Cerebras CEO Rejects AI Doom Talk, Points to Faster Inference
Cerebras Systems CEO and co-founder Andrew Feldman rejected AI doomsday warnings at TechCrunch Disrupt 2026. Speaking in a session titled 'Can AI Keep Scaling?', he said some AI leaders "maybe shouldn't be allowed to talk in public" about doom scenarios. TechCrunch
He offered a different future. Feldman said AI "could lead to a world where your grandkids will never meet anybody who dies of cancer." He linked that idea to large models as tools for biology research and clinical progress. TechCrunch
Feldman did not argue for building without limits. He said "powerful technologies come with the challenge of what happens when they are in the hands of bad people." He added that AI technology "is powerful, must be approached thoughtfully and regulated reasonably." TechCrunch
The business context is inference. Inference is the step when a trained model handles live requests. Cerebras lists customers including OpenAI, Block, Figma, GSK and Lovable. That mix spans frontier-model serving, fintech, design tools, pharma and AI tools that generate code. In each case, low delay matters.
The company describes the Cerebras CS-4 as a rack-scale system, a full rack that acts as one machine, that delivers up to 30x faster inference than GPUs. OpenAI's GPT-5.6 Sol is available in Ultrafast mode running on Cerebras at up to 750 tokens per second, tokens being small chunks of words used to measure output speed. Cerebras has partnered with Lovable to power faster AI software creation and with AMD on disaggregated inference, a design that keeps compute, memory and networking separate, combining AMD Helios and the Cerebras Wafer-Scale Engine, a chip built at wafer size. Cerebras
Feldman also turned to siting and power. He said "the industry was not good at working with local communities in the U.S. on AI-driven data center buildouts." He named Brazil, Canada, Finland, Iceland, Norway, Mexico and Sweden as markets encouraging companies to build data centers.
TechCrunch Disrupt 2026 featured six editorially curated stages. Feldman's scaling session was part of the enterprise and infrastructure programming announced on Sept. 30.
The broader context here is that two debates are now joined. The first is who gets access to strong models and under what guardrails. The second is where to place capacity, how to power it, and how to keep local consent. Both now shape roadmaps as much as silicon or memory bandwidth.
In my view, optimism about cures and caution about misuse can exist together. We have seen this pattern before, when the PC, the commercial internet, mobile and cloud each widened both helpful and harmful uses. What changed each time was how quickly engineering and policy teams learned to manage the tradeoffs. For inference builders, pressure has shifted from training size to serving efficiency, to token throughput, time to first token, and cost per million tokens under steady load. Rack-scale wafer systems and split-apart designs are answers to that pressure. So is building where grids and permitting are clear. The local issues are operational, covering noise, water, connection queues and rate effects, and they call for early disclosure and long-term presence. Jurisdictions that provide that clarity will attract builds and make faster products and research possible.
