NVIDIA's annual tech extravaganza, GTC 2026, will be held in San Jose, California, with the entire tech world's attention once again focused on CEO Jensen Huang's keynote address. Slogans like "It All Starts Here" are visible throughout the venue, foreshadowing the event's prestigious status. According to foreign investment analysis and market dynamics, this year's GTC will no longer be limited to GPUs themselves, but will comprehensively expand to next-generation chip architecture roadmaps (from Rubin to Feynman), CPO (Common Packaging Optics) networking technology that breaks through transmission bottlenecks, and...A new lineup of AI inference products integrating Groq team technology.
Chip roadmap covering 2028: Rubin and Feynman's generational shift
With the demand for AI computing power continuing to surge, the industry is most concerned about NVIDIA's product update schedule.
Following Blackwell Ultra, Jensen Huang had already made initial announcements at CES 2026 in January of this year.Chips using the Rubin architectureAccording to the current plan, Rubin expects to...Officially launched in the second half of this year (2026).Then, the second half of 2027 will see...Rubin UltraAnd in 2028, it will be upgraded to the new Feynman architecture.
Bank of America (BofA) predicts that NVIDIA is highly likely to fully outline this ultra-clear product roadmap spanning three generations (Rubin to Feynman) at this year's GTC. This long-term product visibility not only provides developers and enterprise customers with a strong commitment but is also a key strategy for NVIDIA to outpace its competitors.
Network infrastructure revolution: CPO and Spectrum-X become key players
As AI model parameters expand without limit, the "data transfer speed" between GPUs has become the biggest bottleneck besides computing power.
At CES 2026, NVIDIA showcased six Rubin-generation chips, including CPUs and GPUs, as well as the NVLink 6 switch chip, the ConnectX-9 Spectrum-X super network card, the BlueField-4 DPU, and the Spectrum-6 Ethernet switch, strongly suggesting that network and storage layout will be the highlight of this year's GTC.
Coincidentally, this year's Optical Communications Conference (OFC) was held almost concurrently with GTC. Market analysts and institutions such as Guolian Securities and Minsheng Securities generally believe that NVIDIA has built a global CPO supply chain covering core chips, optoelectronic components, and packaging. Investors are highly anticipating NVIDIA's announcement of technical details for scale-out network architectures, particularly the next-generation 102.4T Spectrum-6 switch and the 115T Quantum-X CPO component, which will be the core infrastructure for future large-scale AI cluster deployments.
By adding CPO components to its own ecosystem, it will also further enhance...BroadcomThis poses a greater market threat and strengthens NVIDIA's competitive advantage in proprietary technologies.
Filling the gaps in the inference: A new ASIC integrating Groq technology
Besides GPUs, NVIDIA's moves in the field of Application Specific Integrated Circuits (ASICs) have also sparked heated discussions.
Last December, NVIDIA reached a non-exclusive IP agreement with Groq, an AI chip startup known for its ultra-fast inference capabilities, and core team members, including founder Jonathan Ross, also moved to NVIDIA. Groq's LPU (Language Processing Unit) uses SRAM memory to process token generation and is considered one of the best solutions to the latency problem in real-time AI applications.
Recently, there have been widespread rumors that OpenAI is preparing to adopt NVIDIA's AI inference chip based on Groq technology. Analysts expect that NVIDIA will launch a "portfolio of inference-specific products" at GTC 2026, including CPX chips and low-latency LPUs. These customized chips are very likely to be directly integrated into NVIDIA's next-generation rack systems, further consolidating its dominant position in the AI inference market.
Physical AI Deployment: Self-driving Cars and Robots
On the eve of the conference, NVIDIA released a 22-minute demonstration video. In the video, Jensen Huang and Xinzhou Wu, NVIDIA's Vice President of Automotive Business, rode together in a self-driving car equipped with the NVIDIA DRIVE AV full-end autonomous driving software platform, navigating the complex streets of San Francisco without any human intervention. This demonstration foreshadows that, in addition to data center business, "physical AI" combining sensing, computing, and control (including autonomous driving and general robotics) will be another eye-catching focus of the GTC 2026 keynote address.
Update: Intel announces participation in GTC 2026, potentially showcasing next-generation data center solutions.
Intel has confirmed its participation in GTC 2026, where it may announce its new server processor products.
Due to NVIDIA's announcement last yearInvest $50 billion in IntelFurthermore, Intel provides NVLink technology to help it build multi-generation PC and data center CPUs. Therefore, it is expected that Intel will explain its new server processor at GTC 2026, and also see the next-generation data center solution that incorporates NVLink technology. Whether Intel CEO Li-Wu Chen will also share the stage with NVIDIA CEO Jensen Huang will attract a lot of attention from the attendees.
NVIDIA has announced that it will license its NVLink Fusion technology to other companies, claiming that the design architecture does not necessarily need to use NVIDIA's CPU or GPU. However, this is clearly an attempt to further integrate NVIDIA's technology into diverse computing architectures, thereby making the market more reliant on NVIDIA's solutions and ensuring its competitive advantage.
Intel at @NVIDIAGTC? Yep!
This is just the next step in our partnership as we combine strengths to unlock new innovations, efficiencies, and opportunities for partners building AI at scale.https://t.co/iq1PCZyDQt pic.twitter.com/QDdmsyEW99
— Intel Business (@IntelBusiness) March 13, 2026
Analysis of viewpoints
This year's GTC 2026 sent a very clear signal: NVIDIA is completely transforming from a "company that simply sells GPUs" into a "system behemoth" that "sells AI data center solutions".
Faced with AMD's relentless pressure and the threat posed by cloud giants like Google (TPU) and AWS (Trainium) actively investing in self-developed chips, NVIDIA has chosen to take proactive measures to directly address its weaknesses. Absorbing Groq's technology aims to solve the problems of high cost and speed limitations in the "inference phase" of large language models; while vigorously promoting CPO and Spectrum-X/Quantum-X network architectures ensures that congestion does not occur when hundreds of thousands or even millions of GPUs are connected in series.
When NVIDIA binds computing, networking, optical transmission, and even inference ASICs all to its own ecosystem, the depth and breadth of this moat will likely be difficult for any competitor to cross in the short term.
