20VC: Cerebras CEO on the Future of Data Centres, Token Costs and Memory | We are Not in an Infra Bubble & Dario Got a Bad Deal with Elon for Compute | Should US Companies Sell to China & Why Most Layoffs are AI Washed with Andrew Feldman

What It Covers

Cerebras CEO Andrew Feldman, following the company's landmark semiconductor IPO that priced at $185 and closed at $311, examines AI infrastructure demand outpacing supply, memory shortages persisting for years, why most corporate layoffs are COVID-hiring corrections rather than AI displacement, and why selling advanced chips to China represents a strategic mistake regardless of revenue opportunity.

Key Questions Answered

• •AI Infrastructure vs. Bubble Logic: Cerebras carries a $25 billion backlog because data center construction cannot match current demand — the opposite of historical bubbles like 1990s fiber optics, where supply preceded demand. Builders are not speculating on future customers; they are already behind on confirmed orders. Recognizing this distinction helps investors and operators avoid misreading infrastructure spending as speculative overreach.
• •Memory Supply Constraint Timeline: HBM memory, produced only by Samsung, Micron, and SK Hynix, faces shortages expected to last several years because fab capacity expands in $40 billion step-function increments requiring five-year build timelines. Micron currently earns 85% gross margins — software-level profitability on hardware — signaling extreme supply tightness. Companies architecting around HBM dependency, as Cerebras does using on-chip SRAM, avoid this bottleneck entirely.
• •Layoff Misattribution to AI: Feldman argues 90–95% of current corporate layoffs stem from COVID-era overhiring and productivity gains from pre-AI tooling, not generative AI displacement. Middle management roles built around information gathering and synthesis are shrinking due to software consolidation, not LLMs. Leaders should separate genuine AI workforce impact from restructuring that would have occurred regardless, avoiding both panic and complacency about actual AI-driven job transformation.
• •Enterprise AI Adoption Blockers: Legal and security teams represent the primary brake on enterprise AI adoption, not data quality or infrastructure readiness. Their incentive structure — zero credit for successes, full blame for failures — creates systematic risk aversion toward unproven technology. Organizations that want faster adoption should have senior leaders set explicit governance frameworks and risk tolerances, removing the decision burden from legal and security functions not designed to approve novel tools.
• •China Chip Sales Strategic Error: Selling leading-edge semiconductors to China benefits adversaries militarily and industrially, regardless of revenue upside. Feldman points to solar, lithium batteries, and automotive sectors as evidence of China's industrial policy effectiveness when given technology access. Credible chokepoints exist through TSMC and ASML dependencies, making export controls enforceable. US chip companies arguing for continued sales to preserve ecosystem influence underestimate the compounding industrial disadvantage created over time.
• •Speed as Infinite-Value Differentiator: Cerebras runs models like Kimi K2 at 6.7x faster throughput than competing GPU clouds, and Feldman frames inference speed as having no upper bound on value — analogous to how there is zero market for slow internet regardless of price. For agentic workflows and coding tasks specifically, a 3-minute versus 20-minute task completion gap compounds dramatically across workdays, making speed a structural competitive moat rather than a marginal performance feature.