Key Takeaways:
Five chip and data center executives publicly rebutted fears of an AI computing surplus, arguing energy supply — not demand — is the binding constraint.
The selloff in semiconductor stocks this month has been driven by concerns that AI infrastructure spending is peaking. Meta Platforms Inc. said it would sell idle AI computing capacity, and Samsung Electronics Co. warned on profit despite a 360% gain over 12 months. But executives who build and supply the infrastructure told CNBC the opposite is true.
"AI demand is almost unlimited," Pat Gelsinger, former Intel Corp. chief executive and now general partner at Playground Global, said. "The real bottleneck is energy — not compute."
Marc Boroditsky, chief revenue officer at Nebius, a cloud platform building data centers with Nvidia Corp. graphics processing units, said demand "far exceeds our ability to fulfill it, and that has been the case for some time." Andrew Feldman, chief executive of Cerebras Systems Inc., a maker of AI chips that went public this year, dismissed the Meta and xAI capacity sales as isolated cases. "Across the industry, demand for compute far exceeds available supply," he said. "We have a shortage of data centers and a shortage of many of the inputs needed to build compute."
The most striking signal came from Lumentum Holdings Inc., a supplier of photonic and optical interconnect products. Chief Executive Michael Hurlston said the company's order book is filled for the next five years. "We are doing everything we can to expand capacity to meet what we now see as five years of demand," he said. Lumentum shares have risen about 600% over the past 12 months.
The Energy Constraint
The International Energy Agency projects global data center electricity demand will more than double to 945 terawatt-hours by 2030, up from about 415 TWh in 2024. AI is the primary driver, with data center power consumption growing at roughly 15% annually — more than four times the pace of other sectors. McKinsey & Co. estimates global data centers will require nearly $7 trillion in capital investment by 2030, with about $1.3 trillion earmarked for power generation, transformers, substations and transmission infrastructure.
Gelsinger's framing aligns with a growing consensus among industry planners: the next phase of the AI race will be won not by chip architecture alone but by the power grid that sustains it. Countries including the U.S., China and Gulf states are racing to secure clean, reliable electricity for AI clusters, with India's Ministry of Power estimating data center electricity demand could reach 13.56 gigawatts by 2032.
From Token-Maxxing to Value-Maxxing
A separate concern has been whether enterprise AI spending can sustain its pace. Many companies have moved through a phase Boroditsky called "token-maxxing" — encouraging employees to use AI tools as much as possible, often on frontier models from OpenAI and Anthropic. As chief financial officers scrutinize costs, that approach is giving way to what he termed "value-maxxing."
"CFOs tightening spending are actually looking for value," Boroditsky said. "We are seeing a more rational shift. Every technology cycle goes through this, and this rationalization will actually sustain demand."
Cerebras's Feldman offered a complementary view: as enterprises mature in their AI deployment, different workloads will migrate to different tiers of compute. "You don't take a bus to buy groceries," he said. "Some workloads will move to one class of compute, simpler workloads to another." That suggests frontier models and open-source alternatives from companies like DeepSeek and Alibaba Group Holding Ltd. will coexist rather than cannibalize each other, keeping aggregate compute demand on an upward trajectory.
Sungyun Park, chief executive of South Korean chip startup Rebellions Inc., which counts Samsung and SK Hynix Inc. among its backers, said the momentum behind AI infrastructure remains strong. He does not see the Meta and xAI moves as evidence of overinvestment by hyperscale cloud providers.
For investors, the debate carries direct implications. Nvidia trades at about 22 times forward earnings after a 13.6% decline from its 52-week high, while Advanced Micro Devices Inc. has drawn analyst price targets as high as $615 on expectations its MI450 series and Helios platform will begin ramping in the third quarter. Marvell Technology Inc., whose optical connectivity products face extended lead times beyond six months, has analysts projecting more than 40% growth over the next three years.
The risk is not that AI demand fades, the executives argued. It is that the industry cannot build power infrastructure fast enough to keep up.
This article is for informational purposes only and does not constitute investment advice.
