Jensen Huang at Berlin GTC: AI Factory Deployment Shortened to as Fast as 14 Days, OpenShell Agent Security Architecture Released

Article body EM_StockImg_Start

EM_StockImg_End

On October 21 local time, Nvidia CEO Jensen Huang delivered a speech at the GTC conference in Berlin, highlighting a system architecture for next-generation AI infrastructure, a modular AI factory solution, and a security protection system built for autonomous agents.

Huang reviewed Nvidia's technological evolution over the past decade and more. He said that over the past 11 years, through "extreme co-design," the company has improved AI performance by a factor of 1 million.

Huang noted that the AI chip platform Vera Rubin is now "in full production," with each node taking less than 5 minutes to assemble, compared with 2 hours for the previous-generation Grace Blackwell. At the same time, Vera Rubin requires no cables and uses a 100% liquid-cooled design, meaning its assembly speed can be more than 24 times faster.

Huang shows the Vera Rubin compute node. Source: livestream

Vera Rubin has also been optimized for the complete agent workflow. The company said that under agent workloads, the per-megawatt throughput of the Vera Rubin NVL72 system can reach 30 times that of the previous-generation Blackwell platform.

However, chip performance is only part of AI infrastructure. In his speech, Huang stressed that "an AI factory only creates value once it starts producing tokens." He said that a 1-gigawatt (GW) AI factory costs about $50 billion, involving land, power, buildings, computing, networking, storage and cooling, among other elements.

By calculation, such an AI factory could generate about $40 billion in revenue per year, and a three-month deployment delay could correspond to about $10 billion in lost revenue.

To shorten the cycle from construction to actually generating computing revenue for AI infrastructure, Nvidia on that day launched the new DSX POD AI factory solution. According to the introduction, a DSX POD solution includes seven liquid-cooled computer racks and about 17 kilometers of Spectrum-X cabling, using 800-volt DC power. The relevant systems will be integrated, cabled and tested before delivery, and upon arrival at the site will connect to power, cooling and networking facilities that are already in place.

Huang said that when on-site supporting conditions are already available, DSX POD is expected to shorten the on-site deployment cycle from about six months to as fast as 14 days. The solution is planned to begin delivery by global partners in early 2027. Moving more integration and testing work forward to the factory can not only shorten construction time, but also is expected to reduce the uncertainty of on-site construction and the costs caused by idle computing resources.

At this Berlin GTC conference, Huang also "called out" to Europe, urging countries to accelerate investment in AI: "For Europe to benefit from AI and truly build an AI ecosystem and AI economy, all end industries need to embrace AI... every industry must transform into an AI company."

According to the introduction, in the previous week, Nvidia held events in multiple European cities; the company has invested nearly $10 billion in Europe's technology ecosystem and works with more than 12,000 partners.

Agents drive simultaneous upgrades in computing power and security demand

In addition, Nvidia announced that day that the world's leading telecom operators are transforming their networks into distributed intelligent platforms, bringing Nvidia accelerated computing into centralized data centers, metro sites and cellular sites.

Agents were another major focus of Huang's speech. Huang explained: "Because agents can think, reason, iterate and use tools, they consume massive amounts of tokens, and demand for GPUs is thus experiencing explosive growth."

Huang cited OpenRouter data showing that token demand from agents on the platform has surged. Source: livestream

However, as the scope of what agents can operate grows, security risks have extended from generating erroneous content to unauthorized data access and system operations.

Huang pointedly noted: "Today, the only way to prevent an agent from abusing permissions is actually to politely ask it: 'Please don't delete anything, please don't touch my SSH keys.' This kind of protection is called 'do not disturb.' But prayer is not security; security is not a vision, it is a technical and engineering problem."

To address this issue, Nvidia released the OpenShell reference architecture for enterprise deployment, providing runtime isolation and behavior monitoring for enterprise-grade agents. The architecture integrates Vera Rubin infrastructure, Vera CPU, the BlueField-4 data processing unit (DPU) and OpenShell secure runtime software. According to the introduction, Anthropic, Amazon Web Services, Microsoft, Google Cloud, OpenAI and other companies are participating in cooperation on this security architecture.

At the model level, Huang stressed that in the future enterprises will not rely on a single model alone, but will simultaneously use closed-source models, open-weight models and models customized based on their own data according to task needs. He also particularly stressed that enterprise knowledge will become an important factor in improving the actual working capabilities of agents.

Eastmoney infographic · extra insights

(Article source: The Paper) Body middle embedded advertisement