Intel CEO Lip-Bu Tan: Inference AI Shifts CPU/GPU Ratio Toward 1:2
Alina Collins
Intel CEO Lip-Bu Tan says the shift from AI training to inference is driving CPU demand sharply higher, pulling the CPU-to-GPU ratio from 1:8 toward 1:2; he is simultaneously betting on foundry and new materials, targeting a 10× return in five to ten years.
Why are CPUs suddenly in demand again?
AI is moving from the training phase to inference and multi-agent orchestration — coordinating swarms of agents, a task where CPUs outperform GPUs.
Tan cites developer feedback: the CPU-to-GPU ratio in training was roughly 1:8; he now sees it shifting to 1:4 and even 1:2. In plain terms = where eight GPUs once needed just one CPU, two GPUs now require one — CPU volumes have multiplied.
Intel's data-center server demand is "very high," the core reason Tan says the company's "game is not over."
Foundry burns cash — why not walk away?
"Many voices" urged Intel to exit foundry, arguing it is too capital-intensive and the gap with TSMC too wide. Tan says he chose to "grit his teeth and stay."
His rationale: supply-chain resilience. Global chip manufacturing cannot depend on one or two regions. This reflects Washington's own bet — the U.S. government has become a major Intel shareholder.
He concedes Intel still trails TSMC meaningfully in IP, yield, defect density, and cycle time, stressing that foundry is a "trust business" requiring humility.
Where is the money coming from?
The U.S. government has taken a stake in Intel. Tan told President Trump that TSMC also started with government backing, as did fabs in Japan and Singapore.
Long-time friend Jensen Huang, Nvidia's CEO, invested $5 billion; that stake is now worth roughly $25 billion. SoftBank's Masayoshi Son also contributed funding.
Fourteen months into the job, Tan says he has delivered roughly a 6× return for shareholders — but calls it "just the beginning."
How do you push past the physical limits?
Intel is advancing its 14A (1.4 nm) process node and has begun planning 1 nm and 0.7 nm. In plain terms = circuits on a chip are being etched ever finer, approaching physical limits — each step forward is costlier and harder.
Tan says Intel is going back to materials science, investing in three new material families: gallium nitride, silicon carbide, and indium phosphide.
In advanced packaging, he favors glass substrates and synthetic diamond as next-generation thermal insulators, and has invested in startups in both areas. Intel also announced a packaging-manufacturing partnership with the Indian government, with sites in India and New Mexico.
What about Musk's fab and AI's bottlenecks?
Tan confirms Intel is working with Elon Musk on TerraFab — Musk's plan to build his own wafer fab, with Intel lending process expertise to speed production. Musk's robotics and automotive businesses need large chip volumes.
AI growth faces three bottlenecks: power constraints, tight helium supply, and a memory shortage. This means → whether the industry builds new memory fabs or expands CPU/GPU capacity, each takes years, and rising costs will pass through to customers.
Tan believes AI's impact will surpass the internet's. He expects Intel's full potential to emerge around 2030–2032, noting that servers once served human users but must now serve millions of AI agents requiring compute and software-stack access.
Content is for reference only, not financial advice.