Former Intel (INTC, Financials) Chief Executive Officer Pat Gelsinger said Nvidia's (NVDA, Financials) pricing strategy for artificial intelligence chips is excessively high and not viable for AI inference at scale.
Speaking on the Acquired podcast at Nvidia's 2025 GPU Technology Conference, Gelsinger said inference—a vital step in deploying AI models—is where the industry is headed and Nvidia's present technology is ill prepared to service that demand cost-effectively.
Gelsinger said that the processors Nvidia uses for artificial intelligence training cost 10,000 times more than what is realistically needed for inference. Although he admitted that most of the early surge in generative artificial intelligence was driven by Nvidia's graphics processing units, he contended that the company's strong point—powered by CUDA software platform—may not persist once inference takes front stage.
Gelsinger pointed out flaws, but he praised Nvidia CEO Jensen Huang for his ten-year vision and tenacity. He remarked that while Huang's early prediction on general-purpose graphics processors and AI workloads paid out, success also resulted from good timing. "Jensen got lucky," Gelsinger stated in the interview.
Under Gelsinger's direction, Intel battled to stand out in the AI hardware space. The Gaudi accelerator chips of the firm fell behind the Hopper and Advanced Micro Devices' (AMD, Financials) Instinct products of Nvidia. Intel has since shelved its Falcon Shores artificial intelligence platform and is currently concentrating on a next-generation project under Jaguar Shores.
Gelsinger also noted a possible change in computer architecture as quantum computing may become commercially feasible by the end of the decade. He left out any indication of Intel's possible posture in that change-over.
Reflecting Intel's larger difficulties leveraging the explosion in demand for machine learning infrastructure, Intel's AI revenue still lags significantly behind those of its rivals.
