NVIDIA Unveils Rubin Ultra GPU at Computex 2026

NVIDIA CEO Jensen Huang took the stage at Computex 2026 in Taipei to unveil the company's next-generation Rubin Ultra GPU architecture, promising a massive leap in AI training and inference performance. The new architecture delivers up to 3x the performance of its predecessor, Blackwell Ultra, and introduces support for HBM4 memory — positioning NVIDIA to maintain its stranglehold on the $300 billion AI accelerator market.

The announcement, made during Huang's signature leather-jacket keynote, drew roaring applause from an audience of thousands of developers, enterprise leaders, and hardware partners. Rubin Ultra represents the culmination of NVIDIA's 2-year architecture cadence and arrives as demand for AI compute continues to outstrip supply across hyperscale data centers worldwide.

Key Takeaways From the Rubin Ultra Announcement

• 3x AI training performance over Blackwell Ultra on large language model workloads
• HBM4 memory support delivering up to 512 GB of memory per GPU with 12 TB/s bandwidth
• New NVLink 6 interconnect enabling 3.6 TB/s GPU-to-GPU communication
• 4nm enhanced process node manufactured by TSMC using advanced CoWoS-L packaging
• Rubin Ultra NVL144 rack-scale system combining 144 GPUs in a single liquid-cooled rack
• Expected availability in Q1 2027 for hyperscale partners, with broader availability by mid-2027

Rubin Ultra Pushes the Boundaries of AI Compute

The Rubin Ultra architecture marks NVIDIA's most ambitious chip design to date. Each GPU packs an estimated 208 billion transistors — nearly double the transistor count of Blackwell Ultra's 104 billion. NVIDIA achieves this through a multi-die chiplet design that stitches together multiple compute dies using TSMC's latest CoWoS-L advanced packaging technology.

At the heart of Rubin Ultra sits a redesigned Tensor Core engine, now in its 7th generation. These new cores introduce native support for FP4 precision at full throughput, enabling AI models to train and run inference with dramatically reduced memory footprints. Huang claimed that a single Rubin Ultra GPU can deliver over 2,500 TFLOPS of FP4 performance, compared to roughly 800 TFLOPS on Blackwell Ultra.

The memory subsystem represents an equally significant upgrade. Rubin Ultra is the first NVIDIA GPU to support HBM4, the latest generation of high-bandwidth memory co-developed with SK Hynix and Samsung. Each GPU module features 8 HBM4 stacks, providing 512 GB of total memory capacity and a staggering 12 TB/s of memory bandwidth — a 2.4x improvement over Blackwell Ultra's 5 TB/s.

NVLink 6 and the Rubin Ultra NVL144 System

Connectivity has always been a differentiator for NVIDIA's data center GPUs, and Rubin Ultra takes this to an entirely new level. The new NVLink 6 interconnect provides 3.6 TB/s of bidirectional bandwidth between GPUs, effectively allowing multiple GPUs to operate as a single unified compute fabric.

NVIDIA introduced the Rubin Ultra NVL144, a rack-scale AI supercomputer that combines 144 Rubin Ultra GPUs into a single, liquid-cooled system. This configuration delivers an aggregate of 73,728 GB of HBM4 memory and what NVIDIA claims is enough compute to train a 50-trillion-parameter model. For context, the Blackwell-based GB200 NVL72 system — which only began shipping at scale in late 2025 — supported 72 GPUs per rack.

'Every AI factory in the world will want to upgrade to Rubin Ultra,' Huang declared on stage. 'We are building the engine of the next industrial revolution.'

The NVL144 system is designed to integrate seamlessly with NVIDIA's Spectrum-X networking platform and the company's CUDA 14 software stack. NVIDIA emphasized backward compatibility, noting that existing CUDA applications will run on Rubin Ultra without modification, though developers can unlock additional performance through new APIs.

How Rubin Ultra Compares to the Competition

NVIDIA's announcement comes amid intensifying competition in the AI accelerator space. AMD has been gaining traction with its Instinct MI450 series, which offers competitive performance on inference workloads at lower price points. Intel recently launched its Falcon Shores GPU, targeting enterprise AI deployments. Meanwhile, custom silicon from cloud giants — including Google's TPU v7, Amazon's Trainium 3, and Microsoft's Maia 2 — continues to chip away at NVIDIA's dominance in specific workloads.

