Broadcom has announced Thor Ultra, its new 800G AI Ethernet Network Interface Card that follows the Ultra Ethernet Consortium (UEC) standard. The company says this NIC can link hundreds of thousands of XPUs to handle trillion-parameter AI workloads, all while staying in an open, vendor-neutral ecosystem.
Instead of relying on proprietary networking like NVIDIA’s InfiniBand, Broadcom is betting on open Ethernet for AI data centers. Thor Ultra supports advanced RDMA features such as packet-level multipathing, out-of-order packet delivery straight to XPU memory, and selective retransmission to improve fabric efficiency. It also introduces programmable congestion control and line-rate encryption that offloads work from the host processor.

The NIC connects over PCIe Gen6 x16 and supports both 100G and 200G PAM4 SerDes with long-reach passive copper options. Broadcom says it delivers the industry’s lowest bit error rate, which helps prevent link drops and reduces job completion time in large-scale clusters. It’s also designed for secure boot and signed firmware, offering a layer of trust that hyperscale customers will expect.
Thor Ultra integrates with Broadcom’s Tomahawk 5, Tomahawk 6, and other UEC-compliant switches, making it a core part of the company’s push for open AI networking. That strategy is drawing support from vendors such as Arista, Dell, Lenovo, and HPE, who all see potential in an Ethernet-based approach that breaks away from closed ecosystems.
The real test, however, will be whether data center operators embrace this move toward open AI fabrics. For now, Broadcom says sampling is already underway and expects the card to serve as a foundation for the next generation of AI infrastructure.