Pat Gelsinger just told a room full of AI infrastructure buyers that today’s GPUs aren’t good enough, and he’d know—he ran Intel until 2025. At Ai4 2026, the former CEO and OpenAI’s head of compute laid out why buying more accelerators won’t fix AI’s real problems: power inefficiency, memory bottlenecks, network limits, and economics that still don’t pencil out. We walk through their case for rebuilding the entire compute stack, not just the chip, and why they think silicon vendors are the actual winners here.
(Source: Ai4)
Pat Gelsinger delivered a blunt verdict on AI hardware at Ai4 2026 in Las Vegas: Today’s GPUs use power inefficiently and are computationally limited. The former Intel CEO, now a general partner at the VC firm Playground Global after his ambitious Intel comeback stalled and he left the company in 2025, framed the comment as a starting point, not a conclusion. His real argument targeted the infrastructure surrounding the chip, not the chip alone. Gideon Lewis-Kraus of The New Yorker moderated the keynote conversation, which centered on AI’s physical foundation instead of the models running on top of it.
Gelsinger’s central claim: Every AI breakthrough rests on semiconductors first. He called chips the underlying fuel of a token-driven economy and described demand for intelligence as effectively unlimited, which, in turn, creates near-unlimited demand for compute, infrastructure, and memory underneath it. That framing matters for anyone sizing a procurement budget: Unlimited demand on one side of the equation means the constraint sits entirely on the supply side—chips, power, and infrastructure—not on whether customers want the compute. AI’s current pace of development, he argued, only exists because it builds on decades of prior investment in the Internet, semiconductors, and cloud infrastructure that nobody had to build from scratch this time around.
Gelsinger organizes his current premise around four priorities: infrastructure, efficiency, energy, and economics. On infrastructure specifically, he called for more semiconductors, more data centers, and more communications capacity, full stop. He singled out high-bandwidth memory for direct criticism, calling HBM a genuinely inefficient technology and, at the same time, the best option the industry currently has for AI training and inference. Real gains, in his view, require redesigning far more than the processor: Packaging, memory, networking, and system architecture all need work, together, not one piece at a time.
Sachin Katti, OpenAI’s head of compute, joined Gelsinger on stage and reframed the industry’s real challenge. Convincing people that AI has value is no longer the hard part, he said; convincing governments, utilities, and manufacturers to build enough physical infrastructure is. Katti described bottlenecks at nearly every layer of the supply chain: data centers that can’t get built quickly enough in the US, energy that’s hard to source, fab and memory capacity that falls short of demand, and integration work that doesn’t happen fast enough even when the individual pieces exist. OpenAI’s stated goal is to compress deployment timelines from roughly three years down to three quarters, a target that depends on faster chip manufacturing, standardized data center construction, expanded electricity infrastructure, and more efficient deployment of the compute that already exists.
Energy emerged as a shared concern for both speakers. Gelsinger framed it directly: In a digital AI economy, economic capacity equals energy capacity, meaning future AI growth depends as much on expanding electricity generation as on advancing semiconductors themselves. Nobody builds a data center for chips they can’t power, he said, and nobody puts billions of dollars into chips without the power available to run them. Katti agreed that energy availability, manufacturing capacity, and deployment timelines have become one intertwined problem, requiring a joint solution, not three separate fixes.
Economics rounded out Gelsinger’s list of concerns, and he didn’t soften the assessment: The economics today don’t work, even with record investment in AI infrastructure. The cost of producing intelligence remains far too high, and closing that gap requires roughly a 10,000-times improvement, not an incremental one. That improvement, he argued, won’t come from a single breakthrough. It comes from cumulative advances stacked across semiconductors, memory, networking, power systems, and data center architecture, all moving together. That reframes the competitive question for anyone buying AI infrastructure right now: The winning vendor is the one solving the whole equation, not the one with the fastest chip in isolation. Earlier in the conversation, Gelsinger pointed to where value has already shifted: toward the companies building AI’s physical foundation. Being a silicon vendor, in his framing, has never carried more upside than it does right now.
For silicon teams, Gelsinger’s critique reads as a roadmap of where the next design wins sit: not in raw GPU throughput alone, but in packaging, memory interconnect, and power delivery working as one system. For CIOs and infrastructure buyers, the message cuts the other direction: Procurement decisions built purely around GPU count or peak FLOPS miss the actual constraint, which is whether the surrounding stack—power, networking, memory bandwidth—can keep those chips fed and running at a cost that makes economic sense.
The bottom line both speakers landed on: The next phase of AI infrastructure competition doesn’t come from deploying more GPUs. It comes from building a compute stack—chips, memory, networking, power, and deployment process—together that delivers meaningfully more intelligence per unit of energy and dollar than today’s hardware can manage. Hard to imagine anyone disagreeing with that.
None of this changes what ships next quarter. It changes what companies and VCs fund, what companies design, and what they prioritize over the next several years, and it puts memory vendors, power system designers, and networking teams on equal footing with the GPU makers who’ve dominated the AI infrastructure conversation until now.
What do we think?
Gelsinger isn’t wrong, and he’s also talking his book: Playground Global invests in exactly the infrastructure layers he’s describing (one of its major bets is NextSilicon). That doesn’t make the diagnosis inaccurate. GPU count-driven procurement genuinely ignores power, memory, and networking constraints that determine real-world throughput. For CIOs, the practical takeaway is to model total-stack efficiency, not accelerator count, before approving the next infrastructure purchase.
Inflection point
Gelsinger’s “10,000 times better” framing marks an inflection point in how the industry talks about AI economics: Efficiency, not raw compute scale, becomes the metric that matters. When a former Intel CEO and OpenAI’s head of compute both say GPU count isn’t the constraint anymore, silicon teams should read that as a signal to invest in packaging, memory, and power systems, not next-generation accelerator cores. The infrastructure stack, not the chip alone, is what forms the next competitive advantage.
LIKE WHAT YOU’RE READING? TELL YOUR FRIENDS; WE DO THIS EVERY DAY, ALL DAY.