Fireside Chat, Panel, Conference Presentation
The Chip Revolution: Beyond GPUs to Tomorrow's Processors | RAISE Summit 2026
- Inference spend is projected to dominate accelerator expenditures by next year, with energy consumption for inference deployments expected to grow substantially over subsequent years.
- The global economy is forecast to undergo a multi-decade convergence where nearly all global GDP activity centers on inference tasks.
- Hardware designed prior to the ChatGPT era is expected to appear obsolete compared to future trillion-dollar data center requirements.
- A new hardware design from Unconventional.ai is scheduled for tape-out every four months, aiming to achieve 1,000x energy efficiency.
- Products from Unconventional.ai and Etched are slated to begin shipping in the summer, with underlying models designed for autonomous self-orchestration and generation of hundreds of millions of tokens.
- Inference workloads are expected to increasingly concentrate in giant data center clusters, driving market consolidation into larger scale-up domains.
- Fabrication economics will dictate that only facilities at the $40 billion level or higher make sense for advanced nodes like two nanometer, leading to fabs significantly larger than current standards.
- Supply chain constraints remain critical, as a four to five-month fab cycle persists regardless of future improvements in EDA tools, while capital expenditures for hardware must be amortized over three to five years.
- Significant capacity expansion is required to handle demand, necessitating immediate ramp-up of hundreds of megawatts to gigawatts of compute to avoid becoming a bottleneck.
- Nvidia has reportedly booked orders solid through the end of the decade, while hyperscale chips from major cloud providers generally lag one generation behind frontier developments.
- The energy efficiency gap may prompt customers struggling with power requirements to adopt alternative frames, although many large companies remain hesitant to make major strategic shifts.
- Software development barriers are lowering with automated kernel generation becoming the industry standard within one to two years, enabling faster porting and recursive model improvement.
- While larger process nodes offer faster iteration, the industry must balance prototyping costs where chips may become cheaper than GPUs sooner than anticipated despite high training costs.
- Market entry for edge computing is contingent upon models reaching capability saturation, whereas current bets remain strictly focused on the data center.
- Time to market is identified as the most crucial factor, with decisions often favoring incremental improvements over rebuilding systems for marginal gains.