NVIDIA Vera Rubin NVL72: The Next Peak of AI Compute and Its Impact on Domestic Chinese AI Chips

I. Introduction: The Dawn of a New AI Compute Era

On March 16, 2026, at the GTC 2026 conference in San Jose, California, NVIDIA founder and CEO Jensen Huang took the stage to unveil the next-generation AI computing system — Vera Rubin NVL72. This is no ordinary product iteration, but a technological revolution that can only be described as a "generational leap." During his speech, Huang revealed that market demand for NVIDIA's Blackwell and Vera Rubin systems would generate at least $\mathrm{<span\; style="color:\#FFFFFF;">1</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">v</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">u</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">b</span>}\mathrm{<span\; style="color:\#FFFFFF;">y</span>}\mathrm{<span\; style="color:\#FFFFFF;">2027</span>}<span\; style="color:\#FFFFFF;">,</span>\mathrm{<span\; style="color:\#FFFFFF;">d</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">u</span>}\mathrm{<span\; style="color:\#FFFFFF;">b</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">g</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">p</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">v</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">u</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">f</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">f</span>}$500 billion for 2026. Behind this astonishing prediction lies the unprecedented technical power and disruptive potential of Vera Rubin NVL72.

This article will comprehensively analyze this "AI supercomputer" from multiple dimensions, including technical specifications, design philosophy, impact on the global compute industry, and challenges to domestic Chinese AI chips.

II. Technical Details of Vera Rubin NVL72

2.1 A Seven-Chip Integrated System Architecture

Vera Rubin is not a single chip, but a complete AI supercomputer platform composed of seven specialized chips. These seven chips are: Rubin GPU, Vera CPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch, plus a newly added Groq 3 LPU inference chip. NVIDIA emphasizes this as "seven new chips, building one AI supercomputer" — the ultimate in collaborative design.

In the Vera Rubin NVL72 rack-level configuration, a single rack integrates 72 Rubin GPUs, 36 Vera CPUs, and 18 BlueField-4 DPUs. The entire system contains approximately 220 trillion transistors, with about 1.3 million components supplied by over 80 suppliers from more than 20 countries and regions, including China, Vietnam, and Thailand.

2.2 Extraordinary Compute Specifications

The raw numbers convey the sheer power of Vera Rubin NVL72: the system delivers 3.6 EFLOPS of FP4 inference compute and 2.5 EFLOPS of FP8 training compute. More specifically, the Rubin GPU is built on TSMC's 3nm process, integrates 336 billion transistors, and features 288 GB of HBM4 memory with 22 TB/s of bandwidth. A single Rubin GPU delivers 50 PFLOPS of FP4 inference compute — five times that of the Blackwell architecture — and 35 PFLOPS of training compute, 3.5 times that of Blackwell.

At the system level, the Vera Rubin NVL72 rack achieves up to 260 TB/s of system bandwidth via NVLink 6, surpassing the entire global internet combined. Each GPU has 3.6 TB/s of fully interconnected bandwidth, ensuring that the 72 GPUs work together seamlessly as a single unit.

Memory has also seen a generational leap. Each Rubin GPU is equipped with 288 GB of HBM4 memory, bringing the total system HBM4 memory to 20.7 TB. Each Vera CPU comes with 1.5 TB of LPDDR5X memory, for a total of 54 TB of LPDDR5X memory per rack. The system also introduces a shared context memory pool of 150 TB for the first time, breaking through memory bottlenecks for long-context inference scenarios.

2.3 Revolutionary Energy Efficiency

While the performance gains are impressive, the leap in energy efficiency is equally stunning. The Vera Rubin system consumes about twice the power of the previous Blackwell architecture, but delivers 10x the performance per watt — a qualitative improvement in overall energy efficiency. This is NVIDIA's first AI infrastructure to adopt 100% liquid cooling, and the company recommends that future AI factories generally adopt closed-loop liquid cooling to improve heat dissipation and conserve water.

Compared to the previous Blackwell generation, Vera Rubin reduces the cost per token to one-tenth, increases inference throughput per watt by 10x, and requires only one-quarter the GPU resources to train a mixture-of-experts large model. This means that for the same AI task, customers pay less and consume less energy — not just a technical victory, but a profound commercial transformation.

