Jensen Huang delivered on the hype. In his highly anticipated CES 2026 keynote on Monday evening, the Nvidia CEO unveiled the Rubin platform—a comprehensive six-chip system designed to power the next generation of artificial intelligence. The announcement represents the most significant hardware roadmap update from the world's most valuable company since the Blackwell architecture was revealed in 2024.

The platform, named after astronomer Vera Rubin who discovered evidence of dark matter, is now in full production. Cloud partners including AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure will begin deploying Rubin-based instances in the second half of 2026.

The Six-Chip Architecture

Nvidia's Rubin platform comprises six distinct chips, each optimized for a specific role in the AI computing stack:

  • Vera CPU: A custom ARM-based processor with 88 Olympus cores and 176 threads. The Vera CPU replaces the current Grace CPUs and features 1.5TB of system memory with 1.8 TB/s NVLink core-to-core interface to link with Rubin GPUs. The chip contains 227 billion transistors.
  • Rubin GPU: The next-generation graphics processing unit manufactured on TSMC's 3nm process with HBM4 memory. Nvidia claims up to 10x lower cost per token compared to Blackwell for inference workloads.
  • NVLink 6 Switch: An upgraded interconnect chip enabling faster communication between GPUs in multi-GPU configurations.
  • ConnectX-9 SuperNIC: A networking chip designed for high-bandwidth, low-latency connections in data center environments.
  • BlueField-4 DPU: A data processing unit for offloading networking, storage, and security functions from the main CPUs.
  • Spectrum-6 Ethernet Switch: An advanced network switch for managing data center traffic at scale.

System Configurations

Nvidia will offer two primary system configurations for enterprise customers:

Vera Rubin NVL72: A rack-scale solution featuring 72 GPUs with full NVLink interconnect, designed for the largest AI training workloads. This configuration targets hyperscale cloud providers and major AI research labs.

HGX Rubin NVL8: A smaller eight-GPU system suitable for enterprise data centers and mid-scale AI deployments. Each system includes two Vera CPUs and two Rubin GPUs in its base configuration.

Performance Claims

Huang made aggressive performance claims during the keynote, positioning Rubin as a generational leap:

"Wherever the universe has information, we can teach a large language model to understand its representation and turn it into AI," Huang declared from the CES stage. "Rubin is the infrastructure that makes this possible at unprecedented scale."

Key performance metrics include:

  • 10x lower cost per token: For inference workloads compared to Blackwell, driven by improvements in memory bandwidth, compute efficiency, and power management.
  • Agentic AI optimization: The platform is specifically designed to accelerate multi-step reasoning and autonomous AI agents.
  • Massive mixture-of-experts support: Enhanced capability for running the largest AI models with hundreds of billions of parameters.

Cloud Partner Deployments

Nvidia announced that major cloud providers will be among the first to deploy Vera Rubin systems:

  • AWS: Will offer Rubin-based EC2 instances in the second half of 2026
  • Google Cloud: Integration with Google's AI platform announced
  • Microsoft Azure: Rubin systems will power next-generation Azure AI infrastructure
  • Oracle Cloud Infrastructure: Commitment to deploy Rubin at scale

Additionally, Nvidia Cloud Partners including CoreWeave, Lambda, Nebius, and Nscale will offer Rubin-based instances, expanding access beyond the hyperscale providers.

The Rubin CPX: A New GPU Class

Beyond the standard Rubin GPU, Nvidia also unveiled the Rubin CPX—a specialized chip designed for massive-context inference workloads. As AI models increasingly require the ability to process longer inputs (context windows now reaching millions of tokens), the CPX addresses a specific bottleneck in current architectures.

The CPX is positioned for applications like document analysis, code generation across entire repositories, and conversational AI with extensive memory.

Manufacturing and Timeline

The Rubin platform will be manufactured by TSMC using advanced 3nm process technology. HBM4 memory from Samsung and SK Hynix will provide the bandwidth necessary for AI workloads.

Key dates:

  • Q1 2026: Full production ramp (announced today)
  • H2 2026: Cloud partner availability begins
  • 2027: Rubin Ultra variants expected, following Nvidia's annual cadence

What Wasn't Announced: Consumer GPUs

Notably absent from Huang's keynote was any consumer GPU news. The RTX 50 series launched at last year's CES, and enthusiasts hoping for Super variants were disappointed. Reports suggest memory shortages have delayed RTX 50 Super cards, with availability now expected later in 2026.

Nvidia's consumer gaming division remains profitable, but the company's strategic focus has clearly shifted to data center AI, where margins are higher and demand seems insatiable.

Market Implications

Nvidia's stock closed Monday up 2.3%, with the after-hours reaction to the keynote still developing. The announcement reinforces several investment themes:

  • Sustained AI capital spending: Cloud providers will need to continue heavy investment to access Rubin capabilities.
  • Margin protection: The 10x efficiency claim suggests Nvidia can maintain pricing power even as volumes grow.
  • Competitive moat: The integrated six-chip platform creates switching costs that competitors will struggle to match.

The Feynman Tease

In a brief mention, Huang referenced "Feynman" as the codename for Nvidia's post-Rubin architecture, adding it to the company's public roadmap. Named after physicist Richard Feynman, the platform is expected around 2028—demonstrating Nvidia's commitment to annual architectural advancement.

The Bottom Line

Nvidia's Rubin platform announcement confirms the company's position at the center of the AI revolution. With six new chips, full production status, and committed cloud partners, Huang has given investors exactly what they wanted: evidence that the $500 billion pipeline of anticipated GPU sales is not just a forecast, but a reality in the making. For competitors, the message is clear: catching Nvidia just got harder.