Big Tech Bets Big on Custom Chips as Broadcom Becomes the Go-To Partner

The Custom Chip Revolution

Broadcom has cemented its position as the go-to partner for custom AI accelerators, landing landmark deals with two of the most influential AI companies in the world. OpenAI has committed to a four-year collaboration worth billions to deploy ten gigawatts of custom chips, while Anthropic has placed orders totalling twenty-one billion dollars for nearly one million Google TPU units.

Breaking Free from Nvidia

The deals underscore a growing trend among hyperscalers to diversify away from Nvidia's dominant position in the AI chip market. While Nvidia controls over ninety percent of the data centre GPU space, major players including Google, Amazon, Microsoft, and Meta are increasingly investing in custom silicon designed specifically for their workloads.

The Numbers Behind the Shift

Broadcom's AI semiconductor revenue grew sixty-five percent year-over-year in fiscal twenty twenty-four, with management projecting first-quarter fiscal twenty twenty-six revenue to exceed eight point two billion dollars. The company's stock has risen nearly five hundred percent since early twenty twenty-three, pushing its market capitalisation past one point five trillion dollars.

Why Custom Chips Make Sense

Custom application-specific integrated circuits offer significant advantages for companies running AI at scale. While less flexible than general-purpose GPUs, ASICs consume less power and deliver better cost efficiency for specific tasks. Analysis suggests Google's TPU version seven offers roughly thirty percent lower total cost of ownership than Nvidia's GB200 platform.

The Road Ahead

Deployments from the OpenAI partnership begin in the second half of twenty twenty-six, with completion targeted for twenty twenty-nine. Anthropic's first phase covers four hundred thousand TPU units worth approximately ten billion dollars in finished racks. Analysts expect Nvidia to retain over seventy percent market share for the next three years, but the competitive landscape is clearly shifting as inference workloads grow and custom silicon becomes increasingly attractive.