2.4 Innovative Hierarchical Inference Architecture

Another major architectural innovation of Vera Rubin NVL72 is its disaggregated compute architecture. For inference tasks, the system implements hierarchical inference through the Groq 3 LPX dedicated rack, where "GPU handles Prefill, LPU handles Decode." Prefill refers to the model processing the entire input at once to generate key-value caches, while Decode generates tokens word by word. The Groq 3 LPU chip is dedicated to the decode stage, working in tandem with the GPU to achieve 35x higher inference throughput per watt and, for the first time on the Pareto curve, unity of high throughput and low latency.

This division effectively solves the inference latency problems brought by trillion-parameter models and million-token contexts, providing unprecedented hardware support for cutting-edge workloads such as agentic AI and long-context inference.

III. Far-Reaching Impact on the Global Compute Industry

3.1 Reshaping the Cost Structure of Compute

The most direct impact of Vera Rubin NVL72 is how it redefines the economics of compute. Reducing the cost per token to one-tenth that of Blackwell dramatically lowers the barrier to scaling AI applications. For cloud service providers, the same investment can serve more users or offer equivalent service at lower prices — profound implications for the entire AI application ecosystem.

From an industry chain perspective, the ramp of Vera Rubin systems will drive demand across multiple segments, including co-packaged optics, optical modules, and liquid cooling. The system adopts Spectrum-X Ethernet CPO technology, which delivers 5x better power efficiency and 10x greater network resilience compared to traditional networking. Jensen Huang explicitly laid out a "copper + optics" roadmap: short-reach intra-rack connections stick with the cost-effectiveness and reliability of copper, while long-range inter-rack connections rely on optics. This roadmap will deeply influence investment directions across the optical communications industry.

3.2 Accelerating the Shift from Training to Inference

The launch of Vera Rubin marks a turning point where the center of gravity in the AI industry moves decisively from the expensive model-training phase to the inference and large-scale production era. Jensen Huang made clear that the next inflection point in AI will significantly boost demand for accelerated computing, software, and AI factories. Lenovo has become a global launch partner for Vera Rubin NVL72, and Lenovo CEO Yang Yuanqing stated that as agentic AI drives exponential growth in inference workloads, cost control and per-token performance will become critically important.

This means future competition will no longer be a simple race to pile up raw compute, but a holistic battle involving "cost per token" and "performance per watt." The advantage NVIDIA has built on this front will be difficult for any competitor to challenge in the near term.

3.3 Widening the Technology Gap

Looking at the historical roadmap, NVIDIA's iteration cadence is breathtaking: Blackwell in 2024, Blackwell Ultra in 2025, Vera Rubin in 2026, and Vera Rubin Ultra (featuring 144 GPUs) in 2027. Huang even previewed the even more distant Feynman architecture, which will use custom high-bandwidth memory.

This rapid iteration itself is a formidable competitive moat. Competitors are still struggling to catch up to Blackwell, while NVIDIA has already launched Vera Rubin, which outperforms Blackwell by multiples. Jensen Huang claimed that Vera Rubin represents a 40 million-fold increase in computing power over 10 years — the implication being that even if a competitor doubled its performance every year, it would take 22 years just to reach that starting line.

IV. The Impact on Domestic Chinese AI Chips

4.1 The Harsh Reality of the Performance Gap

The arrival of Vera Rubin NVL72 poses a severe challenge to China's domestic AI chip industry, which is currently in a catch-up phase. The first reality that must be faced is that Vera Rubin and the more advanced Blackwell family remain under U.S. export controls to China. The most advanced NVIDIA chip that China can legally obtain today is the H200 (whose performance is roughly one-third that of the B200), while Vera Rubin's inference performance is 5x that of Blackwell.

This means Chinese AI companies and research institutions face an increasingly high "ceiling" when it comes to access to cutting-edge compute. The H200's overall performance still leads most domestic AI chips — for example, the overall processing performance of Huawei's Ascend 910C reaches only about 76% of the H200, and its memory bandwidth is roughly two-thirds that of the H200. The performance gap between Vera Rubin and the H200 is then a matter of orders of magnitude.

