Jack Luo — 7 min read

NVIDIA’s CES 2026 keynote framed AI as the next foundational computing platform, arguing that we are in the midst of overlapping platform shifts. Historically, computing has progressed from mainframes to PCs, the Internet, cloud, mobile, and now AI. What’s different in this era is that applications are no longer just “using” AI as a feature; they are increasingly built on top of AI models as the core runtime. This is driving a reinvention of the entire five‑layer computing stack: software is trained rather than hand‑programmed, GPUs replace CPUs as the primary compute engine, and applications generate outputs (tokens, pixels, decisions) dynamically based on context rather than executing pre‑compiled logic.

Economically, NVIDIA positioned this as a wholesale modernization of roughly $10T of existing compute infrastructure toward AI‑centric architectures. The company pointed to hundreds of billions of dollars in annual venture funding and a $100T global industry that is re‑allocating a meaningful portion of its R&D budgets away from classical methods toward AI‑driven approaches. This capital shift is what NVIDIA claims is fueling both the rapid advancement of AI and the intense demand for accelerated computing.

Huang traced a timeline of key AI milestones to contextualize the current inflection. BERT in 2015 was highlighted as the first language model he believed would be truly transformative, followed by the introduction of Transformers in 2017. The 2022 ChatGPT moment “awakened the world” to AI’s potential, but a second critical step came in 2024 with the first O1‑style reasoning model that introduced “test‑time scaling” — effectively giving models the ability to think longer at inference time rather than just emit a one‑shot answer. By 2025, these capabilities had evolved into full agentic systems: models that can reason, call tools and other models, look up information, plan multi‑step workflows, and simulate outcomes. NVIDIA’s view is that these reasoning and agentic paradigms fundamentally reshape both how software is developed and how users will interact with applications.

A major pillar of the keynote was NVIDIA’s open model strategy. The company operates large DGX Cloud AI supercomputers — representing billions of dollars of infrastructure — primarily for its own frontier AI research rather than to compete as a generic cloud provider. Crucially, NVIDIA emphasized that it is releasing these models and the associated training data openly, arguing that transparency about training data is key to trust and that open innovation will ensure AI reaches every company, researcher, and country.

NVIDIA showcased a portfolio of domain‑specific open models. In digital biology, models such as Protein design systems (e.g., successors to AlphaFold‑style work) and EVO2 tackle protein structure understanding and multi‑protein interactions, aiming toward cellular‑level representations. In physics and climate, Earth2AI, ForecastNet, and CorrDiff are used for high‑fidelity, accelerated weather and physical simulations. For general language and reasoning, Nemotron 3 — a hybrid Transformer/SSM model — is designed to be extremely fast while maintaining strong reasoning performance, with more variants promised. Cosmos was introduced as a “world foundation model” focused on understanding how the physical world works, aligned with language so it can explain scenes and dynamics in natural language. In robotics, models like Groot (for humanoid locomotion, articulation, and mobility) and AlphaMile (for self‑driving) are positioned as open, frontier‑level physical AI and autonomy stacks.

Huang claimed that these models are already topping public leaderboards in categories like general intelligence benchmarks, multimodal understanding, document (PDF) parsing, speech recognition, and semantic retrieval/search — what he called the “database engine of the AI era.” All of this, he argued, exists to enable robust AI agents rather than isolated models.

On the application side, NVIDIA described a standard “agentic AI” architecture that is rapidly becoming the default pattern for modern software. In this framework, a smart routing layer inspects user intent and routes subtasks to the most appropriate model — combining proprietary frontier models with heavily customized local or domain‑specific models. This makes applications inherently multi‑model and, by extension, multi‑cloud and hybrid: tasks may run on public cloud APIs, on‑prem GPUs, or edge devices depending on latency, privacy, and cost constraints.

NVIDIA illustrated this architecture with examples of personal assistants that orchestrate email tools, calendars, robots, and text‑to‑speech systems, with an agent deciding when to ground responses via search, when to call external APIs, and when to invoke local models for private data. The company framed this as the new “UI layer” for enterprise platforms. Rather than interacting through spreadsheets, rigid forms, or command‑line tools, users will increasingly converse with an agent that understands text, speech, images, video, 3D data, and specialized formats like PDFs and proteins. This agent becomes both the interface and the orchestrator of complex workflows.

To underscore the ecosystem shift, NVIDIA announced deep integrations of its AI stacks with leading enterprise and infrastructure platforms: Palantir (for AI‑accelerated data and decision platforms), ServiceNow (customer and employee service automation), Snowflake (cloud data warehousing with semantic AI), CrowdStrike (AI‑driven threat detection), and NetApp (data platforms enhanced with semantic and agentic AI for operations and support). In each case, NVIDIA’s agent frameworks, open models, and accelerators are embedded so that the platform’s own “front door” becomes an AI agent.

