Pat Gelsinger told CNBC that he thinks artificial-intelligence demand is "almost unlimited" and energy availability is "the only real limiter." Marc Boroditsky of Nebius said demand exceeded what his company could fulfill, Cerebras chief Andrew Feldman said industry compute demand far outstripped capacity, and Rebellions chief Sungyun Park described infrastructure momentum as huge. These are attributable supplier claims, not metered utilization. [1]
They sharpen the paper's July 10 account of Meta's named generation bill without measured load. That project attached a proposed 1.8-gigawatt campus to a 970-megawatt gas plant, making power scale visible while leaving contracts, construction, water and operating demand unresolved. Sunday's testimony says the constraint is energy; it still does not supply a meter. [1]
The speakers have direct interests in the buildout. Nebius constructs data centers with Nvidia processors, Cerebras sells competing AI systems and Rebellions is a chip startup. Their view can be informed and newsworthy while remaining testimony from suppliers and investors. A request for capacity may be reserved, contracted, cancellable, delivered or paid; each stage says something different about the quality of demand. [1]
CNBC also reports the apparent counterexample. Meta plans to sell computing power it is not using, and xAI has rented excess capacity. Meta's cloud move followed the possibility that it had overbuilt infrastructure, although the company can also treat idle capacity as a new product. Spare machines at one company, in one region and at one moment do not establish broad overcapacity, just as a supplier queue does not prove every installed machine is full. [1][2]
The reconciliation requires dimensions that the rhetoric leaves out: accelerator type, location, workload, time of day, contract length, price and available power. Demand for frontier training on a particular cluster can exceed supply while older chips or capacity elsewhere sit idle. Energy can bind at a named site while networking, memory, cooling, permits, software or customers constrain another. "Shortage" and "excess" do not contradict until they describe the same thing.
A common utilization table would make the disagreement testable. It would separate installed from energized servers, reserved from used hours and requested from paid capacity, then attach each figure to a geography and price. The interviews contain none of that common denominator.
Market prices add heat but not measurement. CNBC notes volatile chip and data-center shares during the debate. A share rally or selloff records investors changing their expectations; it does not show server utilization, customer return or an energized megawatt. The same discipline applies to corporate spending: enterprises scrutinizing cost may shift workloads or models without eliminating demand for useful systems. [1]
This article owns executive demand claims, power and utilization. A companion owns the five-year product queue, and another owns enterprise cost and return. Backlog can reveal supplier visibility without proving shipment. Customer anecdotes can show scrutiny without becoming measured return. Combining all three would make a single bullish or bearish story from receipts that answer different questions.
No qualifying X status survived the recorded Gelsinger, executive and company searches. That does not mean investors or AI executives had no online view. It prevents a fabricated post from resolving a disagreement that the published interviews themselves leave open.
The next decisive evidence is mundane: utilization by chip and region, energized capacity, customer and contract terms, prices, cancellations, revenue and power delivery. Until those numbers arrive, "almost unlimited" is executive testimony and rented excess is a counterexample. Neither is a system-wide demand meter.
-- DAVID CHEN, Beijing