Interview, Fireside Chat
Cerebras CEO on the Future of Data Centres, Token Costs & Memory | Should US Companies Sell to China
- Memory shortages and construction delays driven by multi-year fab timelines are expected to persist for several years, with HBM production bottlenecks requiring billions in "step function" investments over five years, leading to a "metered" market that prevents supply gluts despite high demand.
- Current AI infrastructure lags behind demand rather than oversupplying, creating a landscape where exponential demand growth driven by improved model usefulness and broad demographic adoption is projected to continue without peaking soon.
- Companies capable of forecasting and acting on exponential demand growth one to three years in advance possess a competitive advantage, while those purchasing legacy hardware on the spot market will likely fall 1.5 to two generations behind current technology.
- The industry anticipates a massive reduction in the cost per unit of compute over the next few years due to design efficiency, with the performance gap between specific entities and competitors like NVIDIA and AMD expected to widen as all players deliver faster, more power-efficient chips in three to four years.
- Strategic risks include a significant drag on enterprise AI deployment as technological adoption outpaces legal and security frameworks, which are unlikely to adapt quickly to open-source models due to a lack of precedent and regulatory complexity.
- Organizations with decades of disciplined data organization strategies will likely hold a significant advantage, whereas Google's full-stack ownership model may limit market size compared to open-market sales that drive lower costs through volume.
- The US requires aggressive onshoring of TSMC-like capabilities and packaging expertise to maintain strategic advantage, potentially necessitating a 20-year exemption from local ordinances to bypass bureaucratic obstruction, as current infrastructure is described as a "patchwork" of 1950s technology.
- Europe is expected to remain slower in adoption and innovation compared to the US and China due to a cultural preference for regulation over entrepreneurship, while the US semiconductor industry is projected to benefit from current administration policies described as favorable to business.
- Employment trends predict that "AI washed" layoffs result from pandemic-era over-hiring and automation of information gathering, while engineering hiring will likely increase as productivity gains enable teams to tackle 50 times more tasks, accompanied by the emergence of new AI governance roles and the disappearance of HR information manager positions.
- Commercial dynamics include an influx of vendors targeting newly public companies with aggressive pricing and a prediction that the market for "slow search" or slow internet will vanish, as speed remains the essential differentiator for solving hard problems in an AI-driven world.