News

Broadcom CEO Hock Tan says hyperscaler internal AI chip projects face extreme challenges in packaging and volume yields

Thursday, March 5, 2026 at 01:58 AM

Broadcom CEO Hock Tan addressed the threat of hyperscalers moving to internal customer-owned tooling (COT) models for AI chips. Tan argued that Broadcom's 20-year history in silicon design, proprietary SerDes technology, advanced packaging expertise, and ability to achieve high-volume production yields create a significant moat. He emphasized that LLM developers need chips competitive with Nvidia and that lab-scale designs are insufficient for mass production of 100,000+ units at affordable yields.

Context

Broadcom CEO Hock Tan issued a firm defense of the company’s custom silicon moat during a recent earnings call, projecting AI chip revenue to reach significantly in excess of $100 billion by 2027. Addressing the rise of customer-owned tooling (COT) programs from hyperscalers like Google, Tan dismissed these internal efforts as "science projects" that face "tremendous challenges" in technical execution. He specifically cited the extreme difficulty of mastering advanced packaging, 224G SerDes connectivity, and achieving high-volume production yields at scale. Tan argued that while competitors can design chips in a lab, few can reliably produce 100,000 units with affordable yields. This positioning is critical as Broadcom expands its portfolio, including its 51.2-Tbps Bailly co-packaged optics platform and the upcoming Tomahawk 6 switch. By securing the supply chain through 2028 and adding new partners like OpenAI for 2027 volume ramps, Broadcom aims to maintain a 12 to 18-month lead over internal hyperscaler designs.

Related Companies

Broadcom
Broadcom
AVGO
US
Nvidia
Nvidia
NVDA
US
Google
Google
GOOGL
US