Power costs near data centers have surged 267% in five years as AI compute demand outstrips grid capacity. Memory supply can support just 15 GW of the 100 GW planned buildout.
The AI trade is shifting. For the last two years, the bet was on the chipmakers – the companies that could design and manufacture the processors that run large language models. That trade still works. Nvidia Corp. (NVDA) is up more than 400% from the start of 2023. But a growing number of investors and analysts say the next wave of winners will not be the companies making the chips. They will be the companies supplying the things the chips need to run: electricity and memory.
Data centers filled with Nvidia and Advanced Micro Devices Inc. (AMD) graphics processing units consume enormous amounts of power. In some regions, the cost of electricity near data center hubs has risen 267% over the last five years, according to regional grid data. The surge reflects a simple arithmetic problem: AI compute demand is doubling roughly every six to nine months, but new power generation takes years to bring online. Grid interconnection queues already stretch into the 2030s in parts of Virginia and California, where much of the nation's data center capacity sits.
Hyperscalers – Microsoft Corp. (MSFT), Alphabet Inc. (GOOGL), and Amazon.com Inc. (AMZN) – have begun signing power purchase agreements directly with nuclear, natural gas, and solar developers to guarantee supply for future facilities. Microsoft has backed a restart of the Three Mile Island nuclear plant. Amazon bought a 960-megawatt data center campus connected to a Pennsylvania nuclear plant. These deals signal that electricity, not just compute, is becoming a competitive advantage, said analysts at Goldman Sachs in a June note.
But power is only one chokepoint. The second is memory. Every AI system uses dynamic random-access memory to hold the data it is processing. Without enough DRAM, the chips starve. Nvidia CEO Jensen Huang called the memory bottleneck "severe" during the company's most recent earnings call. Elon Musk made the same point in SpaceX's first quarterly earnings report, saying "the limiting factor currently is memory."
The numbers back them up. Nearly 100 gigawatts of new data center capacity is scheduled to come online globally over the next four years, according to McKinsey & Co. estimates. But DRAM supply, even with new fabrication plants from Samsung Electronics Co. and SK Hynix Inc., can support roughly 15 gigawatts of that capacity over the next two years, according to supply chain data from TrendForce. The gap between planned buildout and available memory components is the widest industry participants say they have seen since the dot-com era.
That period offers a playbook of sorts. During the late-1990s tech boom, the rapid expansion of internet infrastructure, personal computers, and networking hardware created a shortage of metals and raw materials. Copper, tantalum, and germanium prices soared as mining and refining capacity struggled to keep pace. The companies that had built capacity before the shortage did especially well. Antofagasta plc, a copper miner, rose roughly 1,200% between December 1998 and early 2001, after it began production from a mine built during the investment phase of the earlier decade.
Macro-investing analyst Eric Fry, editor of The Speculator at InvestorPlace, drew the parallel in a recent note to subscribers. The same pattern is repeating, he wrote. The companies that solve AI's physical constraints – power and memory – may generate returns that rival the chipmakers themselves. "When supply runs short, the companies supplying the resources could emerge as the biggest winners," Fry wrote.
Investors have already started to price this in. Utilities that serve data center hubs, including Dominion Energy Inc. and Constellation Energy Corp., have outperformed the broader S&P 500 by roughly 30 percentage points over the last 12 months. Memory chip makers Samsung Electronics and SK Hynix have seen their stocks rise as customers sign long-term supply agreements at elevated prices. SK Hynix, which supplies high-bandwidth memory to Nvidia, reported record quarterly profit in July.
The AlphaScala stock analysis page for NVDA shows a score of 75 out of 100, with a Moderate label, as of the latest data. The stock traded at $217.48, unchanged on the session. Microsoft, which trades at $503.81 with an Alpha Score of 71, has been the most aggressive hyperscaler in signing dedicated power deals, according to SEC filings reviewed by industry analysts.
The two bottlenecks feed into a third that Fry identified but did not name in his public note: the supply of the physical land and construction capacity to build the data centers themselves. Real estate investment trusts that own industrial land near power substations, and construction firms that specialize in data center builds, have seen order books swell. Turner Construction, part of the Hochtief group, said in June that data center work now accounts for more than 30% of its North American pipeline.
Investors watching the AI trade today face a different set of questions than they did a year ago. The chip shortage that defined the first phase of the boom has largely eased. Nvidia's supply chain is now delivering GPUs on time, and lead times for server components have shortened. The bottleneck has moved downstream, to the infrastructure that turns those chips into working systems. The companies that own the power, the memory, and the land beneath the data centers are the ones that may define the next phase.
Drafted by a large language model from the source reporting linked above, then screened by automated publishing checks. It is not read by a journalist before publication. Some articles cite our Alpha Score. Verify prices and figures against the original source. Educational coverage, not personalized advice.