Nvidia intends to introduce a new processor platform optimized for AI inference at its GTC developer conference in San Jose during March 2026. The system incorporates technology from Groq's Language Processing Unit (LPU) architecture under a $20 billion non-exclusive licensing agreement finalized on December 24, 2025. This follows Nvidia's largest transaction to date, which included licensing Groq's inference technology and hiring key personnel such as founder Jonathan Ross and President Sunny Madra. The announcement arrives amid Nvidia's stock declining approximately 7 percent over two trading sessions despite record quarterly earnings. OpenAI has committed to 3GW of dedicated inference capacity using the new platform, positioning it as a primary customer. The platform targets faster, more efficient query responses in deployed AI models, addressing the shift from training to inference workloads.
Background on the Nvidia-Groq Transaction
The December 24, 2025, agreement structured as a non-exclusive perpetual license granted Nvidia access to Groq's full patent portfolio and software stack for inference optimization. Nvidia acquired physical assets and transferred roughly 80 to 90 percent of Groq's workforce, including core engineering teams. Groq continues operations independently with its GroqCloud service under new leadership. This transaction marked Nvidia's entry into specialized inference hardware beyond traditional GPUs, following Groq's prior valuation of $6.9 billion after a $750 million round in September 2025.
Groq's LPU Architecture and Inference Focus
Groq's Language Processing Unit employs deterministic execution with large on-chip SRAM, hundreds of megabits per chip, to eliminate bandwidth bottlenecks common in GPU-based inference. Demonstrations showed 10,000 thought tokens generated in approximately 2 seconds. The architecture optimizes for sequential inference workloads, reduces latency compared to GPU architectures reliant on high-bandwidth memory, and is designed for high-throughput, low-cost query processing in deployed models. Nvidia's integration positions the technology as an “accelerator” similar to its Mellanox acquisition role.
Planned GTC 2026 Unveil: LPX Inference Platform Details
The new platform, referred to as LPX in industry analyses, builds on Groq's LPU for dedicated inference racks. Reported specifications include initial racks with 64 LPUs packaged as 32 RealScale ASIC tiles, scaled versions targeting 256 LPUs per rack (a 4x increase), and focus on low-latency decode and agentic AI tasks requiring real-time responses. OpenAI's 3GW allocation underscores projected demand for the system.
Market and Competitive Context for Inference Hardware
Inference has emerged as a distinct market segment from training, with demand for specialized chips rising due to high GPU costs and power consumption. Prior to the deal, OpenAI explored options with Cerebras and Groq directly. The transaction neutralized a competitor while accelerating Nvidia's inference portfolio. Other inference-focused players include Cerebras, SambaNova, and D-Matrix.
When a $20 billion Christmas Eve license turns GPUs into yesterday's news for tomorrow's queries, the only surprise left at GTC might be how quietly inference rewrote the architecture rules.