As the tech world prepares for GTC 2026, excitement is building around the anticipated advancements in artificial intelligence and computing. Notably, this year’s event is poised to feature a pivotal shift in the perception and application of computing technologies, particularly in the realm of AI infrastructure.
Recent years have seen an intensifying race among major companies like NVIDIA and AMD to innovate their computing offerings in response to evolving AI workloads. Since 2022, the focus has significantly transitioned toward training workloads, a trend that NVIDIA capitalized on with its Hopper and Blackwell architectures. However, as we look ahead to 2026, the spotlight is shifting to agentic workloads, which are expected to dominate the conversation at GTC.
Moreover, a notable collaboration is on the horizon as NVIDIA looks set to formalize its partnership with Groq. This agreement marks a strategic departure from the traditional GPU-centric computing model. Anticipations suggest that NVIDIA will blend Groq’s LPU (Logical Processing Unit) technology with its Vera Rubin systems, potentially leading to a hybrid compute configuration that enhances NVIDIA’s capacity for disaggregated inference.
Details surrounding the integration of LPUs into the Rubin architecture remain speculative, but experts expect a range of configurations—64, 128, and 256-unit setups—linked to Rubin GPUs via NVLink Fusion. NVIDIA CEO Jensen Huang has previously indicated that the Groq partnership could play a transformative role akin to that of Mellanox, highlighting the anticipated improvements in managing various workload stages, particularly decoding tasks. The introduction of the Rubin CPX has already enabled the company to address significant segments of traditional inference requests.
In addition to these advancements, GTC 2026 is expected to offer a comprehensive look at NVIDIA’s forthcoming Feynman AI chips. Initially hinted at during GTC 2025, the Feynman architecture aims to leverage Moore’s Law for scaling compute capabilities, reportedly utilizing TSMC’s A16 process. Introduced alongside new design features, the Feynman chips may also incorporate Groq’s LPUs, marking an evolution in NVIDIA’s approach to microarchitecture.
As for existing systems, NVIDIA is not yet done with its Vera Rubin lineup. The company demonstrated the NVL72—featuring a 72-chip configuration—at CES 2026, with plans to expand this series further. While scaling up to NVL144 and NVL576, reports suggest that the former may not reach the market due to customer compute needs. The Rubin CPX targets prefill applications, although details on customer deployments have been limited.
Eyes will particularly be on the NVL576, which signals a transition to the “Kyber” generation, featuring vertical stacking of compute trays akin to organized shelves. This design is expected to support a new power delivery model, enhancing efficiency and thermal management for larger-scale configurations. NVIDIA’s forthcoming CPO (Co-Packaged Optics) technology aims to further revolutionize interconnects, promising significant improvements in throughput and latency while reducing reliance on copper cabling.
While the roadmap anticipates various innovative announcements, including potential collaborations with Intel, it is the integration of the Rubin and Rubin Ultra architectures that will likely dominate discussions leading up to the event. The anticipation builds for what could be groundbreaking demonstrations at GTC 2026, as NVIDIA prepares to unveil transformative solutions for the future of AI-driven computing.
