Nvidia Groq $20B Inference Deal | LPX Platform, 3GW OpenAI Commitment at GTC 2026

What Happened | The $20 Billion Christmas Eve Deal

Nvidia finalized a $20 billion non-exclusive perpetual license with Groq on December 24, 2025, gaining access to Groq's full patent portfolio and software stack for inference optimization. The agreement is structured as a license, not an acquisition, meaning Groq continues to operate independently under new leadership with its GroqCloud service.

The deal included transfer of physical assets and roughly 80 to 90 percent of Groq's workforce, including core engineering teams. Key personnel who moved to Nvidia include founder Jonathan Ross and President Sunny Madra. Groq had reached a $6.9 billion valuation after a $750 million funding round in September 2025, just three months before the deal closed.

What Is Groq's LPU | Why Nvidia Paid $20 Billion for It

Groq's Language Processing Unit (LPU) employs deterministic execution with large on-chip SRAM, hundreds of megabits per chip, to eliminate the bandwidth bottlenecks common in GPU-based inference. Unlike GPUs, which rely on high-bandwidth memory (HBM) that creates latency during sequential token generation, the LPU keeps all active data on-die.

In public demonstrations prior to the deal, Groq showed 10,000 thought tokens generated in approximately 2 seconds, a throughput rate that outpaced GPU-based inference by an order of magnitude for sequential decode workloads. The architecture is purpose-built for the kind of token-by-token generation that dominates deployed AI applications: chatbots, code assistants, search, and agentic AI systems requiring real-time responses.

LPX Inference Platform | What Nvidia Will Unveil at GTC 2026

The new platform, referred to as LPX in industry analyses, builds on Groq's LPU for dedicated inference racks. Nvidia is expected to reveal the full specifications at its GTC developer conference in San Jose during March 2026.

The LPX platform is positioned as complementary to Nvidia's existing Blackwell B300 GPU racks , which dominate training workloads. Jensen Huang has described the strategy as offering “the right silicon for the right workload”: Blackwell for training and prefill, LPX for decode and real-time inference. This mirrors the approach Nvidia took after acquiring Mellanox in 2020 for networking, where the acquired technology became an accelerator within the broader Nvidia ecosystem rather than a standalone product line.

OpenAI's 3GW Commitment | Why It Matters

OpenAI has committed to 3 gigawatts of dedicated inference capacity using the LPX platform, positioning itself as the anchor customer. To contextualize that number: 3GW is roughly equivalent to the power output of three nuclear power plants or the total electricity consumption of a city of 2 million people.

Prior to the Nvidia-Groq deal, OpenAI had been exploring inference alternatives with both Cerebras and Groq directly. The deal effectively channeled those relationships through Nvidia, giving OpenAI a single vendor for both training (Blackwell) and inference (LPX) hardware.

Deal Timeline | From Groq Funding to GTC Unveil

Nvidia's Stock Reaction | Why Markets Sold Off Despite Record Earnings

Nvidia's stock declined approximately 7 percent over two trading sessions following the deal announcement, despite the company reporting record quarterly earnings in the same period. The market reaction reflected several concerns.

Competitive Landscape | Who Else Builds Inference Hardware

The Nvidia-Groq deal neutralized one of the most promising inference-focused competitors, but several others remain. The inference hardware market is fragmenting as the AI industry recognizes that training and inference require fundamentally different chip architectures.

Training vs Inference | Why the AI Industry Is Splitting

The AI hardware market is undergoing a structural shift. For the first five years of the deep learning era (2018-2023), nearly all compute spending went to training: building models. But as models reach production scale, inference, running those models to answer queries, has become the dominant cost center.

This is why Nvidia paid $20 billion for Groq's technology: the company that dominates training (Nvidia holds >80% GPU market share) recognized it needed a fundamentally different architecture to win the inference side. The $4 billion photonics deals with Lumentum and Coherent address the networking bottleneck; the Groq deal addresses the compute bottleneck for inference specifically.