Cerebras CEO on the Future of Data Centres, Token Costs & Memory | Should US Companies Sell to China
TL;DR
Cerebras CEO Andrew Feldman argues AI infrastructure is not a bubble because supply is struggling to catch up with explosive demand, evidenced by $25 billion in data center backlogs and severe memory shortages. Cerebras' recent IPO reflects this reality, with the company positioned advantageously using SRAM instead of scarce HBM memory while industry-wide compute costs continue their historical decline.
🏗️ The 'No Bubble' Infrastructure Reality 3 insights
Infrastructure chases demand rather than leading it
Unlike fiber optic or railroad bubbles where infrastructure preceded demand, AI data centers face a $25 billion backlog because supply cannot keep up with current needs.
Data center metering prevents market oversupply
Construction delays and permitting challenges act as necessary traffic meters that smooth demand and prevent the market from gorging on excess capacity too quickly.
2025 marked the usefulness inflection point
AI models crossed a threshold in early 2025 to become genuinely useful tools rather than novelties, driving exponential demand across all demographics from Silicon Valley to 85-year-olds.
⚡ Supply Chain Constraints & Cerebras' Advantage 3 insights
HBM memory shortage will persist for years
With only three suppliers (Samsung, Micron, Hynix) enjoying 80-85% gross margins, high-bandwidth memory shortages will continue as building new fab capacity requires $40 billion and five years.
Cerebras avoids critical supply bottlenecks
Unlike GPUs, Cerebras chips use SRAM etched by TSMC during logic manufacturing, avoiding both the HBM shortage and CoWoS packaging constraints while utilizing available 5nm capacity.
GPU prices skyrocketing amid component scarcity
Supply constraints have driven GPU prices through the roof, while Cerebras maintains cost advantages by not paying premiums for HBM or advanced packaging.
🎯 Market Strategy & Competitive Dynamics 3 insights
Nvidia's Neo cloud strategy creates dependence
Nvidia has funded and over-allocated to Neo clouds to create hyperscaler competitors, fostering an unhealthy dependence while hyperscalers retain advantages in security and software ecosystems.
Vertical integration limits hardware market size
Google's full-stack ownership from TPU to data center constrains their hardware market to internal demand only, historically limiting volume and cost reduction compared to merchant silicon vendors.
OpenAI forced into disadvantageous supply deals
Facing supply constraints, OpenAI purchased down-revision H100s from Elon Musk rather than current-generation B200s, leaving them one to two generations behind despite early contracting efforts.
📉 The Trajectory of Compute Costs 2 insights
Cost per compute will continue historic decline
Despite current shortages, the industry will see massive reductions in cost per unit compute over three to four years as all chipmakers improve designs to deliver more tokens per dollar and watt.
Performance gap expected to widen significantly
Cerebras currently delivers 15x faster performance through architecture alone and expects this advantage over GPUs to increase as designs evolve.
Bottom Line
The AI infrastructure buildout is supply-constrained rather than a demand bubble, meaning companies that secure compute capacity now—even if not the latest generation—gain competitive advantage, while specialized architectures that bypass HBM and CoWoS bottlenecks are positioned to capture disproportionate value during the shortage.
More from 20VC with Harry Stebbings
View all
Now is the Time for the App Layer | OpenAI & Anthropic Won't Win the App Layer | Mike Mignano, USV
Mike Mignano argues that the AI infrastructure buildout is complete, making now the ideal time to build at the application layer, while predicting the model landscape will either consolidate through recursive self-improvement or commoditize into an S-curve plateau favoring open weights and cost optimization.
The $100,000 token budget EVERY engineer will need | Sierra Co-Founder
Sierra Co-Founder Clay Bavor explains why the future of enterprise AI involves mixing frontier and open-weight models, predicts unbounded demand for frontier intelligence despite cost pressures, and reveals how AI-native engineering teams are achieving 3-20x productivity gains.
Coinbase Cuts AI Spend by 50% | Kalshi's $40B Valuation & Impending IPO | The Year for SaaS Roll-Ups
Coinbase's reduction of AI spend by 50% through open-source adoption signals a broader enterprise shift from experimental 'token maxing' to rigorous cost discipline, raising urgent questions about the revenue sustainability of frontier AI models facing cheaper commoditized alternatives.
Bloom Energy CEO: Why Electricity, Not AI Models, Will Decide the Winners of the AI Race
Bloom Energy CEO KR Sridhar argues that electricity infrastructure, not AI models, will determine the winners of the AI revolution, sharing how his NASA background and Andy Grove's 'walk the floor' mentorship shaped his 25-year mission to power the digital age.