Andrew Feldman, founder and CEO of Cerebras Systems, joins the podcast fresh from the company's $5.5 billion IPO to discuss their revolutionary wafer-scale chip technology. Cerebras has developed dinner plate-sized chips that are 58 times larger than traditional processors, enabling dramatically faster AI inference.

The conversation covers Cerebras's technical breakthrough in using fast memory instead of slow HBM, their $20 billion OpenAI contract, and the economics of AI inference. Feldman explains how their giant wafers avoid key supply chain bottlenecks while delivering 15x performance improvements over GPUs.

The discussion explores the competitive dynamics between open source and closed source AI models, the role of speed in inference economics, and the challenges of scaling semiconductor manufacturing in the US under the CHIPS Act.

Giant Wafers Enable 15x Faster AI Inference

Cerebras builds wafer-scale chips that are 58 times larger than any other chip ever made, roughly the size of a dinner plate, enabling the use of fast memory instead of slow HBM memory that GPUs rely on.

"We're fifteen times faster than the fastest GPU. On some problems we're fifty, one hundred, even one thousand times faster than graphics processing units" - Andrew, explaining the performance advantage of their architecture.

The technical breakthrough required solving fundamental problems across lithography, materials, packaging, power delivery, and cooling that had never been accomplished in the 75-year history of the industry.

Previous wafer-scale attempts, including Gene Amdahl's Trilogy in the mid-1980s, had all failed spectacularly, making Cerebras the first successful implementation.

Open Source Models Deliver Better Economics

Open source AI models cost significantly less than closed source alternatives while delivering only 3-5% less capability, creating a major economic advantage for users.

"What you're not paying for was the cost to train it" - Andrew, explaining why open source models like Kimmy Kate 2 are dramatically cheaper to run than proprietary alternatives.

Two of the three leading frontier models (Gemini and Claude) now use no CUDA, representing a 70% market share loss for Nvidia's software stack in training.

Companies with unique datasets like Mayo Clinic and GlaxoSmithKline can extract value without sharing proprietary information with foundation model makers.

Speed Commands Premium Pricing in AI Markets

Anthropic offered tokens twice as fast for six times the price and sold out completely, demonstrating strong demand for speed despite higher costs.

"How big is the market for slow search? Zero. How big is the market for dial up internet? Zero. Nobody wants to wait" - Andrew, on why speed matters across all AI applications.

GPUs become extremely expensive and power-hungry when trying to achieve useful speeds, while Cerebras makes fast tokens at vastly lower cost and power consumption.

Speed advantages compound over time in competitive scenarios where one party can complete three units of work while competitors finish one.

Supply Chain Advantages and Market Constraints

Cerebras avoids the three major supply constraints limiting AI chip vendors: HBM memory shortages, TSMC's CoWoS packaging bottleneck, and 3nm fab capacity limits.

"Business today is constrained by data centers... what are we all constrained by? Buildings. The data centers right now are everybody's constrained" - Andrew, identifying real estate as the primary growth limitation.

The company uses 5nm processes at TSMC instead of the heavily constrained 3nm node, providing better supply chain access.

Data center capacity constraints will persist for 15-18 months across the entire industry, limiting all players regardless of chip technology.

Major Contracts and Cloud Strategy

Cerebras signed a $20 billion contract with OpenAI in December 2024, representing one of the largest deals ever signed in Silicon Valley.

AWS will offer Cerebras inference through their Bedrock service using a disaggregated solution combining Trainium and CS-3 systems for ultra-fast performance.

G42, the UAE's national AI champion, was a major customer and minority investor, using Cerebras for English-Arabic models, genomics work, and cloud services across universities and companies like ADNOC.

Moving from GPU-based models to Cerebras requires only "ten keystrokes" through their API, eliminating software migration barriers.

Export Controls and Geopolitical Considerations

Feldman believes limiting distribution of precious technologies to industrial competitors makes sense, despite foreclosing some markets to US companies.

"China is an industrial enemy" and export controls represent "an extraordinarily difficult job" requiring balance between security and commercial interests - Andrew.

The company resolved all CFIUS national security issues in March 2025, before taking investment from Seventeen Eighty Nine Capital associated with Donald Trump Jr.

Cerebras works with "right leaning and left leaning investors" but has never asked for political access or favors from any political connections.

Resources Mentioned

Silent Spring

Rachel Carson (1962) was discussed as an example of early environmental warning, where her predictions about pesticides were initially dismissed but later vindicated, illustrating how scientific warnings can be overlooked before being proven correct.