Fireside Chat, Interview
Andrew Feldman, Cerebras Co-Founder and CEO: The AI Chip Wars & The Plan to Break Nvidia's Dominance
- Cerebrus aims to reduce transformer dependency within three to five years by addressing data movement bottlenecks to build faster, lower-power AI computers utilizing wafer-scale integration with hundreds of thousands of tiled, redundant components.
- Future architectures will leverage in-silicon data retention and increased utilization to deliver higher tokens per unit of time, with wafer yield techniques expected to match or exceed those used for smaller chips.
- The AI market is projected to grow over 100 times larger in five years, reaching a population level of AI-hungry users and achieving smartphone-like penetration within one to two years.
- In five years, almost all training data is expected to be synthetic, while scaling laws for inference will continue to yield better answers through increased compute, driving algorithmic efficiency and lower costs.
- Ten-year forecasts predict the discovery of therapeutics for afflictions affecting over a million people annually, with Cerebrus inference powering currently non-existent apps and reaching an inadvertent portion of the U.S. and European populations.
- Hardware is expected to diffuse into everyday devices like cars, pockets, dishwashers, and TVs as AI becomes faster and cheaper, with chip providers anticipated to hold greater enterprise value than model providers in five years.
- Cerebrus plans to expand manufacturing capacity by 2x to 5x, secure a dozen G42-style strategic partnerships within two years, and replicate its strategic partner model with other companies over the next 24 months.
- Data center construction will surge, with Texas benefiting from reduced regulatory burdens while Niagara faces power shortages, leading to uncertainty regarding the quality of new builds and the role of Bitcoin miners like Crusoe in gigawatt-scale projects.
- The company predicts NVIDIA will retain 50% to 60% of the market in five years while its share declines, viewing NVIDIA's business as meaningful for both training and inference despite the absence of CUDA locking in inference.
- Cerebrus expects to capture market growth alongside NVIDIA and others by offering performance that surpasses competitors due to current customer unhappiness with NVIDIA delays and outdated chips.
- Geopolitical strategies include refusing sales to China due to potential misuse, while acknowledging China's substantial infrastructure investments and engineering talent output; the U.S. administration is viewed as more favorable for AI than the prior one.
- Long-term competitive advantage is derived from enduring value and top-tier performance rather than competing on release cadences, with public listing status expected to appeal to large enterprises.
- Compliance for hardware is deemed easier to manage than software due to physical constraints, though delaying technical progress in the Chinese chip market is described as enormously challenging.