Broadcom Lands Shepherding Deal for OpenAI "Titan" XPU (www.nextplatform.com)

Broadcom has secured a major deal to shepherd OpenAI's upcoming “Titan” AI inference chip, marking its fourth custom XPU customer alongside hyperscalers like Google and Meta. This partnership is significant as OpenAI aims to reduce its heavy reliance on Nvidia GPUs by co-designing bespoke AI hardware optimized for its models and workloads. The Titan chip, expected to launch in late 2024 or early 2025, will play a critical role in OpenAI’s ambitious Project Stargate to develop next-generation AI infrastructure at scale and lower the astronomical costs of cloud AI computing. Broadcom’s Q3 results underscore the AI business’s growing impact, with AI-related chip sales surging 63 percent year-over-year to an estimated $5.18 billion, driven by strong demand for custom XPUs and AI networking ASICs. CEO Hock Tan revealed over $10 billion in AI rack orders tied to these custom chips, with Broadcom forecasting AI chip revenue to grow another 66 percent next quarter. Although the deal with OpenAI involves complete AI rack systems beyond just the Titan XPU, it signals Broadcom’s expanding role in shaping AI hardware beyond conventional GPU reliance. OpenAI’s new head of hardware, Richard Ho—who brings deep expertise from Google’s TPU development and silicon photonics startup Lightmatter—may help push innovation with advanced optical interconnects in their AI server design. This could position OpenAI to leapfrog Nvidia not only by custom chip design but also by pioneering novel hardware architectures, setting the stage for a new era of large-scale, efficient AI computing tailored to OpenAI’s cutting-edge models.