Nvidia Integrates $20B Groq IP for Low-Latency AI Inference
At its GTC 2026 conference, Nvidia launched the Groq LPX inference rack, the first major product emerging from its $20 billion intellectual property and team acquisition of Groq. The system integrates Groq's low-latency processing unit (LPU) architecture with Nvidia GPUs to optimize AI inference. Using a technique called "Attention FFN Disaggregation," the system delegates tasks based on hardware strengths: GPUs manage dynamic attention calculations, while the new LP30 LPU chips handle the static feed-forward network (FFN) computations, significantly reducing response delays in interactive AI applications.
Crucially for investors, the LP30 chip is manufactured on Samsung's SF4 process and does not use High Bandwidth Memory (HBM). This means the new LPX system represents incremental production capacity and revenue for Nvidia, as it does not consume the company's scarce TSMC N3 manufacturing slots or HBM supply, a key competitive advantage that cannot be easily replicated.
Vera ETL256 Rack Packs 256 CPUs to Break AI Bottlenecks
To address the growing CPU bottleneck in large-scale AI operations, Nvidia introduced the Vera ETL256. This high-density, liquid-cooled system packs 256 of its new Vera CPUs into a single rack. The design directly targets the massive parallel processing demands of tasks like data preparation and reinforcement learning, where CPU availability can limit overall GPU utilization. By integrating the compute density to a point where all intra-rack connections can be made with copper cabling, Nvidia eliminates the need for more expensive optical transceivers within the rack, offsetting the cost of liquid cooling.
Alongside the CPU rack, Nvidia unveiled the STX storage reference architecture. This standardizes the configuration of storage systems for AI, specifying the required combination of drives, Vera CPUs, BlueField DPUs, and networking components. Backed by major storage vendors including Dell, HPE, and IBM, the STX architecture solidifies Nvidia's expansion from compute and networking into the storage layer, a domain previously controlled by other companies.
Nvidia Aims to Capture Entire AI Infrastructure Market
The combined launch of the LPX, Vera ETL256, and STX systems signals a clear strategic pivot. Nvidia is moving aggressively to provide the entire AI infrastructure stack, creating a deeply integrated ecosystem that extends its market dominance. This platform strategy is already gaining traction, with partners like Cadence and HPE announcing new solutions built on Nvidia's latest hardware. The announcements provide a concrete roadmap for how Nvidia plans to achieve CEO Jensen Huang's forecast of securing $1 trillion in orders for its systems through 2027.
By systematically entering the CPU and storage markets, Nvidia is positioning itself to capture a much larger share of the total spending on AI hardware. This move intensifies competition for established players in these segments and reinforces Nvidia's role as the central provider for the AI industry's massive infrastructure buildout.
