Elon Musk recently unveiled ambitious plans to expand xAI’s Colossus AI supercomputer from its current capacity of 100,000 GPUs to an impressive 1 million GPUs in the foreseeable future. This expansion isn’t unique to xAI, as Broadcom has announced that by fiscal 2027, three of its major clients, all prominent hyperscalers, are expected to launch AI supercomputers each equipped with one million XPUs.

“Currently, we are supporting three hyperscale customers who have each developed their own multi-generational AI XPU roadmap, set to roll out progressively over the next three years,” announced Hock Tan, President and CEO of Broadcom, during the company’s Q4 2024 earnings call. “By 2027, we anticipate that each of these customers will deploy supercomputers with clusters of 1,000,000 XPUs interconnected across a unified network.”

Beyond these three clients, Broadcom revealed they have secured contracts with two additional hyperscalers and are in the advanced stages of developing next-generation AI XPUs for them. There are whispers in the industry about collaborations between Broadcom, ByteDance, and OpenAI for AI chip development, although Broadcom has not confirmed any specific partnerships.

Broadcom’s expertise spans the development of AI chips, general-purpose computing, and customized data center solutions for several major corporations, including tech giants like Google and Meta. The process begins with identifying specific computational needs—such as AI training or data analysis—after which Broadcom and its clients co-design the chip specifications. Leveraging its silicon design prowess, Broadcom crafts the architecture of these processing units, incorporating proprietary IP, cache memory, inter-chip connections, and interfaces. The high-performance XPUs crafted by Broadcom are then produced by TSMC.

Depending on the agreements in place, Broadcom might sell these XPUs or custom ASICs directly to customers under long-term supply contracts. Additionally, Broadcom often plays a role in collaborative product development, offering services that might include engineering cooperation and research and development, typically at a fee.

The XPU division is a strategic component of Broadcom’s business model, aimed at tapping into the burgeoning demand for AI and cloud technologies. Looking ahead to 2027, Broadcom estimates the serviceable addressable market (SAM) for AI XPU and networking hardware to be between $60 billion and $90 billion. The company positions itself to capture a significant market share in this sector. It remains unclear, however, how Broadcom calculates its SAM figures, especially considering that this year alone, Nvidia is projected to generate around $100 billion from sales of GPUs, DPUs, and networking equipment to the AI sector.