A new alliance between OpenAI and Broadcom could redefine how artificial intelligence data centers are built and interconnected. The collaboration blends custom-designed accelerators with open networking technologies, pointing toward a future where dependency on Nvidia’s proprietary ecosystem is reduced.
Building a foundation for custom AI hardware
Under a multi-year agreement, OpenAI is developing its first proprietary AI chips with Broadcom providing the supporting Ethernet-based networking architecture. Large-scale deployment—estimated at around 10 gigawatts of computing capacity—is expected to begin in 2026.
This initiative reflects a move toward vertically integrated AI systems, where model development, silicon design, and networking are optimized together. The combination of custom processors and open networking fabric is intended to meet the growing global demand for compute power and support more flexible, cost-efficient scaling strategies across OpenAI’s own and partner data centers.
From InfiniBand to Ethernet
By selecting Ethernet as its primary interconnect, OpenAI is breaking away from Nvidia’s InfiniBand, which has long dominated high-performance AI computing. Ethernet’s open and standardized nature offers greater interoperability across hardware vendors, allowing hyperscalers and enterprises to mix and match components rather than depend on a single vendor.
This change aligns with a broader industry trend: organizations are increasingly adopting open networking standards to improve scalability, reduce costs, and strengthen digital sovereignty. Ethernet’s expanding role in AI workloads indicates its readiness to support disaggregated infrastructures—where compute, storage, and networking elements evolve independently but cohesively.
A heterogeneous compute landscape
The growing complexity of AI models is pushing data centers toward heterogeneous computing, combining GPUs, ASICs, and other accelerators from various vendors. While Nvidia remains the backbone of most AI training clusters, hyperscalers are exploring alternative silicon designs, including those based on AMD, Arm, and RISC-V architectures.
This diversification supports supply chain resilience and cost optimization—key motivations behind OpenAI’s collaboration with Broadcom. The partnership also complements Nvidia’s recent decision to open its NVLink interconnect to third parties, enabling GPUs to operate more easily alongside custom chips from other vendors.
Networking becomes a strategic layer
The OpenAI–Broadcom collaboration highlights that the network fabric of AI infrastructure is now as critical as the chips themselves. With workloads increasingly distributed across massive clusters, efficient interconnects determine not just performance but also energy efficiency and operational cost.
Adopting Ethernet-based systems allows organizations to standardize hardware, simplify scaling, and reduce the proprietary lock-in often associated with closed interconnect technologies. As a result, the partnership positions both companies at the forefront of a movement toward open, interoperable AI data centers.
Industry-wide implications
The shift toward custom silicon and open networking reflects a broader evolution in AI infrastructure strategy. Major technology companies are striving to balance performance with cost efficiency by designing purpose-built hardware stacks. However, only a few hyperscalers and large generative AI vendors currently possess the expertise and resources to build such systems in-house.
For the majority of enterprises, Nvidia’s full-stack offerings—combining GPUs, software, and interconnects—will likely remain the default choice in the near term. Still, OpenAI’s initiative may accelerate competition among chipmakers and network vendors, fostering greater innovation and reducing the industry’s dependency on a single ecosystem.
