Behind the $20 Billion Non-Exclusive Deal: NVIDIA Acquires Groq's Core Team to Reduce Reliance on HBM

Behind the $20 billion non-exclusive licensing agreement, NVIDIA's acquisition of Groq's core team is widely seen as a strategic move to reduce its reliance on High Bandwidth Memory (HBM).

Recently, NVIDIA struck a non-exclusive licensing agreement worth up to $20 billion with Groq, an AI chip startup, covering the latter's core technologies in the inference chip segment. As part of the deal, Groq's founding team including founder Jonathan Ross and President Sunny Madra will join NVIDIA, further strengthening NVIDIA's layout in inference architecture research and development.

Notably, the core target of this mega deal is Groq's self-developed Language Processing Unit (LPU) chip. Unveiled in 2024, the LPU is the world’s first inference chip fully integrated with on-chip SRAM, eliminating the need for external HBM memory. With SRAM delivering read and write speeds up to 100 times faster than HBM, data access and computing are processed entirely on the chip, which theoretically eliminates latency issues caused by external memory fetching.

Powered by Groq's proprietary TSA architecture, the LPU features computing paths precisely pre-programmed by software and compilers, enabling deterministic computing — all data processing stages are perfectly synchronized in time, forming a seamless, zero-wait high-efficiency pipeline.

Meanwhile, global demand for HBM continues to surge rapidly. Intensified competition from chips such as Google's TPU, AMD's MI400 series and AWS Trainium has widened the global HBM supply gap. At present, all HBM production capacity of major suppliers for 2026 has been fully booked, and the industry is closely monitoring the capacity allocation schedule for 2027. Against this backdrop, HBM manufacturers have seen a steady rise in their bargaining power.

For NVIDIA, the adoption of Groq's LPU technology means it can ease its dependence on the HBM supply chain for lightweight inference scenarios. This not only helps secure its computing power delivery and revenue stability in the inference chip market, but also provides a critical alternative technical pathway to address potential HBM supply constraints in the future.