Fireside Chat, Interview, Conference Presentation
Former Intel CEO: Why This is the Best Time to Build Hardware
- Major new data center projects face a high risk of defaults or abandonment due to insufficient national energy capacity growth, which is currently limited to approximately 4% annually, necessitating a resurgence in nuclear energy as a base load solution.
- The memory and compute sector is poised for a significant innovation shift after 30 years of stagnation, with material needs driving a transition to stacked memory configurations and a stabilization of integration at 2-to-4 layer stacks rather than deeper 16 or 32-layer arrangements.
- Interconnect architecture is expected to undergo a paradigm shift where copper interconnects become obsolete and optical solutions dominate Input/Output functions within the 2028 to 2029 timeframe, specifically moving large-scale AI clusters toward optical circuit switching (OCS) instead of traditional packet switching.
- The processor and hardware landscape is forecast to converge from its current state of approximately 100 competing vendors back to a few dominant players, driven by the unsustainability of extreme specialization and a blurring distinction between training and inference workloads that favors continuous learning.
- Future hardware ecosystems will likely see consolidation where major AI companies select a limited number of optimized hardware platforms from specific partners, reducing the fragmentation of the current market.
- Power infrastructure will evolve through the adoption of 800-volt DC systems and vertical GaN to minimize conversion steps, alongside anticipated superconducting technology improvements that could enhance power performance by a factor of one thousand.
- The primary operational constraint will remain the design-to-production bottleneck, characterized by a current nine-month lag between design completion and scalable deployment, which persists despite AI tools overcoming analog design limitations.
- Virtual machine technology is set to be reimagined specifically for managing AI agents rather than humans, introducing new security protocols and policy constraints tailored to autonomous workloads.
- The industry will reject widespread Optical-to-Electrical-to-Optical (OEO) interfaces within core compute complexes due to their power inefficiency, favoring direct optical integration into the compute fabric.