Cisco takes on Broadcom with new AI supercomputing chips

Everyone is obsessing over buying up GPUs, but building an AI supercomputer has a dirty little secret: if your networking chips can't shuttle data between those cards fast enough, the whole billion-dollar cluster basically sits around waiting.

Now, Cisco wants a bigger slice of that AI backbone. The networking titan just took the wraps off its new G200 and G202 networking chips, taking direct aim at incumbents Broadcom and Marvell.

Wiring 32,000 GPUs together

The new chips belong to Cisco's next-gen ethernet switch lineup, built specifically to tackle massive AI and machine learning workloads like training ChatGPT-class models.

According to Cisco, the G200 and G202 double the performance of their predecessors and can link up to 32,000 GPUs within a single high-speed network.

Even more important is power efficiency. Cisco fellow Rakesh Chopra noted that by cutting down the total number of switches required and trimming latency, the silicon can slash overall networking power consumption by up to 40%.

Cisco also confirmed that five out of the six major cloud computing providers are already testing the silicon, though it didn't name names. The top tier of that market, of course, includes hyperscalers like AWS, Microsoft Azure, and Google Cloud.

The timing lines up with Cisco's broader push into generative AI and cloud infrastructure. In its latest quarterly results, the company posted record revenue and cash flow, with CEO Chuck Robbins pointing to annual recurring revenue near $24 billion and a growing reliance on AI-driven cloud workloads.

Why it matters

The AI race is quietly shifting from GPU supply to cluster networking. Training a multi-hundred-billion-parameter model requires thousands of GPUs to act like a single brain, and standard data center fabrics simply choke on the traffic.

For major cloud providers and hyperscalers, Cisco's entry offers a badly needed alternative to Broadcom and Marvell. More silicon suppliers mean better bargaining power, less single-vendor lock-in, and more leverage when negotiating supply delivery dates.

In our view, Cisco played this smart by targeting total switch count and a 40% power reduction. When AI data centers are already maxing out regional power grids, wattage savings matter just as much as raw speed.

The challenge, however, is execution. Broadcom has had a lock on AI ethernet switches, and Nvidia continues to push its proprietary InfiniBand fabric. Convincing cloud titans to move from pilot testing to mission-critical, large-scale deployment is where Cisco will actually have to prove its silicon holds up under real-world fire.