4.2 Rapid Catch-up by Domestic Chips

Despite the significant gap, the pace of catch-up by domestic compute should not be underestimated. In 2026, Huawei launched the Ascend 950PR chip, whose FP4 performance reached 2.87 times that of NVIDIA's H20, shattering the stereotype that "domestic chips are inferior to imports." Huawei also plans to release the Atlas 950 supernode in the fourth quarter of 2026, supporting 8,192 Ascend 950DT chips — 56.8 times the scale of NVIDIA's NVL144 and 6.7 times the total compute power.

In terms of market share, domestic AI chips have already made breakthroughs. Huawei Ascend shipped approximately 812,000 units in 2025, accounting for about 20% of the domestic AI accelerator card market. Multiple domestic chip manufacturers (including Huawei Ascend, Cambricon, Hygon, Moore Threads, and M-Xing) have formed a tiered competitive landscape, achieving scaled substitution in both training and inference scenarios. Some market research reports even indicate that with breakthroughs in original architectures, the market share of domestic GPUs has exceeded 60%, while NVIDIA's share in the Chinese market has dropped from a peak of 95% to 8%.

4.3 The Double-Edged Sword of Export Controls

The policy trajectory of U.S. export controls on chips to China warrants close attention. In December 2025, the Trump administration approved NVIDIA to export H200 chips to China, but with the onerous condition that 25% of revenue be paid to the U.S. government. This adjustment is essentially a "dual calculation of technology gap control and economic benefit extraction" — the more advanced Blackwell and Rubin chips remain subject to the export ban.

Such "cream-skimming" controls are both a challenge and an opportunity for China. Some analysts suggest that the introduction of the H200 may actually accelerate domestic chip autonomy: a tangible NVIDIA product to benchmark provides clearer, more realistic performance targets for domestic manufacturers, helping to shorten technology feedback cycles and drive more aggressive performance goals. More critically, China's government procurement lists have clearly favored domestic AI chip vendors such as Huawei and Cambricon, using policy power to create market space for domestic substitution.

4.4 Paths for Chinese Compute to Break Through

Facing the performance shock of Vera Rubin NVL72, China's compute industry is exploring multiple breakthrough paths:

First, system-level competition instead of single-chip competition. With lagging process technology, it is difficult to close the gap quickly on a single-chip basis, but achieving greater overall performance through supernode architectures is a feasible route. Huawei's Atlas 950 supernode exemplifies this approach: using tens of thousands of cards in coordinated systems to achieve total compute power far exceeding that of a single rack.

Second, differentiated market positioning. Domestic chips are already competitive in high-throughput inference scenarios, where they can build advantages in cost-sensitive, large-scale inference markets, leaving the capital-intensive training frontier to NVIDIA's leading chips, forming a temporary "coexistence equilibrium."

Third, diversified technical routes. Domestic companies like ZOOM.ai (SinoGist) focus on TPU routes, which offer high energy efficiency for specific AI compute scenarios and can complement general-purpose GPUs. Running multiple routes in parallel helps reduce dependence on any single technology.

V. Outlook: Reshaping the Compute Landscape

The launch of Vera Rubin NVL72 marks a new phase in the global AI compute race. On one hand, NVIDIA, leveraging its technology gap and cost advantages, is rapidly building a full-stack barrier from chip to system, from hardware to ecosystem. For most competitors, Vera Rubin represents a summit that is difficult to scale.

On the other hand, China's compute industry is advancing its autonomy with unprecedented determination and speed. While continued export controls have created short-term difficulties in accessing technology, they have also opened up vast market space for domestic chips. Domestic compute chip shipments are expected to more than double in 2026, and the global AI compute landscape is shifting from single dominance by NVIDIA to a multi-track, parallel structure.

What is foreseeable is that future AI compute competition will be not only a battle of hardware performance, but a comprehensive contest of system architecture, energy efficiency optimization, cost control, and ecosystem development. Vera Rubin NVL72, with its philosophy of "10x efficiency, 1/10 cost," has set a new benchmark in this arena. Global competitors, including China, will redefine their positions and roles in this magnificent technological revolution.

Technology blockade will ultimately become the "whetstone" for independent innovation — the greater the challenge, the stronger the momentum for breakthrough. This is a law repeatedly validated by history, and it should also be the clear-eyed awareness and firm confidence of China's compute industry in the face of the Vera Rubin shock.