Overview of 20VC with Andrew Feldman, CEO of Cerebras
In this wide-ranging conversation, Harry Stebbings speaks with Andrew Feldman shortly after Cerebras’ blockbuster IPO about the state of AI infrastructure, chip economics, data center constraints, geopolitics, enterprise adoption, and what the AI era means for companies, workers, and founders. Feldman argues that we are not in an infra bubble because AI demand is still outrunning supply, and he makes the case that the industry is only getting started on the road to cheaper, faster compute.
AI Infrastructure: Why Feldman Says This Is Not a Bubble
Demand is outrunning supply
- Feldman rejects the idea that AI infrastructure is like past bubbles in rail or fiber, where companies built too far ahead of demand.
- His core point: today’s AI buildout is behind demand, not ahead of it.
- Cerebras, NVIDIA, AMD, and others all have backlogs because data centers, power, and supply chain capacity cannot keep up.
Memory is a major bottleneck
- He says the industry is seeing memory shortages for years, not months.
- The shortage is driven by massive demand for HBM memory used in GPUs.
- Pricing has surged sharply, with memory makers earning unusually high margins because supply is so tight.
- Cerebras is relatively insulated because it uses SRAM, not HBM.
Data center buildouts are slow and lumpy
- Large-scale capacity cannot be added quickly:
- fabs can cost around $40B
- take years to build
- and expand in big step-changes, not increments
- Feldman argues that if demand stays elevated, shortages will persist for several more years.
Why Compute Demand Keeps Rising
AI became genuinely useful in 2025
- Feldman says a turning point happened when models became useful enough that people started using them daily.
- Demand is now being driven by:
- more users
- harder tasks
- more frequent usage
- broader adoption across age groups and industries
Speed matters enormously
- He stresses that for hard problems, there is essentially no ceiling on the value of faster inference.
- A system that solves problems in 3 minutes instead of 20 can create massive productivity and competitive advantage.
- He frames slow inference as almost worthless, likening it to slow internet: nobody wants it.
Cerebras, Token Economics, and Competitive Advantage
Cerebras’ performance edge
- Feldman says Cerebras is around 15x faster for certain workloads due to architectural advantages.
- He points to public demonstrations showing strong performance on models like Kimi K2, where Cerebras was reportedly 6.7x faster than the next-fastest GPU cloud.
Costs should keep falling over time
- He believes the entire industry will continue to improve:
- more tokens per unit time
- lower cost per unit compute
- better performance per watt and per dollar
- In his view, the long-term history of the chip industry is a massive reduction in compute cost.
Hyperscalers, Neoclouds, and Who Wins
Hyperscalers still have a strong value proposition
- AWS and Azure provide:
- security
- credibility
- software layers
- integrated data and tooling
- For many enterprises, that bundled offering is valuable enough to justify the premium.
But some customers want only cheap compute
- For others, the value is in raw cost and speed, not the full platform.
- Feldman sees the market as segmented rather than winner-take-all.
Thoughts on neoclouds
- He gives credit to companies like CoreWeave for innovative financial engineering and rapid deployment.
- But he implies many neocloud economics depend heavily on expensive underlying hardware and margins in the supply chain.
Google, Vertical Integration, and Whether They Should Sell Chips
- Feldman says full-stack ownership can lower costs, but there’s a tradeoff:
- if you only sell to yourself, your market is limited
- external sales can expand volume and potentially improve economics
- He notes that Google may already be moving beyond pure internal use for this reason.
Enterprise AI: Adoption, Security, and New Job Categories
The biggest blockers are lawyers and security teams
- Feldman says the main enterprise bottlenecks are not data cleanliness but:
- legal
- security
- procurement risk-aversion
- He argues these groups are naturally incentivized to say “no” and slow adoption.
Data organization matters
- Companies that have spent years structuring their data well will have a big advantage.
- He cites organizations like Mayo Clinic as examples of disciplined data builders.
AI will reshape roles
He expects new functions to emerge, similar to how CIO and CSO roles appeared in earlier tech transitions:
- chief AI officer / AI governance roles
- redefined HR functions
- fewer “information gatherer/presenter” roles
- more automation in support, admin, and coordination work
Layoffs are often “AI-washed”
- Feldman says many recent layoffs were really about:
- overhiring during COVID
- reorganizations
- productivity gains already underway
- Real AI-driven impact is starting now, especially in engineering-heavy organizations.
China, Fabs, and US Industrial Policy
On selling leading-edge chips to China
- Feldman is firmly against selling frontier chips to China.
- His reasoning:
- China’s military would use them
- Chinese industry would use them to compete more effectively against the US
- He favors keeping China on down-rev tech rather than supplying the latest generation.
Why US onshoring matters
- He argues the US needs not just fabs, but the surrounding ecosystem:
- packaging
- suppliers
- technical talent
- His ideal policy:
- give TSMC/Samsung a long runway in the US
- minimize local ordinance hurdles
- let them build advanced fabs quickly and predictably
Europe: The Worry is Structural
- Feldman thinks Europe has a broader innovation problem, not just an AI problem.
- His critique:
- too much regulation
- too much caution
- too little entrepreneurship and fast adoption
- He does note exceptions in places like London, Cambridge, Stockholm, and strong application-layer companies.
IPO, Perseverance, and Leadership Lessons
The IPO was deliberate, but timing mattered
- Cerebras’ IPO was years in the making.
- Feldman says it was delayed by CFIUS-related issues and later benefited from a better regulatory environment.
- The lesson: keep building through what you can’t control.
Leadership under pressure
- He emphasizes:
- resilience
- relentless iteration
- learning from repeated failure
- He describes a brutal 18-month period where Cerebras burned cash and still couldn’t solve key technical problems.
Personal reflections
- He says money changed him mainly by making him think bigger and worry less about downside.
- He’s proud that Cerebras has created wealth for many employees, saying the company has made 800 millionaires.
- On relationships, his advice is simple:
- choose a partner who understands entrepreneurship
- make the pressure visible
- invest time back into the relationship
Key Takeaways
- AI infra is still supply-constrained, not bubble-driven.
- Memory, power, data centers, and fabs are the major bottlenecks.
- Inference speed is becoming a decisive competitive advantage.
- Enterprise adoption is mostly blocked by legal/security friction, not model quality.
- China policy should favor restraint on frontier tech exports.
- The US needs a stronger domestic semiconductor ecosystem.
- Persistence matters more than perfect timing in company building.
Notable Insights
- “We’re not building ahead of demand. We’re building behind demand.”
- “For hard problems, there is no upper bound to how much faster you want to be.”
- “The biggest inhibitors to enterprise AI adoption are lawyers and security.”
- “The history of our industry is a massive reduction in the cost per unit compute.”