Scaling AI Infrastructure for the Agentic Era | Charlie Kawwas, Broadcom | RAISE Summit 2026

Strategic Shift from General to Purpose-Built AI Hardware

• Frontier labs (OpenAI, Anthropic, Google, Meta, XAI) are rejecting NVIDIA's general-purpose platforms to eliminate two specific "taxes":
  • Efficiency loss from running specific workloads on generic compute.
  • High margin costs associated with NVIDIA's proprietary ecosystem.
• Current spending by the top four U.S. labs ($150B–$200B each) represents a potential lost opportunity of $100B–$150B in token production if a general platform is used.
• Broadcom positions itself as a "couture" design house, allowing clients to customize XPU architectures for specific inference, decode, or training workloads rather than accepting a pre-built rack solution.

Hardware Evolution and Performance Metrics

• Memory Capacity: XPU memory has grown 3x year-over-year (from two memory cubes to six); projected growth indicates a 3x increase again for the following year.
• Compute Power: Intermediate compute capabilities have increased nearly 3x annually; total compute scaling projects a 16x increase in two years and an order-of-magnitude jump within three years.
• Modular Architecture: Chips utilize a "Lego-like" modular system of pre-built, production-ready chiplets (compute, memory, networking) stacked face-to-face to achieve massive density (e.g., 16x capacity of 2022 technology).
• Timeline Projections:
  • 2026: Deployment of current high-density chips.
  • 2027: Further scaling of modular chiplets.
  • 2028: Deployment of "beast" chips designed to enable "agentic AI" and specialized domain experts.

Commercial Scale and Revenue Forecasts

• Broadcom's AI chip business has grown from zero five years ago to projecting over $100B in revenue within the next three years.
• Production volume is ramping rapidly:
  • Baseline deployment: ~1 million chips/year.
  • Near-term target: 40x–50x increase in total compute capacity compared to early generations.
• Power Requirements: Major clients are contracting for massive power expansion:
  • Current deployment: 1.5 gigawatts.
  • Next year: >5 gigawatts.
  • Two years out: >10 gigawatts (totaling ~17 gigawatts of incremental demand across the sector).

Co-Design and Open Standards Strategy

• Deep Co-Design: Broadcom enables tight integration between model development and chip architecture, allowing for faster time-to-market (e.g., "Jalapeno" chip delivered in nine months).
• Networking: Shift toward open Ethernet-based interconnects for both scale-out and scale-up architectures, challenging proprietary silicon-to-silicon connections.
• Hybrid Open/Proprietary Models: Partnerships with startups like Samba Nova demonstrate the ability to run closed frontier models and open-source models (e.g., Minimax, DeepSea) on the same XPU, enabling zero-token-cost inference for open-source tasks.
• Market Dynamics: While NVIDIA accelerates its cadence to one year, Broadcom is the only competitor capable of matching this speed through its modular, integrated XPU platform.

Future Challenges and Edge Migration

• Supply Chain Constraints: The surge in AI demand (gigawatts, wafers, memory) creates supply challenges that may impact non-AI business segments (11 of Broadcom's 16 franchises).
• Manufacturing: Broadcom is establishing its own factories in Singapore to fabricate advanced packaging for high-density chips that existing external partners cannot support.
• Strategic Outlook: AI is projected to migrate from data centers to the edge (smart devices), requiring a dual-focus strategy on massive data center compute and distributed edge AI capabilities.