Nvidia Just Lit the Fuse on “AI 2.0” at CES — Meet Vera Rubin, the Chip Platform That Could Redefine the Next Boom

Nvidia used its highly anticipated CES 2026 keynote to do something Jensen Huang has become increasingly comfortable doing: reset the conversation around what AI demand actually looks like over the next decade—and why Nvidia intends to remain at the center of it. While the event featured plenty of technical detail, the overarching message was clear. AI is no longer just about training large language models, and Nvidia is no longer just a chip company. The future, as Huang framed it, is about compute-hungry systems that span data centers, robotics, autonomous vehicles, and what he repeatedly referred to as “physical AI.”

At the core of the keynote was the formal launch of Nvidia’s Vera Rubin platform, the company’s next-generation computing architecture that will ultimately replace Blackwell. Rubin is already in full production, with a broader ramp expected in the second half of 2026, and will be deployed by essentially every major cloud provider. Nvidia confirmed that Rubin systems are slated for use by partners including OpenAI, Anthropic, Amazon Web Services, and HPE, as well as national lab supercomputers like Lawrence Berkeley’s upcoming Doudna system. This matters because it reinforces a key theme: Nvidia’s product cadence is not slipping, demand remains committed, and hyperscalers are planning infrastructure years in advance.

Rubin is not a single chip, but an integrated platform built around six tightly coordinated components. At the center is the Rubin GPU, paired with a new Vera CPU designed specifically for agentic reasoning workloads. Surrounding those are upgrades to NVLink, BlueField DPUs, ConnectX SuperNICs, Spectrum Ethernet switches, and a new pod-level memory architecture. The goal is not incremental performance gains at the chip level, but system-level efficiency across compute, networking, storage, power, and cooling. Nvidia is designing the entire data center as a single machine, a theme Huang emphasized repeatedly during the keynote.

Performance improvements were substantial. Nvidia said Rubin delivers roughly 3.5x faster training performance and up to 5x faster inference compared with Blackwell, while supporting up to eight times more inference compute per watt. For customers, the economics may be even more compelling. Nvidia claims Rubin can reduce inference token costs by as much as 10x and cut the number of GPUs required to train mixture-of-experts models by roughly fourfold versus prior-generation systems. In a world where AI workloads increasingly involve long-context reasoning, test-time scaling, and continuous inference, those cost reductions matter as much as raw speed.

That shift toward reasoning-heavy workloads was one of the keynote’s central themes. Huang made it clear that AI demand is no longer driven solely by training massive models once, but by inference that runs continuously. Agentic systems plan, search, iterate, and call tools in real time, generating far more tokens per task and placing sustained strain on memory, networking, and power infrastructure. Nvidia’s response is to move aggressively beyond selling GPUs and toward delivering full-stack platforms that can handle long-running, compute-intensive workloads efficiently.

Another major pillar of Nvidia’s growth narrative was physical AI, particularly robotics and autonomous vehicles. Huang reiterated that robotics is now Nvidia’s second most important growth category after AI compute. The company unveiled Alpamayo, a new family of open-source reasoning models designed for autonomous driving. Alpamayo 1, a 10-billion-parameter chain-of-thought model, is designed to break down complex driving scenarios into smaller decision steps and explain its reasoning—an approach that mirrors how humans handle unexpected situations. Nvidia also introduced AlpaSim, a simulation platform that enables closed-loop training for rare driving scenarios that are difficult to capture in the real world.

Nvidia confirmed that the 2025 Mercedes-Benz CLA will be the first vehicle to ship with Nvidia’s full autonomous driving stack, and Huang went further, saying Nvidia plans to test a robotaxi service with a partner as early as 2027. These initiatives underscore Nvidia’s belief that autonomy will require massive amounts of simulation, synthetic data, and continuous inference—again feeding directly into long-term compute demand.

From an industry perspective, Huang addressed concerns investors continue to raise around AI spending sustainability. He argued that AI is being funded not by speculative excess, but by a structural reallocation of capital. Nvidia estimates that $3 trillion to $4 trillion will be spent on AI infrastructure over the next five years, driven by the modernization of legacy computing, changes in R&D workflows, and the expansion of AI into real-world systems. Huang framed this as a repurposing of the roughly $10 trillion spent on computing infrastructure over the past decade, rather than a temporary spending bubble.

Geographically, Nvidia also acknowledged continued demand from China, noting strong interest in the H200 chip, though shipments remain subject to licensing approvals. While regulatory uncertainty persists, Nvidia’s messaging suggested that demand is not the constraint—supply and policy are.

Competition was acknowledged but downplayed. AMD is rolling out its own rack-scale Helios systems, and hyperscalers continue to invest in custom silicon. Still, Nvidia’s annual product cadence, deep software ecosystem, and ability to deliver integrated systems make it difficult for rivals to close the gap meaningfully in the near term. Analysts echoed that sentiment following the keynote, reiterating buy ratings and emphasizing Nvidia’s dominance across compute, networking, and systems.

Ultimately, Huang’s CES keynote was less about flashy surprises and more about reinforcing Nvidia’s long-term strategy. AI demand is broadening, not slowing. Workloads are becoming more complex, not lighter. And Nvidia is positioning itself not just as a supplier of chips, but as the architect of the infrastructure required to sustain AI’s next phase. For investors, the message was simple: the ceiling for compute demand keeps rising, and Nvidia intends to build the ladder.