NVIDIA contributes Blackwell GPU architecture and Ethernet technology to Meta's Open Compute Project consortium.
Nvidia is often perceived as maintaining a closed ecosystem, particularly with CUDA, which Nvidia says must remain closed "in order to develop an optimized software abstraction for new hardware." However, as AI infrastructure evolves beyond individual processor performance toward full-system challenges, Nvidia is increasingly contributing technology to the Open Compute Program to address the next set of bottlenecks driving cost and energy inefficiencies. The company has maintained a long-standing commitment to OCP, which Facebook started over 10 years ago, and has previously provided reference designs such as the HGX-H100 Baseboard, now used in the majority of AI installations.
Nvidia's most significant recent contribution centers on Blackwell GPU architecture and the supporting networking infrastructure. NVSwitch and NVLink, alongside Nvidia's software, represent the company's principal differentiators. NVSwitch connects 72 GPUs into a GPU fabric, allowing software and AI models to operate as a single massive GPU, improving performance and simplifying development. At the Open Compute Program summit, Nvidia highlighted how these technologies unlock Blackwell's potential within a broader ecosystem.
To enable OEMs such as Supermicro, Dell, HPE, and Lenovo to deliver NVL72 DGX-class systems, Nvidia has open-sourced the NVL72 rack, the MGX compute tray, and the Nvidia Switch tray—a significant contribution given the NVL72 rack's complexity. The design includes 5,000 copper wires, consumes 120 KW of power at 1,400 amps, incorporates telemetry-enabled switches for congestion control, and can support models up to 27 trillion parameters. Demonstrating open systems dynamics in practice, Meta took the NVL72 specifications, modified them for their specific data center requirements, and contributed the result—called Catalina—back to the OCP community for broader use.
Blackwell has entered full production, with Nvidia projecting 250,000 to 300,000 units in the fourth quarter generating $5 billion to $10 billion in revenue, and 750,000 to 800,000 units expected in the first quarter. Nvidia anticipates selling DGX B200 8-way servers for approximately half a million dollars each. The scale of Nvidia's ecosystem now encompasses millions of developers, scientists, and research institutions, alongside numerous OEMs, ODMs, and data center infrastructure providers deploying Nvidia hardware and software to power AI infrastructure advancement.