Arista, Broadcom, Cisco, DriveNets and NVIDIA All Message the Growing Importance of Ethernet Products
The past two months were busy for AI Ethernet Providers, with announcements showcasing new products and providing historical revenue growth forecasts. Overall, the announcements highlight the long-term growth of AI Networking into a $30+B market by 2028 and a transition away from InfiniBand towards Ethernet (Figure 1 below). As AI moves from large foundational training towards reinforcement training (T2), inference and additional accelerators will appear, and Ethernet will be the dominant fabric for those networks.
Broadcom Highlights Size of Ethernet Clusters as Tomahawk 5 and Jericho3AI Platforms Begin to Ramp
Broadcom continued to talk about Tomahawk 5 and Jericho3AI and the increasing size of hyperscaler deployments it expects over the next 12-24 months. Networking is directly correlated to the spending on GPUs and AI ASICs. As customers increase CAPEX expectations, we will see an upside to networking port and revenue projections.
NVIDIA Provides Ethernet Switch Guidance for the First Time on Their Earnings Call
NVIDIA changed its public reporting to include networking and called out Ethernet revenue targets for the first time in its earnings call. Surprising to many, but consistent with our forecasts for Ethernet growth, NVIDIA is seeing strong demand for its Spectrum-X Ethernet ASIC for back-end deployments as the company looks towards T2 and enterprise deployments where Ethernet is preferred and a necessity.
Cisco Confirms $1B+ AI Networking Target and Highlights HyperFabric at Cisco Live
Cisco announced Nexus HyperFabric AI clusters with Nvidia to help users design, validate, deploy, and monitor AI infrastructure. At the investor track, Cisco highlighted the $1B+ revenue expectations in networking from a mix of systems, Silicon One, and optics. Cisco is seeing significant traction with Ethernet in the back-end in proofs-of-concept and trials, and we should see the company’s new products ramp up in 2025.
Marvell Starts Shipping Teralynx 10 (51.2 Tbps ASIC)
Marvell announced that it is shipping Teralynx 10 to help usher in the 800 Gbps switch generation. The 800 Gbps ramps will significantly outpace the markets adoption of 400 Gbps and 100 Gbps and is a key enabler to AI/ML networks. Most AI/ML networks will deploy with at least 800 Gbps of bandwidth per GPU/ASIC.
Arista Celebrates 10th Anniversary Since IPO/ Launches New Etherlink AI Portfolio – Making Ethernet Great
Arista announced its new Etherlink AI switching portfolio and Tomahawk 5 switches before its NYSE 10-year IPO celebration. Hock Tan from Broadcom also attended to congratulate Arista and talk about AI networking.
During its second quarter earning call, Arista validated the migration from InfiniBand to Ethernet, discussing four trials with Cloud and AI Titans and showcasing an AI win with a large Tier 2 cloud provider which has been heavily investing in GPUs to increase their revenue and penetrate new markets. Arista CEO Jayshree Ullal also stated having tens of smaller customers, now starting to do AI pilots.
DriveNets confirms AI Hyperscaler customer and AI Ethernet pilots with other customers
DriveNets is also providing AI Ethernet solutions based on Broadcom’s Jericho switches, supporting both Jericho 2C+ and Jericho3AI. In its recent presentation at the OCP regional summit in Lisbon, Broadcom mentioned that DriveNets’ Network Cloud-AI Ethernet fabric was validated and deployed by a tier-1 hyperscaler. Like Arista, the company also mentioned that it is engaged in several new AI-Ethernet customer pilots.
2025 Forecast for Ethernet Continues to Increase
Bringing all these announcements together, it is clear that Ethernet for AI back-end networks is about to explode, and overall, DC Switching will benefit from additional front-end network upgrades. With the inclusion of optics, InfiniBand, and Ethernet, the AI networking market will exceed $30B in 2028, providing one of the largest growth markets that Ethernet has ever experienced.