Power, fibre and memory squeeze

The surprise in artificial intelligence building now is that the bottleneck is no longer just chips. The new choke points are electricity to run the machines, glass fiber to connect them, and memory chips to keep enough data close to the processors. (bloomenergy.com) (corning.com) That is why Meta signed an expanded long-term deal with CoreWeave on April 9, 2026 for about $21 billion of artificial intelligence cloud capacity through December 2032. CoreWeave said the capacity will be spread across multiple locations, which tells you this is a land, power, and network buildout problem as much as a chip-buying problem. (coreweave.com) (finance.yahoo.com) A data center is basically a warehouse full of computers, and those computers are only useful if power arrives first. Bloom Energy’s 2026 data center power report says power availability has become a “defining boundary” on growth and says 73% of operators are embedding on-site power into long-term strategy. (bloomenergy.com 1) (bloomenergy.com 2) Bloom’s pitch is simple: if the electric grid takes too long, put generation next to the servers. The company said in January 2026 that one-third of data centers are expected to be fully off-grid by 2030, and it said Texas data center load is poised to more than double to 30% of total United States demand by 2028. (bloomenergy.com) (investor.bloomenergy.com) Once the power problem is handled, the next problem is moving data between all those machines fast enough. Corning, which makes optical fiber and the connectors around it, said on March 16, 2026 that it was launching multicore fiber, micro cable, and new connectors designed to let artificial intelligence networks pack more links into tighter spaces. (corning.com) Corning has been upgrading its growth targets because artificial intelligence clusters need far more internal cabling than older data centers. In March 2025, Corning raised the expected sales growth rate for its enterprise optical communications business to 30% from 25%, and in February 2026 it upgraded its broader Springboard plan to add $11 billion in incremental annualized sales by the end of 2028. (investor.corning.com 1) (investor.corning.com 2) Corning’s customer list shows how directly this cable demand ties to the biggest artificial intelligence spenders. Corning and Meta said in February 2026 they had a multiyear agreement worth up to $6 billion to accelerate United States data center buildout, and on March 31 they said construction had started on a cable manufacturing expansion in North Carolina to support that work. (investor.corning.com 1) (investor.corning.com 2) Then there is memory, which is the short-term workspace for artificial intelligence chips. Micron has been one of the clearest signs that this part of the stack is tight, with reports in early 2026 that its high-bandwidth memory output for calendar 2026 was already fully committed as cloud companies raced to secure supply. (finance.yahoo.com) (futurumgroup.com) High-bandwidth memory is memory stacked in layers right next to the main processor so data can move in and out faster, like keeping ingredients on the kitchen counter instead of in the basement. When Micron’s high-bandwidth memory sells out that early, it means buyers are not just short of graphics processors from Nvidia; they are short of the memory that lets those processors run large models at full speed. (finance.yahoo.com) (futurumgroup.com) Put those pieces together and the shape of the market changes. The winners are no longer only the companies etching the brains of artificial intelligence systems, but also the companies that can deliver watts on-site, fiber between buildings, and memory beside the chip before the next data center opens. (coreweave.com) (bloomenergy.com) (corning.com)