CES 2026 (Consumer Electronics Show) was held from January 6 to 9 in Las Vegas, Nevada, where global tech giants unveiled their most critical products for the year. As one of the most anticipated speakers, NVIDIA CEO Jensen Huang took the stage to share the company’s latest AI roadmap for 2026. Below is a summary of the major highlights and takeaways from the keynote, including analysis on how these developments could shape AI-related investment trends in 2026.
Highlight 1: AI Becomes the Next Foundational Computing Platform
Roughly every 10 to 15 years, the electronics and computing industry undergoes a major platform reset. This cycle is marked by two simultaneous shifts:
- AI has evolved from a standalone application technology to a general-purpose foundational platform. Moving forward, the core logic behind most applications and services will be built directly on top of AI.
- The software development paradigm is shifting from code-centric programming to model-centric training. Applications are no longer precompiled static programs but dynamic systems capable of understanding context and generating content in real time. This shift is accelerating the transition from CPU-based to GPU-centric computing architectures. GPUs are no longer just accelerators — they are now central execution units in next-gen computing stacks, prompting a near-total redesign of existing system layers.
Since the breakthrough of large language models (LLMs) in 2022 with ChatGPT, AI hardware demand has surged, driven by the scaling law — the idea that expanding compute power, data, and model size leads to consistent capability improvements. On the inference side, demand is growing rapidly due to better inference performance and the rise of agentic AI involving multi-step tasks and tool usage. Even as per-token costs drop, this cost reduction boosts adoption and usage intensity, leading to increased total compute consumption. Ultimately, compute demand driven by AI is shifting from short-term product cycles to long-term capital investment-driven structural growth.
Vera Rubin Enters Production Stage
Huang emphasized that relying on process shrinkage alone can no longer meet growing compute demands. Instead, the key lies in extreme compute stack optimization and cross-component co-design. NVIDIA plans to position Vera Rubin as a full-stack AI supercomputing platform, built around the Vera CPU and Rubin GPU, forming a new "Superchip" architecture. The platform will use NVLink 72 and the next-gen Spectrum-X AI Ethernet for data transmission.
Vera Rubin integrates TSMC’s new Compact Universal Photonic Engine (COUPE) to incorporate optical modules directly into chip packaging, enhancing switching efficiency. It also features CoWoS-L (Local Silicon Interconnect) advanced packaging technology to increase overall chip footprint. By combining cutting-edge technologies with tight cross-component integration, NVIDIA aims to build a scalable rack-level AI supercomputing solution. Deployment is scheduled to begin in the second half of 2026 (2H 2026) among cloud service providers and enterprise clients.
Highlight 2: Open-Source Models Gain Momentum, Driving Long-Term Industry Growth
The emergence of DeepSeek R1 highlights the accelerating spread of open-source models across industries and regions. Though still roughly six months behind leading closed-source models, their capabilities are advancing at a pace of "one generation every six months." This trend signals that AI innovation is no longer monopolized by a few large cloud providers.
As the open-source ecosystem matures, it benefits the broader computing platform and GPU supply chain. AI demand is becoming more widespread and fragmented, extending into a variety of new use cases. With increasingly complete open-source models, datasets, and tooling, organizations can train, deploy, and iterate faster, while significantly lowering adoption barriers. This drives both hardware usage and deeper software-hardware integration. Overall, open source is not eroding the industry's competitive landscape — rather, by making models, data, and tools more accessible, it expands the ecosystem and becomes a structural driver of long-term growth.
Highlight 3: Agentic AI Evolves Toward Autonomous Systems
Agentic AI shifts the paradigm from "tool-based applications" to "autonomous systems." It’s no longer just about content generation — these models can plan, execute, and learn independently. By breaking down tasks into steps, combining search and verification, tool usage, planning, and execution, agentic AI reduces hallucination risks.
This evolution is steering the industry away from omniscient general-purpose models toward task-specific, customized model architectures. System designs now feature multimodal, multi-model, multi-cloud, and hybrid-cloud configurations, with context-aware model selection and dynamic load balancing across cloud and edge environments.
Agentic AI is also redefining enterprise software interfaces. SaaS platforms are transitioning from spreadsheets and command lines to dialogue-based interactions. Through intent-driven model selection, privacy-sensitive tasks can be handled by local models, while general tasks can be routed to frontier models, integrating robotics, speech, and vision applications.
As multi-step reasoning increases context and system loads, inference now requires greater memory and architectural support. Context is being extended beyond GPUs into storage and network layers. Agentic AI marks a structural divergence between inference and training infrastructure, ushering in a new generation of AI systems built around autonomous decision-making and long-duration operations.
Highlight 4: Physical AI Bridges the Gap with the Real World
Physical AI consists of three distinct computing systems: training computers, inference computers, and simulation computers for generating and validating data. In the future, AI will be capable of understanding the physical world’s underlying rules — object permanence, causality, friction, gravity, inertia, and behavioral consequences.
The real world poses challenges like limited data and complex context. It's impossible to collect real-world data for every interaction scenario, which is where simulation becomes essential. Through NVIDIA’s Omniverse digital twin and Cosmos world model, compute power is transformed into synthetic data that complies with physical laws. This allows AI to repeatedly learn, test, and infer within virtual environments before entering the real world.
Highlight 5: Alpamayo — Explainable, End-to-End Autonomous Driving
Autonomous driving is expected to be the first mainstream application of Physical AI. NVIDIA’s Alpamayo is the world’s first explainable, end-to-end self-driving AI — capable of articulating its planned actions, reasons behind those actions, and expected driving trajectory. This is vital for handling the "long tail" of rare edge cases in autonomous driving.
Alpamayo is trained using human driving demonstrations, synthetic data generated by Cosmos, and high-quality manual annotations. This enables the model to deconstruct unfamiliar situations into solvable sub-problems. Crucially, Alpamayo operates alongside a rule-based traditional self-driving stack, with a safety and policy assessor dynamically determining whether control should be handed to Alpamayo or default back to the conservative system.
This dual-stack system, combined with a production partnership with Mercedes-Benz, sets Alpamayo on a path toward real-world deployment and continuous OTA upgrades — paving the way for commercial-grade Physical AI.
Highlight 6: Robotics as the Next AI Industry Wave
In the robotics domain, Huang described the next decade as the “true age of robotics.” He showcased how robots are being trained virtually in Omniverse and Isaac Sim/Lab environments to walk, manipulate, coordinate, and avoid obstacles. Only once these behaviors are stable are robots deployed in the real world — eliminating the need for risky and costly physical trial-and-error.
Taking it a step further, entire factories are now viewed as massive robotic systems — from chip design and simulation to production and operation — all done and validated in virtual environments first. This strategy underpins NVIDIA’s collaborations with industrial software leaders like Cadence, Synopsys, and Siemens, bringing agentic AI into every layer of design, manufacturing, and operations.
Ultimately, the goal is for robots of all types and sizes to share a unified Physical AI methodology. Following autonomous vehicles, this could drive the next and even larger wave of industrial AI adoption.
You can explore more about NVIDIA from the following:
