The Twenty Minute VCCerebras CEO on the Future of Data Centres, Token Costs & Memory | Should US Companies Sell to China
CHAPTERS
Cerebras’ IPO moment and the bigger agenda: chips, geopolitics, and energy
Harry sets the stage with Cerebras’ blockbuster public debut and frames the conversation as a forward-looking exploration of AI infrastructure, compute economics, and US–China dynamics. Feldman positions the discussion around where constraints and leverage points really are in the AI stack.
AI infrastructure bubble—or infrastructure chasing demand?
Feldman rejects the “AI infra bubble” analogy by contrasting today with rail/fiber buildouts where supply led demand. He argues the defining feature now is the opposite: demand is already here and infrastructure is lagging, creating persistent backlogs across the ecosystem.
Why data-center “metering” might actually stabilize the market
They discuss permitting and build delays as a form of ‘metering’ that can smooth adoption and prevent overbuild whiplash. Feldman highlights OpenAI’s early recognition of exponential compute demand as a competitive advantage, while noting not all compute deals are equal.
Memory (HBM) as the chokepoint—and why it won’t clear quickly
Feldman explains that explosive demand is stressing every part of the supply chain, with HBM memory a major bottleneck due to limited suppliers. Because fab capacity is added in huge, slow ‘step functions,’ he expects shortages and elevated pricing to persist for years if demand remains strong.
2025 as the inflection: inference demand explodes when AI becomes truly useful
Feldman argues that around 2025 models crossed a threshold from novelty to daily utility, shifting the center of gravity from training to inference usage at massive scale. He attributes sustained demand growth to AI adoption spreading across demographics and problem types.
Will frontier models commoditize like cloud—or segment like every other market?
The discussion shifts to whether AI model providers will become utilities. Feldman emphasizes market segmentation: hyperscalers win where security, software layers, and credibility matter, while ‘cheap compute’ buyers may prefer leaner providers without enterprise overhead.
Token economics and why compute gets cheaper (but speed still dominates)
Feldman forecasts ongoing reductions in cost per unit compute as architectures improve, even amid near-term supply constraints. He argues speed has compounding value—slow inference has ‘zero market’—and positions performance gains as decisive in competitive workflows like coding and agents.
Full-stack control: can Google become the lowest-cost token producer?
They examine the thesis that owning everything from silicon to power procurement makes Google the cheapest token supplier. Feldman notes the countervailing risk: if you only sell chips to yourself, you may lose volume benefits and constrain economies of scale—though Google is testing ways to broaden reach.
Cerebras’ differentiation: supply-chain advantages and ‘proof by benchmark’ moments
Feldman highlights how Cerebras avoids key GPU bottlenecks (HBM, CoWoS, oversubscribed nodes), turning industry constraints into opportunity. He also describes the strategic value of public benchmark-style proof points (e.g., Kimi K2 speed claims) to counter skepticism and win trust.
Scaling to massive customers: delivery muscle, concentration concerns, and giga-scale thinking
Feldman discusses what it takes to serve very large customers and why early wins build operational capability for subsequent deals. The conversation broadens into the changing mindset of infrastructure scale—from megawatts to multi-gigawatts—and whether electricity becomes the ultimate limiting factor.
Data centers vs local communities: delays are normal, but neighbor relations matter
Feldman downplays schedule slippage as inherent to large construction, but criticizes the industry for poor community engagement. He argues data centers can be clean, job-creating assets if builders are transparent, pay their own way on grid upgrades, and invest locally to earn legitimacy.
AI layoffs, tool spend, and the jobs that will emerge next
Feldman separates ‘AI-washed’ layoffs from true AI-driven displacement, arguing many cuts reflect earlier overhiring and ongoing automation. He expects software tool spend per engineer to rise dramatically (as in hardware EDA), while new governance roles will emerge as AI becomes core to enterprise operations.
The real enterprise blocker: lawyers, security, and open-source risk—especially from China
Feldman argues the main constraint on enterprise AI adoption is organizational risk management—legal and security teams incentivized to say ‘no’ when precedent is unclear. Open source intensifies concerns, particularly when leading models originate from Chinese firms, even as cost pressure pushes adoption forward.
Should US firms sell advanced chips to China—and what ‘chokepoints’ matter
Feldman takes a firm stance against selling leading-edge tech to China, arguing it will be used for military and industrial competition. He acknowledges counterarguments about keeping China in the ecosystem but believes the US and allies retain meaningful choke points via advanced manufacturing and tooling dependencies.
Europe’s innovation gap, Cerebras IPO timing, and leadership lessons from going public
Feldman argues Europe’s broader pattern—fear, regulation, and slower adoption—reduces breakout tech creation, though pockets thrive at the application layer. He describes Cerebras’ IPO as driven by persistence through regulatory obstacles (including CFIUS), and closes with reflections on leadership pressure, relationships, and empathy from boards and partners.
