At CES 2026, NVIDIA made an announcement that will reshape the global AI landscape. Named after Vera Rubin – the American astronomer who discovered dark matter – NVIDIA’s newest AI supercomputer is not just an incremental upgrade. It is a complete architectural reinvention: six co-designed chips, 220 trillion transistors, 100 petaflops of AI compute in a single rack, and the world’s most advanced liquid cooling system. Vera Rubin is now in full production.

Understanding why this matters requires understanding the problem NVIDIA was solving — and the astronomical ambition behind the name itself.

Why Name It After Vera Rubin?

Vera Rubin was an American astronomer who made one of the most profound discoveries in modern physics. She observed something that should have been impossible: the outer edges of galaxies were rotating at roughly the same speed as the stars near their centers — a direct contradiction of Newtonian physics. Just as planets further from the sun orbit more slowly, galactic stars should follow the same rule. They didn’t.

Her conclusion? There must be invisible mass — matter we cannot see — holding galaxies together. She called it dark matter. It remains one of the most important discoveries in astronomy, fundamentally changing our understanding of the universe.

“It makes no sense unless there are invisible bodies — dark matter — that occupy space even though we don’t see it.”

— Jensen Huang, NVIDIA CEO, CES 2026, describing Vera Rubin’s discovery

In naming their most powerful supercomputer after her, NVIDIA honors a scientist who revealed hidden forces shaping the universe. The parallel is deliberate: just as dark matter governs the cosmos invisibly, computation governs the AI revolution from beneath the surface. The faster the compute, the sooner humanity reaches the next frontier.

The AI Compute Explosion Driving Vera Rubin

To appreciate why Vera Rubin was necessary, you need to understand the brutal mathematics of modern AI scaling. Three simultaneous forces are compounding demand for computation at rates that break traditional hardware roadmaps.

The Three Forces Driving AI Compute Demand:

• Model scale: AI models are growing 10× in size every year
• Token explosion: Test-time scaling is generating 5× more tokens annually
• Post-training cost: Reinforcement learning inflates pre/post-training compute dramatically
• Inference shift: O1-style “thinking” models replaced single-shot inference
• Cost race: Token costs are declining 10× per year as competition intensifies
• Frontier race: Every lab is racing to reach the next capability threshold simultaneously

The launch of OpenAI’s O1 model was, in Jensen Huang’s words, “an inflection point for AI.” Instead of answering in a single forward pass, modern inference is a thinking process — the model reasons through a problem step by step. The longer it thinks, the better the answer. That means every inference is now generating far more tokens. Multiply that by millions of users and you get the compute crisis NVIDIA had to solve.

Meanwhile, post-training techniques shifted from supervised fine-tuning (imitation learning) to reinforcement learning — where the model tries thousands of different approaches, fails, learns, and iterates. The compute cost of this approach is orders of magnitude higher than anything before it.

NVIDIA’s response: advance the state of the art every single year, with zero exceptions. Vera Rubin is the result.

Vera Rubin Architecture: 6 Co-Designed Chips

Total: 100 PetaFLOPS FP4 | 220 Trillion Transistors per Rack | 6 Chips Co-Designed | 15,000 Engineer-Years

1. Vera CPU

• 88 cores, 176 spatial multi-threads
• 2× performance per watt vs prior generation
• Designed for supercomputer-scale I/O

2. Rubin GPU

• 5× Blackwell FP performance
• NVFP4 Tensor Core: dynamic precision
• Only 1.6× transistors of Blackwell

3. ConnectX-9 (Networking NIC)

• 6 Tb/s scale-out bandwidth per GPU
• Co-designed with Vera CPU
• Programmable RDMA data path

4. BlueField-4 (DPU)

• Offloads storage and security from compute
• Keeps compute 100% focused on AI
• Integrated in every compute tray

5. NVLink-6 Switch

• Connects 18 compute nodes
• Scales to 72 Rubin GPUs as one unified unit
• 6 TB/s bandwidth — moves more data than the global internet

6. Spectrum-X (Ethernet Switch)

• World’s first 512-lane Ethernet switch
• 200G co-packaged optics
• Scales thousands of racks into an AI factory

The Philosophy of Extreme Co-Design

NVIDIA broke one of its own rules to build Vera Rubin. The company typically changes only one or two chips per generation — a conservative approach that limits risk and preserves compatibility. But with Vera Rubin, they redesigned every single chip from scratch.

Why? Because Moore’s Law has largely stalled. The number of transistors you can add to a chip each year has hit a ceiling. The Rubin GPU delivers 5× the performance of Blackwell with only 1.6× the transistors. That ratio — massive performance gain from a modest transistor increase — is only possible through one mechanism: co-design at every level of the stack simultaneously.

“It is impossible to keep up with those kind of rates unless we deploy aggressive, extreme co-design — innovating across all the chips, across the entire stack, all at the same time.”

— Jensen Huang, NVIDIA CEO

The star innovation enabling this leap is the NVFP4 Tensor Core — a dedicated processor, not just a data format. Unlike conventional FP4 or FP8 implementations that apply fixed precision across a model, the NVFP4 Tensor Core dynamically and adaptively adjusts its precision and structure in real time as it processes different layers of a transformer model. It maximizes throughput wherever precision can be sacrificed, then snaps back to full precision wherever accuracy is critical — all happening at hardware speeds, far too fast for software to control.

