Broadcom has revealed it has added a fourth provider of foundational artificial intelligence (AI) models to its customer base at a time when revenue generated by AI semiconductors for its third fiscal quarter reached $5.2 billion, a 63% year-over-year increase.
Additionally, Broadcom has secured commitments worth more than $10 billion for AI racks based on its XPU portfolio of processors for providers of foundational models, which reportedly now includes a set of custom processors being built in collaboration with OpenAI.
Broadcom president and CEO Hock Tan told industry analysts that, for all practical purposes, there are only seven potential providers of foundational AI models that require AI semiconductors, and Broadcom continues to woo the other three. In its forthcoming fiscal fourth quarter, Broadcom is now projecting revenues of approximately $6.2 billion, up 66% year-on-year, driven largely by a “material improvement” in AI semiconductor revenues, notes Tan.
However, as training of foundational AI models becomes more distributed, many of these organizations will also require high-end network processors that enable them to create logical clusters that can be as far as 100 kilometers apart, noted Tan. That’s critical because an AI rack scales up to 72 GPUs at 28.8 terabits per second bandwidth using a proprietary NVLink. Broadcom is now making a case for a Tomahawk 6 processor based on Ethernet that can scale up to 102 terabits per second per switch. In general, Ethernet switching remains faster than either NVLink from NVIDIA or InfiniBand, says Tan.
Daniel Newman, CEO of The Futurum Group, said it’s likely that Broadcom will eventually hold an 80% share of what will become a $100 billion market for AI semiconductors, which will also drive demand for networking technologies. With its customer base, Broadcom is incredibly well-positioned, he added.
In total, consolidated revenue for Broadcom’s third quarter was a record $16 billion, up 22% from a year ago. Revenue for the company’s semiconductor segment was $9.2 billion, a 26% increase year over year, accounting for 57% of total revenue in the quarter.
Less clear is the degree to which investments in AI will drive demand for additional IT infrastructure among enterprise customers. Many of these organizations are customizing AI models using foundational AI models, but Broadcom is projecting that demand for processors aimed at enterprise customers is expected to remain relatively flat. Non-AI semiconductor revenue for the third quarter, with much of that demand driven by telecommunication carriers investing in broadband projects. Broadcom is forecasting non-AI semiconductor revenue to grow sequentially to approximately $4.6 billion in the next quarter, but it’s not clear how much of that demand will be driven by enterprise applications.
It’s too early to say when AI use cases will drive much larger spending on IT infrastructure, but the potential is significant. In the meantime, it’s clear that a handful of organizations training foundational models are driving the bulk of current spending. Whether that spending is creating an AI market bubble or is a precursor to an even larger wave is difficult to determine, but the one thing that is certain is that billions of dollars will be spent in the months and years ahead to find out.
