Interview, Fireside Chat, Conference Presentation
Andrew Feldman on Building a Chip 58x Larger Than Nvidia's
- Global demand for AI infrastructure (chips, memory, data centers) has exceeded all market forecasts, creating a competitive race for resources.
- Cerebrus announced a partnership with OpenAI in January to supply over $20 billion in hardware over the coming years, described as one of the largest deals in Silicon Valley history.
- Rapid AI adoption has outpaced the traditional speed of real estate development, resulting in a critical bottleneck where data center construction cannot keep up with compute demand.
- Data center infrastructure, largely unchanged for 20 years regarding power generation and cooling, is undergoing a significant architectural overhaul to support high-density AI workloads.
- Cerebrus is pursuing a "co-design" approach where hardware and software are developed simultaneously to maximize performance, a process that involves difficult trade-offs between engineering constraints.
- Concerns exist that NVIDIA is leveraging its balance sheet and market dominance to limit competition by investing in competitors (NeoClouds or model builders) to discourage the adoption of non-NVIDIA chips.
- The speaker recommends that startups avoid dependency on single-vendor ecosystems by sourcing chips from multiple suppliers, including AMD and Cerebrus, rather than accepting free tokens from dominant players.
- The speaker envisions a future where AI contributes to eliminating cancer as a primary cause of death within 25 years, citing the technology's role in accelerating drug discovery and biological research.
- AI is projected to significantly reduce traffic fatalities by enabling autonomous vehicles to outperform human drivers, addressing the leading cause of death for people aged 15–40.
- The speaker advocates for personalized education via AI tutors that adapt to individual learning styles, contrasting this with the rigid, one-size-fits-all classroom model used for the last two millennia.
- Cerebrus's recent IPO created approximately 1,000 millionaires among investors and employees, highlighting the wealth generation potential of hardware-focused AI infrastructure companies.
- The speaker notes that the current hype cycle for hardware differs from previous years where building hardware was considered "uncool," now driven by genuine necessity and a "Cambrian explosion" of architectural innovation.
- While space-based data centers are discussed as a future possibility, the speaker estimates such infrastructure is at least five years away due to the immense logistical challenges of chip communication in orbit.