That matches the post on X. He blamed transformers, wiring, liquid-cooling loops, and high-capacity networking—the kit that is power-and-cooling—and not the accelerators.
20 GW by end-2027 is the slide number. About 15 GW is what Musk called more realistic. Crypto Briefing reports both. A realistic about 15 GW of power-and-cooling against a 20 GW target is a markdown at the company level, delivered in the same year—end-2027—as the industry’s roughly 15 gigawatts of dark AI computing capacity. Infrastructure, not silicon, is what he is marking down.
Memphis Colossus under xAI is central. Nvidia Vera Rubin chips and possible Microsoft offtake are flagged. Those are the named pieces of the build that the power-and-cooling target is meant to serve. Colossus in Memphis is an xAI cluster; Crypto Briefing still places it at the center of the same power story that includes SpaceX’s public compute ramp. Nvidia Vera Rubin chips are the accelerators in the frame—the very chips Musk said are not the bottleneck. Possible Microsoft offtake is flagged as a demand outlet, not as a signed figure. The word is possible. The report does not convert it into a contract size.
Putting those names next to the 20 GW target and the about 15 GW realistic mark is how the outlet ties a public SpaceX power-and-cooling plan to an xAI hall, to a Nvidia generation, and to a possible Microsoft buyer. None of that changes the transformer calendar. High-voltage transformers still carry 48-to-60-month lead times. North American grid interconnection queues and permitting still add further delay. Memphis Colossus still needs wiring, liquid-cooling loops, and high-capacity networking if Vera Rubin chips are going to do more than occupy a dark floor.
A central cluster does not exempt itself from the queue. If Colossus is central to the xAI build, it is also central to the power infrastructure problem Musk described on August 29, 2026. Nvidia Vera Rubin chips can be the right generation and still sit dark if the transformers are on a 48-to-60-month clock. Possible Microsoft offtake can be flagged and still wait on megawatts that permitting has not released. The named demand does not ship the gear.
Depreciation on a dark cluster
Idle clusters burn fast depreciation on expensive GPUs. That is the financial punchline Crypto Briefing attaches to Musk’s August 29, 2026 post. A GPU that cannot be powered still ages. Expensive GPUs—including the Nvidia Vera Rubin chips flagged in the same report—are capital assets with short useful lives in a market that replaces generations quickly. Idle is the state of a cluster that sits dark. Fast depreciation is what that state does to the asset.
The industry picture and the company picture meet there. Analysts estimate U.S. data center power demand could reach 66 GW by 2027. Leaving 15 GW unpowered would slow industry deployment. Roughly 15 gigawatts of AI computing capacity due online in 2027 will likely sit dark because power infrastructure will not be ready. SpaceX, public earlier in 2026, can close Q2 at 1.4 GW of compute, aim to exceed 2 GW by year-end 2026, and target 20 GW of power-and-cooling by end-2027, while Musk calls about 15 GW more realistic. If the realistic power does not arrive, the clusters go idle. Idle clusters burn fast depreciation on expensive GPUs.
That is why the post is about grid gear stranding GPUs. The GPUs are not stranded by a foundry. They are stranded by transformers with 48-to-60-month lead times, by wiring and liquid-cooling loops and high-capacity networking that do not keep pace, and by North American grid interconnection queues and permitting that add further delay. Crypto Briefing reports Musk saying so on X on August 29, 2026, with Memphis Colossus under xAI central, Nvidia Vera Rubin chips and possible Microsoft offtake flagged, and a dark roughly 15 gigawatt slice hanging over both a 66 GW U.S. demand estimate and his own more realistic 15 GW mark.
The servers can be in Memphis. The chips can be Vera Rubin. The buyer can be a possible Microsoft offtake. The public company can be SpaceX, already at 1.4 GW of compute after 1 GW in Q1 and 400 MW a year earlier, aiming to exceed 2 GW by year-end 2026. None of that turns a dark hall into a live one. Power infrastructure has to be ready. Musk’s post, as Crypto Briefing reports it, says that for roughly 15 gigawatts of AI computing capacity due in 2027, it will not be.