Executives Refute Claims of Excess Computing Power: AI Demand Is 'Almost Infinite,' Energy Is the Bottleneck

Recent sharp volatility in chip stocks has sparked concerns about a slowdown in AI demand, but multiple industry executives have clearly stated that these worries are overstated.

Pat Gelsinger, former CEO of Intel and now a General Partner at Playground Global, told CNBC this week that AI demand is "almost infinite," and the true bottleneck constraining industry development is energy supply, not computing power demand itself. Executives from various data center and chip companies have also spoken out, stating that current market demand far exceeds supply capabilities, and claims of excess computing power do not align with their actual operational conditions.

At the same time, corporate AI spending patterns are quietly shifting—from previously encouraging employees to use AI tools extensively regardless of cost, to a "value maximization" strategy that focuses more on return on investment. This trend has raised external doubts about the sustainability of corporate AI spending, but several executives believe that this rational adjustment will precisely help sustain demand in the long run.

Behind Chip Stock Volatility: Where Do Concerns About Excess Come From?

The factors driving this round of chip stock volatility come from multiple directions. Meta announced it would sell its idle AI computing power; although Meta's stock price rose following the news, it triggered associations in the market about overall industry excess in computing power. Elon Musk's xAI also rented out its surplus computing power externally this year.

In addition, Samsung, one of the world's largest memory chip manufacturers, previewed a significant increase in profits this week, yet its stock price fell in response. After accumulating a gain of over 360% in the past 12 months, the market began to question whether its upside potential was already limited.

These market signals, combined, have led investors to doubt the sustainability of AI infrastructure investments, triggering a new round of turbulence in chip and data center-related stocks.

Supply Side: Demand Far Exceeds Capacity, Orders Booked for Five Years

However, executives from companies directly involved in building AI infrastructure have offered a completely different judgment.

"The demand we are experiencing is extraordinary. Demand far exceeds our fulfillment capacity, which has been the norm for us for some time," Marc Boroditsky, Chief Revenue Officer at Nebius, which is building data centers using NVIDIA GPUs, told CNBC.

Andrew Feldman, CEO of Cerebras Systems, characterized the cases of Meta and xAI selling idle computing power as "isolated instances," emphasizing: "For the entire industry, computing power demand far exceeds existing capacity. We have gaps in data centers, as well as in many input elements required for computing power." Cerebras completed its IPO earlier this year and is one of a batch of semiconductor startups attempting to challenge NVIDIA and enter the data center market.

Rebellions, a chip startup from South Korea backed by Samsung and SK Hynix, also reported strong demand. "The momentum in AI infrastructure remains strong," said Rebellions CEO Sungyun Park, who believes that the moves by Meta and xAI do not indicate overinvestment in infrastructure by hyperscale cloud providers.

The situation at Lumentum, a supplier of photonic and optical interconnect products, may be the most illustrative. The company's CEO, Michael Hurlston, revealed that its product orders are fully booked for the next five years. "We are doing everything possible to expand capacity to meet the five-year demand we currently see," he said. Lumentum's stock price has risen approximately 600% cumulatively over the past 12 months.

Demand Side: Shifting from "Maximizing Usage" to "Maximizing Value"

Although signals from the supply side are clear, corporate AI usage is undergoing a structural transformation, which is another source of market concern.

Previously, many companies were in a so-called "tokenmaxxing" (maximizing usage) phase, encouraging employees to use AI tools extensively regardless of cost, primarily adopting frontier models from OpenAI, Anthropic, and others. However, as the cost advantage of these frontier models relative to open-source solutions like DeepSeek and Alibaba comes under increasing scrutiny, corporate CFOs have begun to evaluate the returns on AI investments more strictly.

Nebius's Boroditsky calls this new trend "valuemaxxing" (maximizing value). "CFOs tightening spending are actually looking for value," he said. "We are seeing a more rational shift. Every technology cycle goes through this process, and this rationalization will actually continue to drive demand."

Cerebras's Feldman offered another interpretation from the perspective of layered model usage: As companies become more mature in AI deployment, tasks of varying complexity will be matched with different levels of models and computing power. "You don't need to drive a bus to go grocery shopping," he said. "Some workloads will migrate to certain types of computing power, while simpler workloads will migrate to others." This implies that frontier models and open-source models will form a complementary relationship rather than a simple substitution, meaning overall computing power demand will not shrink as a result.