Analyst Ming-Chi Kuo of TF International Securities indicated that NVIDIA's increased investment in AI inference architecture firm Groq has led to a significant upward revision in Groq's LPU shipment forecasts. The AI inference hardware market is expected to experience exponential growth over the next two years. According to the report, cumulative LPU shipments between 2026 and 2027 are projected to reach approximately 4 to 5 million units. Shipments in 2026 are estimated to account for 30% to 40% of the total, rising to 60% to 70% in 2027. Compared to previous annual shipment volumes, this represents a more than tenfold increase, signaling that AI inference demand is rapidly transitioning into large-scale deployment phases.
In terms of hardware architecture, NVIDIA plans to substantially increase the number of LPUs per AI cabinet from the current 64 units to 256 units. This expansion aims to enhance memory capacity while maintaining ultra-low latency inference performance. The move is viewed as a critical strategy to address the growing demand for generative AI and large language model inference. Supply chain sources indicate that the next-generation architecture cabinets are expected to enter mass production between the fourth quarter of 2026 and the first quarter of 2027.
Regarding corresponding shipment scales, cabinet shipments in 2026 are projected to be around 300 to 500 units, with a significant surge anticipated in 2027 to 15,000 to 20,000 units, reflecting explosive growth. The market widely believes that as AI applications shift from training phases to large-scale inference deployment, specialized inference chips featuring low latency and high throughput will become central to data center upgrades. The collaboration between NVIDIA and Groq is also regarded as a key indicator of AI infrastructure entering a new cycle.
Comments