Summary Bullets:
• NVIDIA believes enterprises are evolving from digital transformation driven by large language models (LLMs), and towards physical AI, or spatial AI, and multimodal AI.
• The industry at large is turning to “reasoning” models, capable of offering more analysis and autonomy than LLMs, especially combined with AI agents.
GTC is NVIDIA’s most important conference, focused on developers, who are the core of the AI market. This year, the event was held on March 17-21 in San Jose, California (US). NVIDIA believes that enterprises should be diversifying from purely digital transformation driven by LLMs, and towards physical AI or spatial AI, as well as multimodal AI, robotics, and AI at the edge. For this reason, the event tilted towards the industrial, automotive, and manufacturing sectors. The industry at large is turning to “reasoning” models, capable of offering more analysis and autonomy than LLMs, especially combined with AI agents. This transition should bring further demand for raw computing power, despite efficiency gains in terms of chipset architectures.
GTC gave the company an opportunity to offer an update on its roadmap for the next few years. The next generation of specialized Blackwell AI chips, called Blackwell Ultra, is slated for delivery from H2 2025. NVIDIA has changed the cadence of its releases, formerly every other year, to follow annual updates, more in tune with the frantic pace of the current AI landscape. Regarding forthcoming generations, the Vera Rubin chip, named after the woman who provided the first observational evidence that supported the existence of dark matter, is planned for 2026. Additionally, NVIDIA debuted a hybrid architecture for agentic AI, the GB300 chip, which blends 72 Blackwell Ultra graphic processing units (GPUs) and 36 Arm Neoverse-based NVIDIA Grace central processing units (CPUs). The GB300 processor is housed within the NVIDIA GB300 NVL72 rack, which is liquid-cooled, and the NVIDIA HGX B300 NVL16 system. The company also unveiled an optical networking system, Spectrum X, to help combat bottlenecks to building AI-focused data centers.
NVIDIA launched its open-source software NVIDIA Dynamo, for AI data centers, and to optimize the performance of Blackwell Ultra chips, just as “reasoning” models such as open-source DeepSeek’s V3 and R1 AI systems gain increasing relevance. The company claims its Dynamo software can boost DeepSeek’s models’ performance by a factor of 30. The US chipmaker saw its own market capitalization suffer a wipe-out of $589 billion back in January, as investors reacted to the news of DeepSeek’s entrance in the market, with an AI model claimed to operate at a fraction of the cost of other equivalent technologies.
In a sign of the current geopolitical climate, NVIDIA highlighted that there will be a shift towards moving the supply chain back from Asia into the US. The company is following the same path carved by some of its largest competitors, and announced eye-popping investments, including plans to spend $500 billion over the next four years in electronic components, chip design and manufacturing. This comes just as Taiwan Semiconductor Manufacturing Company (TMSC), which manufactures the processors designed by NVIDIA, announces plans to invest $100 billion in three more chip manufacturing plants in the US, signaling the company is moving some of its most advanced manufacturing processes to the US.
Fittingly, the company also announced its participation in a consortium bringing together utility companies, technology vendors, and academic researchers, to build AI models that optimize the efficient use of electricity, called the Open Power AI Consortium. The consortium is led by nonprofit energy R&D organization EPRI.
In addition, the world’s largest chipmaker announced the acquisition of synthetic data company Gretel. NVIDIA has been releasing synthetic data generation tools recently and synthetic data is seen as a solution to the problem of diminishing data pools, challenges focused on intellectual property, and so on. However, increasing use of synthetic data comes with its own set of problems, and AI models trained on synthetic data containing flaws will inevitably compound defects and poor quality.
The chipmaker is opening a quantum lab based in Boston, to accelerate innovation and foster collaboration with Harvard University and the Massachusetts Institute of Technology, called the NVIDIA Accelerated Quantum Research Center to collaborate with quantum firms including Quantinuum, Quantum Machines, and QuEra Computing. NVIDIA said the center will begin operating later this year.
NVIDIA struck a bullish tone, but some issues were left unaddressed. The company is making a strong bet that the arrival of “reasoning” AI systems such as DeepSeek’s v3 will spur even greater demand for computing power. But in a world of limited grid capacity, it is the move from raw processing power to efficient processing that lies at the core of growth. It makes sense for the company to highlight technologies such as digital twins, and the possibility of use cases outside a few core industries. This would compound current demands for computing power. However, it remains to be seen how long timeframes will be for adoption at scale.