However, NVIDIA's ecosystem advantage remains formidable. Key competitive differentiators include:

• CUDA ecosystem: Over 5 million developers and 1,000+ accelerated applications
• Full-stack integration: Hardware, networking, software, and frameworks tightly coupled
• NVLink superiority: No competitor currently matches NVLink 6's GPU-to-GPU bandwidth
• Supply chain partnerships: Deep relationships with TSMC, SK Hynix, and major OEMs
• Software moat: Frameworks like TensorRT, Triton, and NeMo optimized for NVIDIA hardware

Analysts at Morgan Stanley estimate that NVIDIA still commands approximately 80% of the AI training accelerator market, though that share has dipped from 90% two years ago. The Rubin Ultra launch is widely seen as NVIDIA's effort to reassert technological dominance and justify premium pricing.

What This Means for Developers and Enterprises

For AI developers and enterprise buyers, Rubin Ultra's arrival carries significant practical implications. The 3x performance improvement means that training runs that previously required weeks on Blackwell clusters could potentially complete in days on equivalent Rubin Ultra configurations. This directly translates to faster iteration cycles and lower total cost of ownership for AI labs.

The 512 GB HBM4 memory per GPU is particularly noteworthy. As frontier models grow beyond 1 trillion parameters, memory capacity has become a critical bottleneck. Rubin Ultra's expanded memory allows larger model shards to fit on a single GPU, reducing the communication overhead that degrades performance in distributed training.

NVIDIA also announced that DGX Cloud — its cloud-hosted AI supercomputing service — will offer Rubin Ultra instances starting in Q2 2027. Pricing was not disclosed, but Huang suggested that performance-per-dollar would improve by approximately 2x compared to current Blackwell-based DGX Cloud offerings. Major cloud providers including AWS, Microsoft Azure, and Google Cloud are expected to offer Rubin Ultra instances shortly after launch.

For startups and smaller AI companies, the improved inference efficiency of Rubin Ultra could be transformative. NVIDIA demonstrated a single Rubin Ultra GPU serving a 400-billion-parameter language model at over 200 tokens per second — performance that previously required a multi-GPU setup.

NVIDIA's Roadmap Signals Relentless Pace of Innovation

Huang also provided a glimpse into NVIDIA's future roadmap beyond Rubin Ultra. The company confirmed that its next architecture, codenamed Feynman, is already in development and expected to arrive in 2028. Feynman will reportedly move to a 3nm process node and introduce support for HBM4e memory.

This aggressive cadence — new architectures roughly every 12 to 18 months — reflects the breakneck pace of AI advancement. NVIDIA's roadmap progression tells the story:

• Hopper (2022): Introduced Transformer Engine, FP8 support
• Blackwell (2024): 208B transistors (dual-die), NVLink 5, FP4 introduction
• Blackwell Ultra (2025): Enhanced Blackwell with HBM3e, improved efficiency
• Rubin (2026): HBM4 support, NVLink 6, new Tensor Core architecture
• Rubin Ultra (2026): Full HBM4 integration, NVL144, 3x Blackwell Ultra performance
• Feynman (2028): 3nm process, HBM4e, next-gen NVLink

NVIDIA's stock rose approximately 4% in after-hours trading following the keynote, pushing its market capitalization further above the $5 trillion mark. The company's data center revenue — which now accounts for over 85% of total revenue — is expected to receive a significant boost as Rubin Ultra systems begin shipping.

Looking Ahead: The AI Infrastructure Arms Race Intensifies

The Rubin Ultra announcement underscores a broader truth about the AI industry in 2026: the appetite for compute shows no signs of slowing. Hyperscalers are collectively spending over $300 billion annually on AI infrastructure, and sovereign AI initiatives from countries including Saudi Arabia, France, Japan, and India are adding billions more in demand.

NVIDIA's challenge will be execution. The Blackwell generation faced well-documented supply constraints and thermal management issues during its initial rollout. Huang acknowledged these lessons on stage, noting that NVIDIA has worked closely with TSMC and its packaging partners to ensure smoother production ramps for Rubin Ultra.

For the broader AI ecosystem, Rubin Ultra represents another step toward making previously impossible AI workloads routine. Models with tens of trillions of parameters, real-time multimodal reasoning systems, and AI agents capable of complex autonomous tasks — all of these become more feasible with each generation of hardware improvement.

The message from Taipei was clear: NVIDIA intends to remain the dominant force in AI computing, and Rubin Ultra is its latest and most powerful argument for that position.