The Biggest Chip Ever Built — Why OpenAI Runs On It | Cerebras CEO Andrew Feldman – The MAD Podcast with Matt Turck

The Biggest Chip Ever Built — Why OpenAI Runs On It | Cerebras CEO Andrew Feldman – The MAD Podcast with Matt Turck — Transcript

Market views

Bullish

CEREBRAS— Cerebras built the largest chip in computing history (58x larger than GPU) optimized for AI inference speed, solving memory bandwidth bottlenecks that plague GPUs. Secured historic $20B+ OpenAI deal for 750 megawatts across 2026-2028, demonstrating product-market fit as AI shifts from novelty to production workloads requiring fast tokens.
OPENAI— OpenAI demonstrated exceptional capacity planning foresight by securing massive deals for memory, compute, and silicon before constraints materialized. Their multi-silicon strategy (Cerebras deal, Jalapeno with Broadcom) shows sophisticated understanding of exponential AI adoption curves and willingness to invest ahead of demand.
TSMC— TSMC demonstrated exceptional partnership and risk-taking ability by agreeing in a single 2017 meeting to modify manufacturing processes for Cerebras' unprecedented wafer-scale chip. Their willingness to innovate with bold customers and maintain close collaboration positions them as the critical enabler of next-generation AI infrastructure.
AIINFRA— AI infrastructure demand is real and supply-constrained across memory (HBM), packaging (CoWoS), and data center capacity. Unlike historical bubbles where supply preceded demand, the industry is chasing existing usage with constraints everywhere - memory sold out, packaging bottlenecked, data centers limiting factor.
POWER— Power infrastructure is a critical bottleneck for AI data centers, with China making strategic investments in grid capacity while US lags. Cerebras building data centers in Nordic regions specifically for access to clean, low-cost power demonstrates power as the binding constraint for AI infrastructure expansion.
CPUS— Agentic AI is driving explosive CPU demand as AI systems increasingly take actions (website visits, data retrieval, ordering) rather than just providing answers. CPUs act as the 'body' executing instructions from AI 'brains', creating through-the-roof demand as AI becomes more agentic.

Bearish

NVDA— NVIDIA's CUDA moat is eroding rapidly - 70% of state-of-the-art models now train without CUDA (Gemini on TPUs, Claude on Tranium, OpenAI without CUDA). GPU architecture fundamentally struggles with fast inference due to HBM memory bottlenecks. NVIDIA's $20B Groq acquisition validates that GPUs cannot do fast inference and specialized chips are winning.
SAAS— Traditional SaaS dashboarding business model faces existential threat as AI can build custom tools instantly. The ability to ask AI to 'build me a tool like Salesforce' and get a working application in 30 seconds makes rigid, cross-silo SaaS products obsolete. Andrew describes this damage as 'unrepairable.'

Not investment advice.

Mentioned

Market views