Broadcom (AVGO, Financials) just rolled out a powerful new networking chip called Tomahawk Ultra, and it's built with one goal in mind — helping AI systems talk to each other faster and at scale. Think of it like a traffic cop for data racing between hundreds of processors packed inside modern data centers.
And it's not just any chip. Broadcom says it can connect four times more chips than Nvidia's (NVDA, Financials) competing NVLink Switch — and does it using a supercharged version of Ethernet instead of proprietary tech.
“We designed this originally for high-performance computing, but it turns out it's perfect for AI, too,” said Ram Velaga, a senior exec at Broadcom. The chip is already shipping and built using Taiwan Semiconductor Manufacturing Co.'s 5-nanometer technology.
AI models are getting bigger, and to keep up, companies need to connect more chips together — fast and efficiently. That's where this chip comes in. Broadcom's solution allows developers and data centers to “scale up,” linking processors sitting just inches apart so they can act as one big brain.
It's also part of a bigger trend: Broadcom, which already helps Google (GOOGL, Financials) build custom AI chips, is quietly becoming one of the few serious challengers to Nvidia in the AI infrastructure race.
