CPUs can't be made fast enough: Tan on Intel's $23B raise, the Nvidia co-opetition and the heterogeneous AI future
Key points
Key takeaways from an AI Infra Summit interview with Intel CEO Lip-Bu Tan (September 2026):
Listen for bad news, delight the customer. Tan, who ran Intel since March 2025 after his Cadence turnaround, says he surrounds himself with people who tell the truth and treats bad news as the most valuable signal; he cut Intel from over 100,000 people to 78,000 and flattened management from 10-12 layers to four or five to keep accountability and direct customer feedback.
AI is bigger than the internet — and a bubble. He has lived through multiple hype cycles and views a correction as natural, but warns Silicon Valley's AI euphoria meets deep anti-AI sentiment elsewhere in America; leaders need a positive narrative on security and privacy. Intel itself runs Gemini, Anthropic and Perplexity internally to redesign how it designs and manufactures chips.
The CPU is the bottleneck nobody saw. Reinforced learning and agentic AI need CPUs for orchestration, control planes and general-purpose processing, and demand now outstrips supply: Intel can fill only about half of customer orders, with CEOs calling him directly for more, while AI data centres move from roughly one CPU per 30 GPUs to a far higher ratio.
Gaudi's lesson: keep the founders. Intel acquired Habana Labs but did not keep the team; the founders left and founded a new chip company doing very well, a mistake Tan contrasts with Annapurna Labs, where Amazon kept the founders and the business flourished. His rule: empower the entrepreneurs or lose the talent.
The heterogeneous future. GPUs are power-hungry, so a decade ago he backed SambaNova (dataflow architecture) and Cerebras (wafer-scale) to deliver performance at a fraction of the power; Intel is also exploring brain-inspired chips, neuromorphic designs and memory-efficient compute, and sees software-defined, full-stack systems — compilers, MLIR, platform engineering and rack-scale — as the way to lift low GPU utilisation.
Co-opetition with Nvidia. He reached out to Jensen Huang after taking the CEO job: CUDA is "fantastic software", Nvidia brought platform and system thinking, and the two are now in an engineering-to-engineering collaboration, "coopetition" — partnering where Intel is strong and competing elsewhere to grow the market together.
The $23B raise, oversubscribed 5.6x. Intel needed to rebuild its balance sheet and its foundry economics — 18A is now in volume production with 14A ramping, 0.9 PDK due in October and 1.0 PDK in Q1, plus advanced packaging with 14x radical interconnects on two layers versus the three-layer interposer approach. Substrates are the tight spot: only four key players, two in Japan and two in Taiwan.
Memory and power are next year's walls. Tan called memory the big bottleneck at the start of the year; it has only worsened, with prices up five to seven times and capacity constrained, so he is backing optical (Celestial.ai sold to Marvell, Dust Photonics to Cradle), liquid and microfluidic cooling, and whole-system design — "listen to the customer workload and tailor to their needs".
Read more: YouTube — Intel Business (The AI Opportunity: Silicon to Systems)
