NVIDIA enters 2026 in a position few technology companies have occupied before. Once best known for gaming graphics, it now sits at the centre of how AI is trained, deployed and scaled across industries.
In early 2025, the giant reached a $4 trillion market capitalisation, followed by $5 trillion just months later in October, making it the first company in history to cross that threshold. Its chips powered the biggest LLMs, cloud data centres and an expanding class of AI systems.
Over the past year, the company has advanced progressively on two fronts. The first is software and models, where open and reasoning-based AI has gained momentum. The second is infrastructure, where the cost and complexity of running AI at scale have emerged as significant constraints for enterprises and governments alike.
NVIDIA’s core argument is that these two problems must now be solved together. That context framed the company’s announcements at the Consumer Electronics Show (CES) 2025, where the company laid out a broad roadmap spanning physical AI models, autonomous systems and a new generation of computing hardware designed around inference and long-running AI workloads.
CEO Jensen Huang called this moment a turning point. “The ChatGPT moment for physical AI is here, when machines begin to understand, reason and act in the real world,” he said in his keynote.
Huang announced a new suite of open-source physical AI models alongside a next-generation computing platform. He outlined how NVIDIA plans to scale autonomous systems, robotics and large-scale AI inference. Commercial rollouts based on these announcements are expected to begin in early 2026.
From Chatbots to Physical AI
The company said it is entering what it describes as the era of physical AI, in which models move beyond perception and language to enable real-world action.
“Everything that moves will ultimately be fully autonomous, powered by physical AI,” Ali Kani, the VP and GM of automotive at NVIDIA, said in a briefing. He added that safety-focused reasoning models are central to that shift.
NVIDIA framed the developments as a response to rising demand for agentic AI, robotics and autonomous driving, alongside growing infrastructure needs for large models with extended context windows.
Central to this shift is Alpamayo, the family of open reasoning models for autonomous vehicles announced at the event. The company described Alpamayo as the first open-source vision-language-action model designed to reason through complex driving scenarios rather than relying solely on pattern matching.
“It allows autonomous vehicles to really think,” Kani said, noting that the model can reason through situations such as traffic light failures without prior examples.
“Alpamayo does something that’s really special. It doesn’t just take sensor input and activate steering, brakes and acceleration; it also reasons about what action it is about to take,” Huang said.
Alpamayo’s flagship model has 10 billion parameters and takes in inputs such as camera feeds, vehicle history and navigation context. It outputs driving trajectories alongside reasoning traces that explain why a decision was taken.
Alongside the model, NVIDIA is releasing more than 1,700 hours of autonomous driving data and AlpaSim, an open-source simulation framework that allows developers to test reasoning-based driving stacks using real and synthetic data.
The company also expanded its robotics portfolio. It introduced updated open models under its Cosmos and Groot families, including Cosmos Reason 2 for physical reasoning, and made GR00T 1.6, its humanoid robotics model, generally available. NVIDIA also said its Nemotron family now includes retrieval-augmented generation, safety and speech models.
NVIDIA named partners such as Boston Dynamics, LG Electronics and NEURA Robotics, which are already building next-generation robots using its Isaac and Jetson platforms.
“We don’t monetise our models,” said Kari Briski, the VP of generative AI software for enterprise at NVIDIA. “We open-source the models. We open-source the data that we use to train those models because only that way can you truly trust how the models came to be,” Huang described.
New Stack for Inference-Heavy AI
Among models and software, Huang also turned to hardware, officially launching the Vera Rubin platform. Named after astronomer Vera Rubin, the platform is NVIDIA’s successor to Blackwell and is designed as a rack-scale system rather than a single chip.
The company describes Rubin as six co-designed chips working as one system. These include the Rubin GPU, Vera CPU, sixth-generation NVLink switches, ConnectX-9 networking, BlueField-4 data processors and Spectrum-X Ethernet. The goal is to reduce the cost of training and inference for mixture-of-experts and agentic AI models.
At the centre is the Rubin GPU, which the company claims delivers up to 10x reduction in inference token cost and 4x reduction in the number of GPUs to train MoE models, compared with the NVIDIA Blackwell platform. It is also the first GPU to use HBM4 memory. The Vera CPU, built with 88 custom Arm cores, handles data movement, scheduling and security across the rack.
One of the more notable additions is a new Inference Context Memory Storage Platform, designed to store the key-value cache generated by long-running AI agents. NVIDIA said traditional storage systems struggle with this workload.
“AI is no longer about one-shot chatbots but intelligent collaborators that understand the physical world, reason over long horizons, stay grounded in facts, use tools to do real work, and retain both short- and long-term memory,” Huang said.
All of this comes together in the Vera Rubin NVL72 rack and a new DGX SuperPOD reference design aimed at what the giant calls ‘AI factories’. Dion Harris, NVIDIA’s senior director of HPC and AI hyperscale infrastructure solutions, said partners are already validating systems and that products will be available in the second half of 2026.
Partners, Clouds and What Comes Next
NVIDIA also took the opportunity to underline the breadth of its ecosystem. Cloud providers, including Microsoft and CoreWeave, were named as early adopters of Rubin-based systems. Storage companies such as NetApp and VAST Data are working on the new context memory tier.
In automotive, NVIDIA confirmed that its full-stack DRIVE platform is launching in the United States, with hands-free highway driving, enhanced level-2 point-to-point driver assistance and planned end-to-end urban driving features.
The new Mercedes-Benz CLA, the brand’s first vehicle featuring the MB.OS platform will be a beneficiary of the full-stack DRIVE AV software, AI infrastructure and accelerated compute, introducing advanced driver-assistance features.
In robotics, multiple partners, including LEM Surgical, AGIBOT, Richtech Robotics, unveiled machines built on NVIDIA hardware and software during the show. Boston Dynamics also unveiled its new E-Atlas humanoid, which runs on Jetson Thor and was trained using Isaac Lab.
As Huang put it on stage in Las Vegas, “Every 10 to 15 years, the computer industry resets. A new platform shift happens.” NVIDIA’s argument is that the next reset is already underway and that it intends to build the full stack behind it.
The post NVIDIA Thinks the Future of AI Has a Body appeared first on Analytics India Magazine.