Marvell Launches Teralynx T100: AI-Purpose-Built 102.4 Tbps Network Switch
Key Takeaways
- ▸Marvell launched the Teralynx T100, a monolithic 102.4 Tbps switch chip with 25% lower power consumption than competitors—critical for hyperscalers managing 120+ KW GPU racks
- ▸The chip supports 512-port radix for scale-out deployments and programmable protocols for emerging scale-up fabrics, enabling operators to consolidate network tiers and reduce latency in large AI clusters
- ▸Nvidia's $2 billion investment and CEO Jensen Huang's public endorsement signal strong industry backing, despite Marvell entering a market where Broadcom and Cisco have already shipped solutions
Summary
Marvell announced the Teralynx T100 switch silicon at Computex 2026, a 102.4 Tbps networking chip purpose-built for AI datacenter infrastructure. The fabless semiconductor firm claims 25 percent lower power consumption than competitive solutions from Broadcom and Cisco, with improved latency for AI training and inference workloads. Manufactured using 3nm process technology, the chip operates at under 1000W typical power and supports configurations ranging from scale-out deployments with up to 512-port radix to scale-up fabric protocols including Ethernet Scale-Up Networking (ESUN) and Ultra Ethernet Consortium (UEC) standards.
Nvidia CEO Jensen Huang publicly praised Marvell at the event, calling it the "next trillion-dollar company" and emphasizing the critical importance of networking and connectivity chips to distributed AI infrastructure. Nvidia's confidence is backed by a $2 billion investment made earlier this year alongside a strategic partnership connecting Marvell's silicon with Nvidia's AI factory initiative. The Teralynx T100 will begin sampling to customers this quarter in multiple package configurations including BGA, co-packaged copper, and co-packaged optics implementations.
- Production sampling begins this quarter with multiple co-packaging options (BGA, CPC, CPO) for flexible datacenter deployment architectures
Editorial Opinion
Marvell's entry into the AI switch market carries real technical merit and institutional backing that matter: a genuine 25% power efficiency improvement is material at hyperscale, and Nvidia's partnership signals coordinated optimization for AI infrastructure. However, arriving after Broadcom and Cisco have established customer relationships, Marvell faces an uphill battle to claim market share in what may become a commoditized segment. The company's success will likely hinge less on chip quality—which appears solid—and more on whether Nvidia's influence can drive adoption at a time when major hyperscalers already have qualified alternatives.
