Conference Presentation, Fireside Chat, Interview, Product Demonstration, Keynote
Scaling AI Infrastructure for the Agentic Era | Charlie Kawwas, Broadcom | RAISE Summit 2026
- Frontier labs are projected to spend $50 billion to $150 billion while facing a "general compute tax" of $100 billion to $150 billion per player in lost tokens, driving a strategy to build custom platforms at a pace matching NVIDIA's annual cycle.
- Broadcom is currently shipping a next-generation XPU with compute performance nearly 3x stronger and memory capacity over 3x that of the previous year's model, with plans to scale production volume from approximately 1 million units annually to 2 million, and subsequently to 4 million to 5 million units.
- Memory capacity for Broadcom's XPU is expected to increase by almost an order of magnitude within two years, supporting a 2028 timeline for a new chip offering 3x the previous generation's compute and 16 times the capacity of chips from two years prior.
- A "beast" chip currently under development is intended for the largest four LLM players to facilitate agentic AI platform capabilities, while Broadcom anticipates enabling foundation labs to build their own silicon within a year to align with the industry's one-year cadence.
- The AI chip business is forecast to grow from less than $1 billion two years ago to over $100 billion in three years, with customers expecting 2x to 3x more compute per dollar to drastically lower token costs through purpose-built platforms.
- Power demand is predicted to become "insatiable," with a single major customer's deployment rising from 1.5 gigawatts this year to over 5 gigawatts next year, reaching over 10 gigawatts by the year after next.
- Across the next three years, major labs are expected to increase power needs from 2 to 5 to 10 gigawatts each, totaling 17 gigawatts starting from a baseline of 1 gigawatt, necessitating the construction of new factories in Singapore for "big beasts and big ships" that existing technologies cannot produce.
- The company is co-designing open, rack-level solutions with Ethernet-based networking capabilities and anticipates a market shift where AI migrates from data centers to the edge, with hybrid solutions combining closed frontier models and open source models already in use on XPUs.
- By 2028, the technology is expected to enable a proliferation of capable domain-specific agents and contribute to AGI, while the market sees significant capacity growth as Broadcom plays in both data center and edge AI sectors.