AI Descends into the 'Real World' at CES 2026: The Next Moves from NVIDIA, AMD, and Boston Dynamics
CES 2026 (Las Vegas) marked a turning point where the main battlefield for AI shifted from 'cloud-based text generation' to cars, robots, and PCs. Symbolically, NVIDIA simultaneously launched (1) open models capable of 'reasoning' for autonomous driving, (2) the next-generation computing infrastructure 'Rubin,' and (3) the 'Android-ization' of general-purpose robots. Meanwhile, AMD put AI PCs at the forefront, and Boston Dynamics is teaming up with Google DeepMind to enhance its 'brains'.
1. The 2026 slogan is 'Physical AI'—moving toward AI that reasons in the 'real world'
NVIDIA announced 'Alpamayo,' a group of open AI models designed for autonomous vehicles to break down and think through complex situations step-by-step. CEO Huang emphasized that we have entered a phase of understanding, reasoning, and acting in the real world, stating, 'The ChatGPT moment for physical AI is here.'
The core 'Alpamayo 1' is a 10-billion parameter 'chain-of-thought' VLA (Vision-Language-Action) model. It is described as being designed to 'identify possibilities and choose the safest route' even in 'rare cases,' such as intersections with broken traffic lights, without needing prior experience. Furthermore, they simultaneously released an open dataset containing over 1,700 hours of driving data and an OSS simulator for verification called 'AlpaSim.' In short, the strategy is to distribute 'models + data + simulation' as a single set to exponentially increase the number of trials by developers.
2. 'Thinking AI' also changes hardware—Rubin is not a 'GPU' but 'one AI supercomputer made of 6 chips'
Another pillar is the next-generation architecture, 'Vera Rubin.' According to TechCrunch, Rubin is a 'rack-scale' design that coordinates six chips, configured to eliminate bottlenecks not just in the GPU, but also in storage, interconnects (NVLink), and DPUs (BlueField).
The performance message is also intense; NVIDIA's tests show 3.5x the training performance and 5x the inference performance (up to 50PF) compared to Blackwell, with significantly increased 'inference compute/watt.' Additionally, it is explained that a new storage tier has been introduced to handle KV cache (memory load) issues, which are problematic for 'agentic AI'.
The Verge summarizes Rubin as an 'integrated platform' that includes Vera CPU / Rubin GPU / NVLink / NIC / DPU / Ethernet, appealing to the leap from Blackwell by even factoring in 'token costs.' In other words, Rubin is not just a 'fast GPU,' but a concept aimed at selling the production equipment of an AI factory itself.
3. NVIDIA's 'Android Strategy'—aiming to become the standard OS for robot development
NVIDIA's metaphor of wanting to become the 'Android of general-purpose robotics' hit home. What was announced was a full stack of foundation models, simulation, and edge hardware for robots to 'see, understand, and move' in diverse environments.
The models listed by TechCrunch include world models for synthetic data generation and evaluation (Cosmos Transfer / Predict), the inference VLM 'Cosmos Reason 2,' and the VLA for humanoids, 'Isaac GR00T N1.6.' Furthermore, they provide the OSS 'Isaac Lab-Arena' for safe verification in virtual space and 'OSMO' as a command center to connect workflows. On the hardware side, new cards in the Jetson series (Thor family) were also brought to the fore, revealing a vision to capture the 'robot development experience'.
This is where the ecosystem comes into play. NVIDIA is deepening its collaboration with Hugging Face, and it is explained that Isaac/GR00T will be integrated into LeRobot. The goal is clear: moving from 'running in a lab' to 'runnable by anyone.' Just as Android bundled device manufacturers and developers, NVIDIA is moving to capture the developer distribution (toolchain) and standardize it.
4. AMD is 'bringing AI to the PC'—capturing the daily battlefield with Ryzen AI
Meanwhile, AMD framed the context of CES as 'AI for everyone,' pushing AI PCs to the forefront. According to TechCrunch, the 'Ryzen AI 400' claims 1.3x the multitasking and 1.7x the content creation performance compared to competitors. With a 12-core/24-thread configuration, it is a route aimed at capturing the experience of 'AI running locally' (generation, editing, gaming) through performance.
However, AMD is not just about 'PCs.' Reuters reports that at CES, AMD also mentioned AI chips for data centers and moves involving its relationship with OpenAI (*a challenge to the market dominated by NVIDIA). In other words, in 2026, AMD is poised to capture the AI adoption phase by eyeing both the edge (PC) and the core (DC).
5. Boston Dynamics x DeepMind—loading a 'general-purpose brain' onto the 'strongest body'
The most important partnership on the robotics side is Boston Dynamics and Google DeepMind. According to TechCrunch, they will use the next-generation humanoid 'Atlas' as the first testbed to integrate DeepMind's AI foundation models and accelerate development to enable more 'human-like' behavior. The head of DeepMind stated on stage, 'We are aiming for a state-of-the-art robot foundation model.'
The weight of this news lies in the fact that it looks toward mass production and on-site implementation, not just 'research.' Boston Dynamics has already deployed the quadruped Spot in over 40 countries, and the warehouse robot Stretch has unloaded over 20 million boxes since its introduction in 2023 (according to Hyundai). Furthermore, it is reported that Atlas has entered production and is headed to the Hyundai factory in Savannah, Georgia, USA.
AP and WIRED also reported on the unveiling of Atlas at CES and the DeepMind partnership, linking it to 'implementation on the factory floor.' What can be read from this is that the victory in the robot race has become a complex battle of not just 'mechanical strength,' but 'general-purpose models + field data + safety design.'
6. What will happen: The points of focus for 2026 are 'data,' 'standards,' and 'safety'
Finally, I will briefly summarize this series of announcements from an 'investor and business perspective'.
Data and simulation as a moat: Alpamayo's 1,700 hours of data, AlpaSim, and the Isaac Lab ecosystem are tools to increase the number of training iterations in domains where 'failure in the real world is not an option.' Winning is more likely to come down to the 'design of trial frequency' rather than model size.
Capturing the market through standardization (Android-ization): NVIDIA is moving to bundle not just 'chips' but the 'development experience.' If this takes hold, peripheral companies will begin optimizing with 'NVIDIA as the prerequisite'.
Safety and accountability as differentiators: Whether in autonomous driving or humanoids, the question asked in the event of an accident is 'why did it make that decision?' NVIDIA's emphasis on the ability to 'explain driving decisions' anticipates the issues that will follow the performance race.
CES 2026 was the starting gun for AI entering industries in a 'visible form.' 2026 will see a shift from 'cloud GPU hegemony' to 'real-world OS hegemony'—the very way that battle is fought will change.