NVIDIA has already published academic papers on this technique and has signaled it may become an industry standard. It is, in their own words, “completely revolutionary.”

The Hardware: Zero Cables, 100% Liquid Cooling

The Vera Rubin system is not just about chips. The mechanical and thermal engineering is equally radical. The previous generation NVL72 rack required 43 cables and 6 tubes per compute tray, took two or more hours to assemble, and demanded skilled technicians who would often need to disassemble and reassemble multiple times before getting it right.

The new Vera Rubin compute tray: zero cables. Two tubes. Assembly time drops from over two hours to five minutes. The entire chassis is 100% liquid cooled — a mandatory requirement, not an option. With rack densities approaching and exceeding 240 kW per rack, air cooling is physically impossible.

Each MVL72 rack contains 18 compute trays, each housing 2 Vera CPUs and 4 Rubin GPUs, connected by 9 NVLink switch trays, collectively operating as a single massive compute unit. A Rubin pod scales this further — 1,152 GPUs across 16 racks, delivering compute at a scale that was science fiction five years ago.

NVIDIA GPU Generation Performance Leap (FP4 AI PetaFLOPS per Rack)

• A100 (2020): ~1.2 PetaFLOPS
• H100 (2022): ~4 PetaFLOPS
• Blackwell (2025): ~20 PetaFLOPS
• Vera Rubin (2026): 100 PetaFLOPS
• 5× vs Blackwell FP Performance
• 83× vs A100 per rack
• 10× token cost reduction per year

Note: Performance figures are indicative FP4 AI petaflops per NVL72 rack. Vera Rubin is in full production as of CES 2026.

What Vera Rubin Means for the AI Industry

The implications of Vera Rubin go far beyond raw benchmark numbers. Three structural changes will ripple across the AI industry:

1. Reinforcement Learning at Scale Becomes Viable

Post-training with reinforcement learning — the technique powering reasoning models like O1 and its successors — requires the model to attempt thousands of variations of a task autonomously. This is compute-intensive to an almost absurd degree. Vera Rubin makes it economically viable to run this at the scale required for frontier models, democratizing access to reasoning AI beyond the handful of labs that can currently afford it.

2. Open Source Models Will Eclipse Proprietary Ones

Jensen Huang made a bold prediction at CES 2026: open source models will ultimately become the largest category of AI usage, surpassing even OpenAI — today’s dominant token generator. With Vera Rubin cutting the cost of computation by roughly 10× per generation, the economics of training and serving large open models will continue to improve, putting frontier-class AI within reach of thousands of companies, researchers, and domains globally.

3. Test-Time Scaling Accelerates

As inference shifts from single-shot answering to extended reasoning chains, token generation rates are rising 5× per year. Models that “think longer” produce better outputs — and users are discovering this rapidly. The infrastructure challenge this creates is enormous: serving a reasoning model requires 5–25× the compute of serving a conventional model. Vera Rubin exists precisely to absorb this demand at scale.

Why Cyfuture Is India’s Only Vera Rubin-Ready Facility

Vera Rubin is not a chip you can drop into an existing data center. At 240 kW per rack, it demands purpose-built infrastructure that most facilities — even hyperscaler-grade ones — simply don’t have. The cooling physics alone are non-negotiable: air cooling fails above roughly 30 kW per rack. Liquid cooling at Vera Rubin densities requires direct-to-chip cold plate deployment, purpose-designed fluid distribution manifolds, and rack architecture that accommodates the fully cable-free tray design.

Cyfuture Cloud’s 10 MW facility, going live October 2026, is the only colocation data center in India engineered from the ground up for this generation of compute. The facility supports:

Cyfuture Cloud — Vera Rubin Infrastructure Checklist:

• 240 kW/rack direct-to-chip liquid cooling — already deployed
• 100% liquid cooled — no air-cooled fallback needed
• NVL72 rack form factor compatible from day one
• SEZ-enabled for import duty advantages on GPU hardware
• MeitY-empanelled for government and sovereign AI workloads
• N+1/2N redundancy across power and cooling
• 10 MW total IT capacity — from single-rack to full campus
• Modular phased design for Rubin Ultra (2027) readiness

The competitive window for Vera Rubin allocations is narrow. NVIDIA’s production is ramping, but enterprise-grade colocation at the required density in India is available from only one provider. Organizations that secure capacity blocks now will be running Vera Rubin workloads the day their GPU allocation ships — those that wait will face 6–12 month delays as infrastructure scrambles to catch up.

Conclusion: A Supercomputer for the Next Frontier

Vera Rubin — the astronomer — revealed invisible forces governing the universe. NVIDIA’s Vera Rubin — the supercomputer — reveals the invisible ceiling on AI advancement and then shatters it. Six breakthrough chips. 15,000 engineer-years. 220 trillion transistors. Zero cables. One extraordinary leap.

The AI compute race is not slowing. Models will continue to grow 10× per year. Inference will continue to expand through test-time scaling. The cost of tokens will continue to fall as competition intensifies. Every organization that wants to compete on AI — whether training foundation models, running large-scale inference, or building sovereign AI capabilities — needs the infrastructure to match.

In India, that infrastructure is Cyfuture Cloud.