Jonathan Ross on Groq: The $20B Nvidia Deal, Three Weeks from Bankruptcy, and Why Speed Is Intelligence
Three weeks from running out of money, Groq's engineers cut their salaries to $50,000. Two years later, Jonathan Ross closed a $20 billion deal with Nvidia - the largest in that company's history.
$20B deal, three weeks. Groq went to Nvidia wanting to buy 100K GPUs. Jensen countered with a partnership making GPU+LPU hybrid inference available to all Nvidia customers Jonathan @ 0:11.
Faster hardware = smarter model. The same AlphaGo model on TPU achieved 300+ ELO points higher than on GPU. Move 37 - the creative move that changed the match - was only reachable on faster hardware Jonathan @ 28:44.
80% took minimum wage to survive. With three weeks of runway left, Groq’s engineers cut to $50-60K from six-figure comp. Attrition stayed under 10% Jonathan @ 51:53.
Dismissed for years, vindicated in a day. Fast inference was worthless according to the market 3-4 years ago. A viral X video converted skeptics in hours Jonathan @ 42:26.
This breakdown is for paid subscribers. Below: the full deal structure and what it means for Nvidia’s inference roadmap, the AlphaGo proof that hardware speed raises model intelligence, and the operating principles that kept Groq alive through a decade of rejection. Subscribe to Podcast Alpha to read on.
