Episode Summary
Executive Summary: Andrew Feldman explains how Cerebras evolved from a contrarian wafer-scale AI chip bet into a fast-inference leader as AI usage surged in 2025. He argues that speed changes markets, enables new business models, and that Cerebras is now positioned for massive demand, partnerships with OpenAI and AWS, and future productivity jumps across software and enterprise work.
Main Topics: Cerebras’s evolution from AI hardware bet to inference leader (Priority: 5/5): Feldman recounts how Cerebras began as a new-architecture chip company for emerging AI workloads and transitioned into a company focused on ultra-fast inference for foundation models as AI became broadly useful. Why speed matters once AI becomes daily infrastructure (Priority: 5/5): He argues that when users rely on AI every day, latency becomes unacceptable and speed becomes a decisive product advantage, comparable to the demise of slow search or dial-up internet. Wafer-scale computing as a contrarian technical strategy (Priority: 5/5): Cerebras chose a radically different chip design: a dinner-plate-sized wafer-scale engine rather than a traditional postage-stamp chip, enabling major performance gains over GPUs. From technical proof to market pull (Priority: 4/5): Feldman describes years of building ahead of demand, followed by a demand explosion once AI models became useful enough for daily work, especially among code-generation and AI application companies. Scaling hardware, supply chain, and software stack (Priority: 4/5): The discussion covers manufacturing bottlenecks, battle-testing with early customers, and the long time required to build the compiler and software ecosystem needed to support scale. OpenAI, AWS, and strategic partnerships (Priority: 5/5): Major deals with OpenAI and AWS validated Cerebras’s positioning and required the company to rapidly expand capacity and supply chain readiness. Leadership, endurance, and when to quit (Priority: 4/5): Feldman reflects on the loneliness of long-term company building, the importance of accountable advisors, and the need to distinguish persistent conviction from stubbornness.
Key Arguments: A truly radical performance jump requires a fundamentally different architecture; incremental changes to GPUs would not deliver 15-20x gains. AI became economically important only once models were smart enough and useful enough for everyday work, which is why inference speed suddenly mattered in 2025. Cerebras’s wafer-scale approach was initially dismissed as impossible, but the company proved it could work and then improved it over time. Hardware companies cannot scale instantly like software companies; manufacturing, QA, and supply chain expansion create real-world timing constraints. Early customers in supercomputing, oil and gas, pharma, and sovereign AI helped bridge the gap between prototype success and mainstream demand. Open source has been important in sustaining the ecosystem and pushing closed-source leaders to keep improving. Fast AI will not just improve existing workflows; it will enable entirely new business models and reorganized ways of working. Founders should keep going only as long as their hypotheses for winning remain plausible; if all core assumptions fail, they should reassess or stop.
Data Points: Inference speed vs GPUs: 15-20x faster - Feldman says Cerebras is faster than GPUs across models and use cases. Cerebras chip size: 46,000 square millimeters - He describes Cerebras’s wafer-scale chip as the size of a dinner plate. Market cap at IPO context: about $63 billion - The intro notes Cerebras’s stock market value after going public. OpenAI deal size: north of $20 billion - Feldman describes the OpenAI agreement as one of Silicon Valley’s biggest deals. AWS agreement timing: signed in March - He says Cerebras later signed an agreement with AWS for deployment in its data centers. Failed development period: mid-2017 to mid-2019 - He recalls a two-year period when the company could not yet make the wafer-scale product work. Monthly spend during failure period: about $8 million per month - The company burned significant capital while debugging the architecture. Early product sales: a dozen, then 300 - He contrasts first- and second-generation sales before the market matured. Company size: 800-850 people - Feldman cites the current headcount when discussing the company’s scale. Software spend on tokens: from $1,000 per engineer to $25,000-$30,000 per engineer - He says AI coding usage internally has grown sharply over eight months. Manufacturing growth target: 10x this year - Feldman says Cerebras aims to expand manufacturing dramatically. Deal timeline: about 4.5 weeks - He says the OpenAI master agreement was signed quickly after initial term-sheet discussions.
Pivotal Quotes: "That's how big the market for slow inference will be." — Andrew Feldman: He compares slow AI inference to slow search and dial-up internet, arguing latency will eliminate demand. "To be radically better, you can't build something that is a similar architecture." — Andrew Feldman: He explains why Cerebras chose wafer-scale computing instead of a derivative GPU-like design. "We would much rather fail in pursuit of the extraordinary than succeed in the ordinary." — Andrew Feldman: He describes the company culture he wants to preserve as Cerebras scales.
Implications: The transcript suggests AI infrastructure winners will be those that deliver dramatic speed gains and can scale hardware fast enough to meet demand. It also implies AI will create new business models, not just better tools, and that companies should rethink product design, hiring, and persistence accordingly.