Marvell launches 102.4 Tbps switch designed to power AI data centers

Marvell Technology just dropped the Teralynx 102T, a networking switch capable of pushing 102.4 terabits per second of bandwidth. To put that in perspective, that’s enough throughput to transfer roughly 12,800 copies of a 4K Blu-ray movie every single second.

The device is being positioned as the industry’s first monolithic die engineered specifically for scale-up switching in AI data center environments. It doubles the bandwidth of Marvell’s previous Teralynx generation, which topped out at 51.2 Tbps and is already in volume production serving AI cloud deployments.

What the Teralynx 102T actually does

The Teralynx 102T addresses this with 64 ports running at 1.6 TbE each. In English: each individual port can handle 1.6 terabits per second of Ethernet traffic, and there are 64 of them working simultaneously.

Marvell is emphasizing two key selling points beyond raw throughput. The first is ultra-low latency, which matters enormously in scale-up AI clusters where even microsecond delays compound across thousands of nodes. The second is thermal efficiency, a factor that’s become increasingly critical as data center operators wrestle with power consumption and cooling costs that are spiraling alongside AI demand.

The “monolithic die” design is worth noting. Rather than stitching together multiple smaller chiplets, Marvell built the entire switch on a single piece of silicon. This approach typically yields lower latency and simpler integration, though it comes with manufacturing complexity tradeoffs, particularly at the leading edge.

Marvell is also leaning into compatibility with SONiC, the open-source network operating system originally developed by Microsoft and now widely adopted across hyperscale data centers.

The Broadcom problem

Marvell isn’t operating in a vacuum here. Broadcom’s Tomahawk 6 series offers identical 102.4 Tbps throughput and has already begun production shipping. That head start matters in a market where hyperscalers like Amazon, Google, and Microsoft tend to qualify and deploy networking hardware on long procurement cycles.

Marvell’s previous Teralynx generation already proved the company could compete at the high end of data center switching. Securing volume production deployments at 51.2 Tbps gave Marvell credibility with the exact customers who will evaluate the 102T.

What this means for investors

Marvell’s move to 102.4 Tbps signals that the bandwidth arms race in AI networking is doubling roughly in line with each generation. The jump from 51.2 Tbps to 102.4 Tbps represents a clean 2x improvement.

The risk, naturally, is execution. Broadcom has deep relationships with hyperscalers and a proven track record in this market. Marvell needs to deliver on the performance and efficiency promises, hit production timelines, and price competitively, all while Broadcom is already shipping. Investors should watch for design win announcements from major cloud providers as the most meaningful signal of whether the Teralynx 102T is gaining traction.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.