Cerebras Claims 20x Speed Advantage Over GPUs Across All AI Models
"Right now we're the fastest at inference, not by a little bit, but by a lot, 15, 18, 20x faster than GPUs. Faster across the board. Big models, small models, US models, Chinese models, trillion-parameter models, 1-billion-parameter models, across the board."
About this episode
On this episode of No Priors, hosts Elad Gil and Sarah Guo interview Andrew Feldman, co-founder and CEO of Cerebras, the AI chip company that recently went public at a $60 billion market cap. Feldman explains how Cerebras built a dinner-plate-sized wafer-scale chip that delivers inference speeds 15-20 times faster than GPUs across all model types, and why that advantage only became commercially valuable in 2025 when AI models reached sufficient utility for daily work integration. The conversation reveals the company's near-death experience burning $8 million monthly for two years while failing to manufacture the chip, a problem that had defeated the computer industry for 70 years. Feldman discloses that after mid-2025 conversations with Sam Altman, Cerebras closed a $20+ billion OpenAI deal in just 4.5 weeks over Thanksgiving, followed by an AWS partnership, creating backlog that now requires a 10x manufacturing scale-up. He shares internal data showing Cerebras engineers increased AI token spending from under $1,000 to $25,000-$30,000 per person in eight months, with elite engineers achieving 100x productivity through AI agent workflows. Feldman argues the AI revolution mirrors earlier infrastructure shifts where speed didn't just improve existing workflows but enabled entirely new business models, comparing it to Netflix's transformation from DVD delivery to movie studio. The episode explores the psychology of contrarian entrepreneurship, with Feldman describing himself as a professional David competing against Goliath across five startups, and offering advice on when founders should persist versus give up on failing ventures.
Key takeaways
- Cerebras signed a $20+ billion OpenAI deal in 4.5 weeks over Thanksgiving 2025, working seven days a week with multiple law firms to close by December 24.
- Feldman revealed engineers now spend $25,000-$30,000 annually on AI tokens, up from under $1,000 eight months ago, with top performers achieving 100x productivity using AI agents.
- Cerebras claims 15-20x faster inference than GPUs across all model types and sizes, an advantage that became commercially relevant only when AI reached daily-use utility in 2025.
- The company burned $8 million monthly for two years failing to build its wafer-scale chip before achieving success in summer 2019, solving a 70-year-old industry problem.
- Cerebras is attempting a 10x manufacturing scale-up in 2025, which Feldman claims would be the fastest in hardware history, enabled by earlier $1 billion G42 strategic partnership.
- Feldman argues fast AI will create entirely new business models rather than incremental improvements, comparing it to Netflix transforming from DVD delivery to movie studio when internet speeds increased.
- The company went public to exchange venture capital for public market investors at lower cost of capital and gain legitimacy as the only pure-play AI infrastructure stock.