A significant portion of the keynote was devoted to physical AI: extending AI beyond screens into machines that can interact with the real world. NVIDIA argued that to build such systems you must teach models common‑sense physics — object permanence, causality, friction, gravity, inertia — and that this requires three distinct classes of computers. First, massive training systems for learning the models. Second, inference computers at the edge — deployed in robots, cars, factories, hospitals, and other real‑time environments. Third, specialized simulation computers used to create digital twins and to evaluate and iterate on physical AI safely before deployment.

Simulation, which NVIDIA described as central to its DNA, underpins synthetic data generation, an area they called a breakthrough. Because real‑world, labeled data for complex physical interactions is scarce and narrow, NVIDIA uses simulators and models like Cosmos to generate synthetic, physically plausible surround video and scenarios conditioned on accurate physics. This turns raw compute into high‑quality training data, letting companies systematically cover long‑tail edge cases for robotics, autonomy, and industrial systems in ways that would be infeasible with purely real‑world capture.

To anchor this in industry, NVIDIA announced expanded partnerships with the core EDA and industrial simulation vendors. Synopsys and Cadence are integrating NVIDIA’s accelerated computing and AI into their chip logic design, physical design, emulation, and verification flows, such that future chips and systems will effectively be “born” and iterated inside these AI‑accelerated environments. Siemens will integrate CUDA‑X, physical AI, agentic AI, and digital twin technologies (via Omniverse and related stacks) across the full industrial lifecycle — from design and virtual commissioning of plants and factories, through to operations. Huang characterized future manufacturing facilities as “gigantic robots” whose behavior is designed, tested, and optimized in simulation before being built in the physical world.

All of this is tied together by NVIDIA’s latest hardware platform, the Vera Rubin AI supercomputer, which was unveiled as being in full production and “arriving just in time” for the next AI frontier. Vera Rubin is built from six co‑designed chips engineered to work as a unified system, a break from NVIDIA’s traditional rule of changing only one or two chips per generation. The company argued that the explosive growth in model size, training compute (pre‑training and reinforcement‑learning‑based post‑training), and inference tokens driven by test‑time scaling makes incremental improvements insufficient; only aggressive, end‑to‑end co‑design can keep up.

At the node level, the Vera CPU is a new, large CPU with 88 physical cores providing 176 hardware threads via spatial multithreading, tuned for supercomputing workloads. It delivers roughly 2x performance per watt versus prior generation high‑end CPUs and offers extremely high‑bandwidth I/O, co‑designed with NVIDIA’s new ConnectX‑9 NICs. The Rubin GPU delivers roughly 5x the floating‑point AI performance of the previous Blackwell generation, despite using only about 1.6x as many transistors — highlighting how much of the gain comes from architectural innovation rather than process scaling.

A key enabling technology is the new NV FP4 tensor core “transformer engine,” not just a 4‑bit data format but a full processing block that dynamically and adaptively adjusts precision and structure layer‑by‑layer inside the model. This allows different portions of a transformer to run at lower precision where safe, and higher precision where necessary, all at hardware speeds. NVIDIA claimed this yields very high throughput while preserving accuracy, and suggested NV FP4 could become an industry standard format and approach for high‑efficiency AI.

On the system side, Vera Rubin introduces a fully redesigned compute tray and rack architecture. Each compute tray includes two Vera CPUs and four Rubin GPUs, BlueField‑4 DPUs for offloading storage and security, and multiple ConnectX‑9 NICs, all with a cable‑free internal design. The sixth‑generation NVLink switch enables 18 such compute nodes (72 Rubin GPUs) to operate as a single coherent GPU complex within a rack, moving more data than the global public Internet. Spectrum‑X Ethernet with co‑packaged photonics then scales these racks out into large AI factories.

A single Vera Rubin rack contains 72 Rubin GPUs and on the order of 220 trillion transistors, weighing roughly two tons. A full Vera Rubin AI supercomputer, as shown on stage, aggregates 1,152 GPUs across 16 racks. NVIDIA emphasized practical improvements as well: prior DGX systems required 43 cables per node and up to two hours of careful assembly with high risk of error; the new cable‑less compute chassis reduces assembly to around five minutes and dramatically simplifies manufacturing and service.

Throughout the keynote, NVIDIA tied these advances back to a central thesis: AI has become the new computing platform, with open models, agentic architectures, physical AI, and deeply co‑designed hardware stacks all reinforcing each other. The company’s strategy is to sit at the intersection of these trends — providing frontier‑grade but open models, the hardware and libraries to train and deploy them, and deep integrations with the software and industrial ecosystems that will actually build and operate the next generation of AI‑native applications, robots, and factories.

Originally published on Medium.