OpenAI Debuts ‘Jalapeño’ Inference ASIC, Kicking Off Multi-Generation Silicon Plan with Broadcom

Big tech writing its own chips is not new—IBM’s 1964 System/360, Apple’s A4 in 2010, and Google’s first TPU in 2015 each marked moments when a software-centric firm seized vertical control to escape supplier choke-points. OpenAI’s leap mirrors those pivots but is compressed into a breathtaking nine-month sprint, enabled by using its own language models as EDA co-pilots—a feedback loop unheard-of even five years ago. The move underscores a structural trend: hyperscalers are re-internalising critical hardware to tame skyrocketing inference costs (OpenAI forecasts a US$1.4 trillion compute bill over eight years) and reduce exposure to Nvidia’s 75% gross-margin toll gate. Whether Jalapeño itself succeeds technically matters less on a century horizon than the precedent it sets: AI models are beginning to design the fabs, networks, and energy systems that will host their successors, hinting at a self-reinforcing industrial stack reminiscent of the shift from railroads to vertically integrated oil in the late 19th century. If this flywheel holds, today’s chip reveal could be remembered not as a product launch but as the moment AI began to own the means of its own production.