Broadcom drops the hammer on AI networking with Thor Ultra

Broadcom has launched Thor Ultra, an 800G Ethernet Network Interface Card (NIC) purpose-built for AI backend networks, specifically targeting rack-to-rack scale-out connectivity. This new NIC implements Ultra Ethernet Consortium specifications and modernizes RDMA with hardware-accelerated features like packet-level multipathing and selective retransmission, projecting a 10-15% improvement in AI job completion times and significantly lower power consumption (50W) compared to competing DPUs. The launch positions Broadcom to capture a critical segment of the AI infrastructure market, directly competing with Nvidia's Ethernet and InfiniBand solutions, with sampling currently underway.

Analysis

Broadcom (AVGO) has launched Thor Ultra, an 800G Ethernet Network Interface Card (NIC) specifically engineered for AI backend networks. This product targets the crucial 'scale-out' connectivity segment, facilitating rack-to-rack communication in large AI clusters, directly positioning Broadcom against Nvidia's (NVDA) Ethernet and InfiniBand solutions. The launch represents a significant milestone in Broadcom's three-year strategic investment into its AI networking portfolio, aiming to address performance constraints in large-scale AI training.

Thor Ultra incorporates Ultra Ethernet Consortium 1.0 specifications and introduces hardware-accelerated capabilities to modernize RDMA, including packet-level multipathing and selective retransmission. These innovations are projected to yield a 10-15% improvement in AI job completion times, offering a compelling performance advantage. Furthermore, the NIC consumes approximately 50 watts, significantly less than the 125-150W of competing DPUs like Nvidia's BlueField 3, due to its specialized design for AI backend networking.

Broadcom is offering Thor Ultra through various consumption models, including standard PCIe and OCP 3.0 form factors, discrete chips, chiplet integration, and IP licensing, indicating a comprehensive market penetration strategy. The company emphasizes that the network investment can pay for itself through the projected efficiency gains, underscoring a strong value proposition for hyperscalers. This strategic product launch, coupled with its technical advantages and flexible deployment, suggests a strong push into the rapidly expanding AI infrastructure market.

Key Decisions for Investors

• Investors should evaluate Broadcom's (AVGO) enhanced competitive position in the high-growth AI infrastructure market, considering Thor Ultra's performance advantages and flexible deployment models.
• Monitor Nvidia's (NVDA) strategic response and competitive dynamics in the 'scale-out' AI networking segment, as Broadcom's new offering directly challenges its existing solutions.
• Assess the broader implications for AI infrastructure spending, as specialized, high-performance networking solutions become increasingly critical for optimizing large-scale AI deployments.