At Computex in Taipei on June 2, Intel’s new CEO Lip-Bu Tan chose not to pretend the company can outmuscle Nvidia in training accelerators. His pitch was narrower and more urgent: as inference, agentic systems and physical AI spread, the missing component may be CPU capacity rather than more GPUs.
The showcase was Xeon 6+, the first Intel data-center CPU slated for volume production on the company’s 18A process. Intel says a liquid-cooled 32U rack can hold 36,864 cores and draw about 100 kilowatts. Tan said Intel is positioned to bring innovation from chip to system level as inference and agents rise.
Intel’s argument is that training often pairs CPUs and GPUs at roughly 1 to 4, while agent workloads can move closer to 1 to 1 because they spend so much time scheduling, queuing, moving data and coordinating subtasks. The company also leaned into partnerships: Xeon with SambaNova SN-50 chips in inference racks built by Foxconn, Vector Core cloud systems mixing Xeon 6, SambaNova and Nvidia Blackwell GPUs, and edge or robotics work with Siemens, Hitachi and others. The deeper test is 18A, because Xeon 6+ is also a manufacturing proof point for Intel’s foundry strategy.
Sources：Intel Newsroom announcement from Computex; CNBC coverage of Nvidia’s broader AI stack push; CocoLoop