# Groq Licenses LPU Architecture to Nvidia in $20B Deal

> In a surprising twist, Nvidia is licensing technology from a smaller rival. It paid roughly $20 billion for Groq's LPU inference architecture, a chip design built for one job: generating tokens as fast as possible. The Groq 3 LPU reaches about 1,500 tokens per second, leaning on 500MB of on-chip SRAM and 150 terabytes per second of internal bandwidth. The trade-offs are deliberate. An LPU gives up trainability and vision work in exchange for raw memory bandwidth and blistering autoregressive generation, which is exactly the profile that agentic AI, with its long chains of tool calls and reasoning steps, is hungry for. Nvidia plans to pair the design with its Vera Rubin systems in five new rack configurations shown at GTC 2026. Groq founders Jonathan Ross and Sunny Madra are moving to Nvidia, though Groq continues independently under new leadership. The deal is a tell about where inference is heading: as agents multiply, the speed and cost of producing tokens, not just training bigger models, becomes the battleground.

_Section: [Daily AI Updates](https://www.wortins.com/daily-ai) · Source: Spheron · Published Thursday, July 16, 2026_

## Wortins' read

In a surprising twist, Nvidia is licensing technology from a smaller rival. It paid roughly $20 billion for Groq's LPU inference architecture, a chip design built for one job: generating tokens as fast as possible. The Groq 3 LPU reaches about 1,500 tokens per second, leaning on 500MB of on-chip SRAM and 150 terabytes per second of internal bandwidth. The trade-offs are deliberate. An LPU gives up trainability and vision work in exchange for raw memory bandwidth and blistering autoregressive generation, which is exactly the profile that agentic AI, with its long chains of tool calls and reasoning steps, is hungry for. Nvidia plans to pair the design with its Vera Rubin systems in five new rack configurations shown at GTC 2026. Groq founders Jonathan Ross and Sunny Madra are moving to Nvidia, though Groq continues independently under new leadership. The deal is a tell about where inference is heading: as agents multiply, the speed and cost of producing tokens, not just training bigger models, becomes the battleground.

## Source

[Read the full story at Spheron](https://www.spheron.network/blog/nvidia-groq-3-lpu-explained/)

## Related coverage

- [Google Gemini 3.5 Pro targets July 17 launch after architectural rebuild](https://www.wortins.com/story/google-gemini-3-5-pro-targets-july-17-launch-after-architect-a33376a4) — [TechTimes](https://www.techtimes.com/articles/320308/20260713/gemini-35-pro-targets-july-17-after-full-rebuild-every-spec-remains-unconfirmed.htm)
- [A Script for Mark Zuckerberg: Meta's AI Bet and the Question of Staying Power](https://www.wortins.com/story/a-script-for-mark-zuckerberg-meta-s-ai-bet-and-the-question--84d71c6f) — [Stratechery](https://stratechery.com/2026/a-script-for-mark-zuckerberg/)
- [Figma Config 2026: AI agents, motion, generative shaders, code layers ship](https://www.wortins.com/story/figma-config-2026-ai-agents-motion-generative-shaders-code-l-5d002fa3) — [Figma](https://www.figma.com/blog/config-2026-recap/)
- [OpenAI proposes handing 5% stake to US government worth $42.6 billion](https://www.wortins.com/story/openai-proposes-handing-5-stake-to-us-government-worth-42-6--4a1dacdc) — [Axios](https://www.axios.com/2026/07/02/openai-stake-trump-altman)
- [Big-5 hyperscalers committing $725B to AI infrastructure capex in 2026, reshaping market](https://www.wortins.com/story/big-5-hyperscalers-committing-725b-to-ai-infrastructure-cape-cbc8f42b) — [Futurum Research](https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/)
- [Perplexity launches SPACE secure sandbox to let AI agents run safely at scale](https://www.wortins.com/story/perplexity-launches-space-secure-sandbox-to-let-ai-agents-ru-b2bfbad8) — [SiliconANGLE](https://siliconangle.com/2026/07/15/perplexity-launches-secure-sandbox-make-ai-agents-secure-powerful/)

---

_Curated and written by [Wortins](https://www.wortins.com) — The daily AI briefing. Every story links to its original source; the "Wortins read" on each is our own original analysis. [About Wortins & our editorial approach](https://www.wortins.com/about)._