At Ai4, Pat Gelsinger says AI’s future hinges on boosting power efficiency, infrastructure, and economics – not simply deploying more GPUs.
LAS VEGAS — “Today’s GPUs suck.”
That was former Intel CEO Pat Gelsinger’s blunt assessment of today’s AI hardware. But his real focus wasn’t the GPUs – it was the infrastructure powering them.
“They’re very power inefficient,” Gelsinger said during his keynote at Ai4 2026 yesterday (August 4). “They’re computationally limited.”
Gelsinger’s comments framed a broader argument that AI’s next breakthrough will depend less on buying more accelerators and more on solving fundamental constraints in compute, memory, networking, energy, and economics.
“We need much greater efficiency in AI processing if it’s going to reach its potential,” he said.
AI’s Foundation Starts With Chips #
Moderated by The New Yorker staff writer Gideon Lewis-Kraus, the discussion at Ai4 focused on the infrastructure underpinning AI rather than the models themselves.
Asked what business leaders most often misunderstand about AI, Gelsinger argued that every breakthrough ultimately rests on semiconductors.
“There ain’t no compute without chips sitting underneath it,” he said.
He described chips as “the underlying fuel” of a token-driven economy and argued that demand for intelligence is effectively unlimited.
“For AI, I think the demand is unlimited for intelligence,” Gelsinger said. “We see this insatiable demand. Underneath that, you have an almost insatiable demand for compute, for infrastructure, for memory.”
The pace of AI development, he said, has only been possible because it builds on decades of investment.
“It didn’t have to build the internet. It didn’t have to build semiconductors. It didn’t have to build clouds. It’s building on all of those.”
Four Priorities for AI Infrastructure #
Gelsinger, who was ousted from Intel in 2025 after an ambitious comeback sputtered and the company fell further behind rivals, is now a general partner at Playground Global.
His work has centered on four priorities for AI: infrastructure, efficiency, energy, and economics.
“We need more semiconductors,” he said. “We need more data centers. More communications infrastructure.”
He also criticized high-bandwidth memory (HBM), describing it as an inefficient technology that remains the industry’s best available option for AI training and inference.
“HBM today is a very inefficient memory,” he told the Ai4 audience. “It’s just the best one we have for AI modeling today.”
Future gains, he argued, will require redesigning far more than the processor itself. Improvements must come across the entire compute stack, including packaging, memory, networking, and system architecture.
OpenAI Sees Constraints Across Every Layer #
Sharing the stage with Gelsinger was Sachin Katti, head of compute at OpenAI. Katti said the industry’s biggest challenge is no longer convincing people that AI is valuable, but convincing governments, utilities and manufacturers to build enough infrastructure to support it.
“The main real focus has been convincing the world that we need to invest in every layer of the infrastructure stack because of the potential this technology has in delivering intelligence to the whole world,” Katti said.
According to Katti, virtually every part of the supply chain has become a bottleneck.
“It’s not easy to build data centers quickly in the US. It’s not easy to find energy in the US that we need. There’s not enough fab and memory capacity to supply the chips we need, and it’s not easy to put it all together into usable compute.”
OpenAI’s objective, he said, is to compress deployment timelines dramatically.
“Our vision is: can we cut that down to three quarters instead of three years?”
Doing so will require faster chip manufacturing, standardized data center construction, expanded electricity infrastructure and more efficient deployment of compute resources across the industry.
Energy Is Becoming AI’s Limiting Resource #
Gelsinger agreed that one of the biggest challenges facing the AI industry is energy.
“In a digital AI economy, economic capacity equals energy capacity,” he said.
He argued that future AI growth depends as much on expanding electricity generation as advancing semiconductors.
“We have an energy issue,” he said. “We need more energy efficiency in getting power delivered to the AI computation. And we just need more energy to power it.”
That challenge extends directly to data center development.
“Nobody’s going to build a data center for chips that you can’t power up,” Gelsinger said. “Nobody’s going to put billions of dollars into chips that you don’t have power for.”
Katti said energy availability, manufacturing capacity and deployment timelines have become intertwined problems that must be solved together if AI infrastructure is to keep pace with model development.
The Economics Still Don’t Work #
Gelsinger’s final concern was economics.
“The economics today suck,” he said.
Despite unprecedented investment in AI infrastructure, he argued that the cost of producing intelligence remains far too high.
“It doesn’t need to get better. It needs to get 10,000 times better.”
He believes that improvement will come not from a single breakthrough but from advances across semiconductors, memory, networking, power systems and data center architecture.
Earlier in the discussion, Gelsinger argued that reality has already shifted value toward the companies building AI’s physical foundation.
“The silicon guys are the big winners,” he said. “It has never been this good to be a silicon guy.”
For data center operators, the message was clear: The next phase of AI will not be won simply by deploying more GPUs. It will be won by building an infrastructure stack that delivers more intelligence dramatically with far greater efficiency than today’s hardware can provide. Ai4 continues this week.