Broadcom Just Dropped a Chip That Could Supercharge 100,000 GPUs

Broadcom (NASDAQ:AVGO) is making a bold move to grab a bigger slice of the red-hot AI infrastructure pie. Over the weekend, the company began shipping its Tomahawk 6 chipits most powerful data center switch to date. Why does this matter? Because when you're running a supercomputer with 100,000+ GPUs, speed isn't just about the chipsit's about how fast they talk to each other. And that's exactly what Tomahawk 6 is built for. Ram Velaga, Broadcom's SVP of Core Switching, says the new chip can do the work of six old ones. That could mean faster AI training, better inference, and lower infrastructure bloat for the same GPU count.

The underlying problem? Most GPUseven the powerful ones from Nvidia (NASDAQ:NVDA)are sitting idle 60% to 70% of the time, waiting for the network to catch up. Think of it like race cars stuck in traffic. Broadcom's switch could unclog those lanes. Early customers, including top cloud providers and networking firms, are already deploying the new chip to stitch together some of the world's largest GPU clusters. No names were dropped, but Velaga hinted at rollouts involving over 100,000 GPUs. Given that you typically need one switch for every 10 GPUs, the upside here could be huge.

And while Tomahawk 6 won't come cheapcosting nearly double its predecessorBroadcom believes the performance payoff justifies the premium. Velaga didn't give an exact price but noted it's below $20,000 per unit, with bulk discounts likely in play. Full availability is coming in July. With AI data centers racing to optimize every watt, every connection, and every dollar, Broadcom might've just positioned itself as an indispensable piece of the next-generation AI stack.

This article first appeared on GuruFocus.