AI executives say demand remains strong as customers scrutinize costs

AI infrastructure executives told CNBC this week that demand for computing capacity remains ahead of available supply, despite recent volatility in chip and data center-related stocks. Their comments come as investors reassess the pace of artificial intelligence spending after a year-long rally in semiconductor shares and signs that some large technology groups are selling spare compute capacity.

Pat Gelsinger, the former Intel chief executive who is now a general partner at Playground Global, told CNBC on Wednesday that he views AI demand as “almost unlimited,” with access to energy the main constraint. He said the economic value of adding intelligence across industries could be substantial, which in his view supports continued demand for AI systems.

The debate has intensified after Meta said it would sell excess AI computing capacity, a move that lifted its shares but raised questions about whether the sector had built too much infrastructure. CNBC also reported that Elon Musk’s xAI rented out spare capacity this year. Executives interviewed by CNBC said those examples did not point to broad overcapacity across the industry.

Supply constraints remain central to the AI buildout

AI compute depends on several scarce inputs: graphics processing units, data center space, power, networking equipment and optical components that move data between servers. When one part of that chain is constrained, companies may struggle to bring new capacity online even if customer demand is available.

Marc Boroditsky, chief revenue officer at Nebius, told CNBC on Thursday that the company is seeing more demand than it can meet. Nebius is building data centers that use Nvidia GPUs, the chips widely used to train and run AI models.

Andrew Feldman, chief executive of Cerebras Systems, told CNBC on Wednesday that Meta and xAI were unusual cases. He said demand for compute across the industry exceeds available capacity and that data centers and other inputs remain in short supply. Cerebras, which went public earlier this year, is among the chip startups seeking to compete in the data center market.

Sungyun Park, chief executive of South Korean AI chip startup Rebellions, also told CNBC on Wednesday that AI infrastructure momentum remains strong. Rebellions is backed by Samsung and SK Hynix.

Lumentum, which sells photonics and optical products used in data center connectivity, told CNBC that its products are sold out for the next five years. Chief Executive Michael Hurlston said Wednesday that the company is trying to expand capacity to meet demand it can see over that period. Lumentum shares are up about 600% over the past 12 months, according to CNBC.

Enterprise buyers focus more on returns

Executives also said corporate customers are taking a closer look at what they spend on AI. Some companies had encouraged employees to use AI tools extensively, a practice CNBC described as “tokenmaxxing,” often through frontier models from companies such as OpenAI and Anthropic.

Those models remain expensive compared with some open-source alternatives from companies including DeepSeek and Alibaba, according to CNBC. As a result, businesses are assessing which tools are suitable for specific tasks and whether the output justifies the cost.

Boroditsky said finance chiefs should look for value rather than usage alone, and that AI spending should be tied to returns. He described the current shift as a move toward more rational spending, while saying demand can continue as customers become more selective.

Feldman said different AI models are likely to be used for different workloads. More advanced models may be reserved for complex tasks, while less demanding work can move to other systems. That allocation could shape demand across chips, data centers and software as enterprises refine how they use AI.

This story draws on original reporting from CNBC.