NVIDIA GTC Conference Preview: Integrating Groq Technology to Launch a Major Offensive in Inference Chips, Samsung to Manufacture for the First Time, OpenAI May Become One of the First Customers

On Monday local time, the highly anticipated NVIDIA annual GTC developer conference will kick off in San Jose, California, USA. CEO Jensen Huang's keynote speech has always been regarded as an important barometer for the AI industry.For investors, the most noteworthy aspect of this conference is NVIDIA's strategic shift from training to inference and its adjustments in supply chain layout.According to information from media outlets such as The Information and The Deep View, this GTC may release three key signals:First, NVIDIA may leverage the integration of Groq technology to make a significant entry into the AI inference market;Second, the chip foundry segment may shift from TSMC to Samsung in the short term;Third, the ecosystem of robotics physical AI and open-source models is expected to further expand.Focusing on the Inference Market, Groq Chips Become Key LeverageThe current AI industry is gradually shifting from "training-first" to "inference-driven." In the training field, NVIDIA has established a solid technological and ecological advantage with its GPUs; however, in the inference market, competitors like Cerebras are vying for market share with faster and lower-cost solutions.In this context, the market is highly focused on NVIDIA's response strategy. According to The Information, Jensen Huang is expected to announce a new chip system that integrates NVIDIA and Groq technology at the conference. Behind this is NVIDIA's investment of approximately $20 billion last year to acquire the licensing rights for Groq technology.The chips developed by Groq are called LPU (Language Processing Unit), specifically optimized for inference workloads. This will also be the first time NVIDIA directly integrates another company's AI processor into its server rack system.Supply Chain Restructuring and First Batch of Major Clients ConfirmedThe manufacturing and commercial rollout of the new system is also a focal point for the capital markets.According to The Information, Groq LPU is expected to be produced by Samsung in the second half of this year. This arrangement is significant as it may mark the first time NVIDIA's server chips are manufactured by a foundry other than TSMC, breaking its long-standing reliance on a single supplier.However, according to sources cited by the aforementioned media, this change may be primarily temporary. As the next-generation LPU needs to be more closely integrated with NVIDIA's future AI chips, subsequent production may still return to TSMC.On the demand side, NVIDIA is expected to announce OpenAI as one of the first customers for this system. This chip system may be used to drive AI agents to perform tasks such as coding.Changes in Underlying Architecture and Future Technology RoadmapFor investors focused on chip technology, the architectural design of the NVIDIA-Groq system also reveals potential integration challenges and opportunities.According to The Information, the new rack structure will be significantly different from existing systems: each rack will be equipped with 256 Groq chips. Meanwhile, Intel processors will be responsible for communication management within the system. This design also suggests that NVIDIA's existing architecture has not yet fully integrated with the LPUHowever, NVIDIA clearly has longer-term plans. According to The Information, two informed sources involved in development revealed that the company is exploring deeper integration of LPU into its future product roadmap. One proposal is to merge the Groq processor with the next-generation Feynman GPU (the successor to the Rubin architecture) into a single chip to enhance performance and reduce overall costs.Physical AI and Expansion of Open Source Model EcosystemIn addition to computing power infrastructure, the AI application ecosystem is also a key highlight of the GTC conference.As the "AI Moore's Law" continues to advance—computing efficiency roughly doubling every four months—NVIDIA's layout in the fields of robotics and physical AI has garnered significant attention. Especially against the backdrop of the accelerated development of the humanoid robot industry in China, whether NVIDIA and its partners can provide more cost-effective solutions in scenarios such as autonomous driving will become a focal point for the market.At the same time, NVIDIA is also rapidly advancing in the field of open-source models. The company previously released the 120 billion parameter Nemotron 3 Super model and stated that it will launch the Nemotron 4 Ultra with a fourfold increase in parameter scale. The enhancement of model capabilities is expected to further reduce enterprise AI inference costs and improve overall return on investment.The signals released at this year's GTC may significantly influence the AI industry landscape in 2026