Nvidia Finalizes $20 Billion Licensing Deal with Groq, Showcasing LPU Technology at GTC 2026

Accelerated computing giant Nvidia confirmed a strategic licensing agreement with artificial intelligence (AI) chip startup Groq in December 2025, reportedly valued at US$20 billion. This significant transaction encompasses licensing of Groq's core technology and includes the acquisition of key talent through an acqui-hire mechanism, notably involving the company's founder, Jonathan Ross. The move signals a strategic pivot by Nvidia to bolster its standing in the increasingly competitive AI inference market, an area historically dominated by model training.

Groq, established by former Google engineers who contributed to the design of Google's first Tensor Processing Unit (TPU), concentrates on delivering rapid and cost-effective AI inference—the real-time application of trained models. This shift in focus aligns with a broader industry trend where major corporations are redirecting capital expenditure toward serving end-users directly. The AI inference chip market has seen explosive funding growth, escalating sixfold from US$822 million in 2022 to over US$5.2 billion in 2025, underscoring a global pursuit for alternatives to Nvidia's GPU dominance.

Competition in the inference segment is intensifying, with major players like Google, which announced its seventh-generation Ironwood chip in April 2025, and Amazon, which is also developing its proprietary accelerators. In a direct response, Nvidia is anticipated to launch new products leveraging Groq's Language Processing Unit (LPU) technology at the GPU Technology Conference (GTC) 2026 in San Jose, scheduled to commence on March 16, 2026. This launch is expected to be a central feature of CEO Jensen Huang's keynote address.

The primary innovation introduced by Groq's LPU is an architecture fundamentally distinct from GPUs, designed specifically for inference workloads. The anticipated new chip is expected to incorporate on-chip Static Random Access Memory (SRAM) instead of the High-Bandwidth Memory (HBM) utilized by traditional GPUs. This SRAM integration aims to achieve improved cost efficiency in inference, potentially offering speed enhancements up to ten times greater than existing solutions. The LPU architecture eliminates the need for round trips to external memory, which can consume hundreds of nanoseconds on GPUs, by accessing on-chip SRAM in a single clock cycle.

Jonathan Ross, Groq's founder and a key architect behind the Google TPU, previously demonstrated that the LPU could provide up to an 18-fold speed advantage for AI inference over conventional GPU solutions. Although Groq's LPU has a limited on-chip SRAM capacity of approximately 230 MB per chip, its deterministic architecture and static scheduling ensure efficient execution, effectively addressing the Memory Wall that constrains Batch-1 inference speeds on GPUs. Nvidia's move to acquire this technology—structurally resembling an acquisition despite no equity transfer—demonstrates the company's commitment to securing leadership in an era where infrastructure efficiency and inference capabilities are paramount in 2026. GTC 2026 will serve as the proving ground for how this LPU integration will strengthen Nvidia's technology stack amidst intense competition from hyperscalers building their own custom silicon.