Cerebras CEO on the Future of Data Centres, Token Costs & Memory | Should US Companies Sell to China
Why It Matters
Supply constraints and geopolitical risks will dictate AI hardware margins, making Cerebras’ SRAM advantage a critical competitive edge for investors.
Key Takeaways
- •AI data‑center buildout lags demand, creating $25B backlog.
- •Memory shortages persist; HBM supply constrained, driving price spikes.
- •Cerebras’ SRAM‑based chips avoid HBM scarcity, gaining cost advantage.
- •AI compute cost per token expected to fall as architectures improve.
- •US‑China tensions may force firms to rethink AI hardware sales abroad.
Summary
Cerebras CEO Andrew Feldman joined a discussion on the future of AI‑driven data centres, the ongoing memory crunch, and the strategic implications of U.S.–China relations for semiconductor sales.
Feldman warned that data‑centre construction is falling behind demand, leaving a $25 billion backlog across Nvidia, AMD and others. The surge in AI workloads has also strained high‑bandwidth memory (HBM), with Samsung, Micron and Hynix unable to keep pace, pushing prices up five‑fold. Cerebras, however, sidesteps HBM by using SRAM integrated at the 5 nm node, insulating it from the shortage and preserving margins.
He noted that AI models became truly useful around 2025, sparking exponential growth in token consumption. As architectures improve, the industry expects a massive reduction in cost per token, even as overall compute demand remains high. Feldman also debated whether Google’s full‑stack approach could make it the lowest‑cost token producer, highlighting the trade‑off between internal consumption and external sales.
The conversation underscores that supply‑chain bottlenecks will persist for years, rewarding companies with alternative memory strategies and prompting U.S. firms to weigh the risks of selling AI hardware to China. Investors should monitor fab capacity expansions, memory pricing trends, and geopolitical policy shifts as they will shape the profitability of the AI infrastructure market.
Comments
Want to join the conversation?
Loading comments...