Nvidia Kindacquires Groq

What It Covers

Nvidia acquires Groq's technology and 90% of staff through a $20 billion licensing deal to access SRAM-based chip architecture for ultra-low latency AI inference and counter Google's TPU success.

Key Questions Answered

• •SRAM Architecture Strategy: Nvidia pays $20 billion for Groq's SRAM chip design that keeps data on-processor, minimizing reliance on high-bandwidth memory chips controlled by a handful of suppliers and enabling faster token-per-second metrics for AI workloads.
• •Hackquisition Structure: The licensing deal allows Nvidia to bypass regulatory scrutiny that would block a traditional acquisition given their 90% AI chip market share, while paying shareholders at $20 billion valuation and hiring CEO Jonathan Ross who created Google's TPU.
• •Inference Disaggregation: AI inference splits into prefill and decode operations, with SRAM architectures offering unique advantages in decode for ultra-low latency agentic reasoning workloads, though at higher cost per token due to smaller batch sizes that users prove willing to pay.
• •Robot Deployment Economics: For every $100 spent deploying robots today, only $20 goes to the actual machine while $80 covers safety equipment and systems to protect humans, making installation costs the biggest barrier to adoption according to McKinsey partner surveys.