The Nvidia GTC 2025 event, which is being held from March 17 to 21 in San Jose, California, takes place against a backdrop of great expectations from the industry and investors.
At the beginning of the year, Nvidia experienced a significant decline in its stock valuation due to the so-called “DeepSeek effect”, caused by the emergence of the Chinese startup DeepSeek, which presented AI models capable of operating with cheaper hardware, calling into question Nvidia’s hegemony in the sector.
Each annual GTC event generates excitement comparable to the launches of the most popular mobile devices. As the benchmark for accelerated computing, anticipation for its new hardware and software announcements is always high, leaving little room for error.
This scenario increased attention on CEO Jensen Huang’s keynote speech on March 18, where announcements were anticipated to reaffirm the company’s leadership in the field of artificial intelligence and accelerated computing. Let’s take a look at what’s new.
Blackwell Ultra Chips
One of the most notable announcements was the Blackwell Ultra chip, an evolution of the Blackwell architecture introduced in 2024. This new chip promises 1.5 times the AI performance of its predecessors, with 208 billion transistors manufactured using a custom TSMC 4NP process.
It includes two arrays connected by a 10 TB/s interconnect, with improvements to the second Transformer engine and Tensor Cores to accelerate the inference and training of large language models (LLMs) and mixture of experts (MoE) models.
Availability is expected in the second half of 2025, with key partnerships with AWS, Google Cloud, Microsoft Azure, Oracle, Cisco, Dell, HPE, Lenovo and Supermicro, ensuring integration into high-performance data centers.
Vera Rubin System
Nvidia also offered details on the Vera Rubin system, a next-generation platform scheduled for 2026, which combines the Vera CPU and the Rubin GPU, named after astronomer Vera Rubin.
This architecture will use HBM4 memory and will be manufactured with a 3nm process from TSMC.
The Vera Rubin system is expected to allow configurations with up to 144 GPUs, while the Rubin Ultra version, expected in 2027, could support up to 576 GPUs, suggesting massive systems for intensive AI applications.
This initiative reflects Nvidia’s focus on scaling computing to meet growing data center demand.
Dynamo Software
Another significant announcement was the launch of Dynamo, an open source inference framework designed to accelerate and scale AI reasoning models in distributed environments.
As the successor to Nvidia’s Triton inference server, Dynamo optimizes inference performance, reducing costs and increasing efficiency, especially for AI factories.
It is reported to double performance for models like Meta’s Llama on the Hopper platform, with partnerships with companies like AWS, Cohere, CoreWeave, Dell, Fireworks, Google Cloud, Lambda, Meta, Microsoft Azure, Nebius, NetApp, OCI, Perplexity, Together AI, and VAST.
This software is now available, marking a step towards the accessibility and scalability of AI.
Partnership with General Motors on autonomous vehicles
Nvidia announced a strategic collaboration with General Motors (GM) to develop fleets of autonomous vehicles.
During the keynote address, Jensen Huang highlighted that GM will use Nvidia’s Drive AGX platform, an in-vehicle computing system that delivers up to 1,000 trillion operations per second, including hardware and software for autonomous driving features and enhanced in-vehicle experiences.
Additionally, the HALO safety system was mentioned, designed to improve safety in autonomous vehicles. This partnership also includes the use of Nvidia’s Omniverse platform for factory simulations and planning, optimizing GM’s manufacturing processes.
Collaboration between Google DeepMind and Disney Research
The most exciting moment of the presentation came when Huang showed a friendly robot that he called “Blue”. This served to showcase the new collaboration between Disney and Google DeepMind, focused on the development of Newton, an open-source physics engine based on Nvidia’s Warp framework.
This engine seeks to improve robot simulation and learning, with applications in humanoid robots and robotic characters. The robot, which looked like something out of a Star Wars movie, had two Nvidia computers inside and walked around Huang, beeping and nodding its head.
Attendees stood to watch the demonstration and recorded with their phones. Disney Research plans to use Newton in its next-generation robots, such as the Star Wars-inspired BDX droids.
Nvidia maintains dominance in data centers
Nvidia has reported significant GPU shipments in recent years, with 1.3 million Hopper GPUs shipped in 2024 to cloud providers such as Microsoft, Alphabet, Amazon and Meta, and a projected 3.6 million Blackwell GPUs by 2025, destined for the four major cloud platforms.
These numbers reflect Nvidia’s dominance in the data center market, with growth driven by demand for AI, especially in large language models and deep learning applications.
Finally, Nvidia predicted that data center spending could surpass $1 trillion by 2028, driven by growing demand for AI computing. Impact and future prospects
Is it enough for investors?
Despite the technological advances presented at Nvidia GTC 2025, the company’s share price experienced a 3.4% drop right after the presentation, reflecting investor dissatisfaction.
One of the reasons is the lack of immediate impact of the ads. Key products like the Blackwell Ultra and Vera Rubin systems, while impressive, won’t be available until mid-2025 and 2026, respectively.
On the other hand, the market seemed to anticipate more revolutionary developments. Instead, many announcements, such as the deployment of the Blackwell Ultra, appeared to build on already known plans rather than introduce unexpected developments, leading to a perception of insufficient progress.
As always with AI, it’s a matter of expectations. Given Nvidia’s dominant position in AI and lofty valuation, investors likely set a high bar for the event. It wouldn’t be the first time.
This post is also available in: