As technology heavyweights including OpenAI, Amazon, Anthropic, Tesla, Google and Microsoft pour massive resources into custom in-house processors, former Intel CEO Pat Gelsinger has offered a reality check: designing specially made silicon with AI will not solve the artificial intelligence industry’s most stubborn physical shortages. In discussions exploring how AI is changing semiconductor design, Gelsinger acknowledged that machine-learning tools are accelerating the creation of microprocessors. However, he warned that the true bottlenecks constraining the global AI infrastructure race extend far beyond the speed of chip design itself.“AI will make chip design faster. It won’t make chips easier. It’s not just compute,” Gelsinger wrote, highlighting a discussion he held with Andreessen Horowitz (a16z) partners Raghu Raghuram and Guido Appenzeller on what is actually limiting the broader rollout. He highlighted four key chokepoints:1) Silicon still takes quarters to manufacture, no matter how fast we design2) Memory bandwidth and advanced packaging are huge choke points3) Copper is giving way to optical4) Energy capacity is becoming economic capacityI also think today’s explosion of specialized AI chips will consolidate around fewer architectures.
Where AI helps in chip design, and where it falls flat
Addressing whether current industry hardware, like Nvidia’s Blackwell, represents the final generation of processors designed prior to automated AI engineering, Gelsinger noted that the shift is already well underway, pointing to OpenAI’s custom silicon effort, code-named “Jalapeno.” He characterized projects like Jalapeno as an example of designing from first principles, where machine learning automates vast portions of the workflow and discards legacy design conventions.Yet, Gelsinger stressed that significant parts of semiconductor engineering remain outside the reach of automated AI tools.“When you look at things like Jalapeno (OpenAI chip) today, you say that’s sort of a first principle’s use of AI in the chip design process where you throw out a lot of those things. Now, there are pieces of design that are still really hard right in that sense and a lot of the analog aspect is probably the best example that there are things that are still not ‘AI-able’. You just need lots of silicon data to get those. So you either get so conservative in your analog design that you’re able to say ‘AI it” or you have to do those the hard way,” he said during the conversation.To cope with these limits, Gelsinger explained that engineers must either design their analog systems with extreme conservatism to allow AI tools to process them, or continue engineering those components “the hard way.”