NVIDIA CEO Jensen Huang took the stage at GTC Europe this week with a singular focus: transforming the timeline of the artificial intelligence revolution. In a keynote that blended high-performance hardware announcements with a strategic outreach to European industry, Huang promised to reduce the "time to token" for new data centers from months to a matter of weeks. The centerpiece of this effort is the DSX Pod, a containerized, "factory-built" version of NVIDIA’s reference architecture designed to bypass the traditional bottlenecks of on-site assembly and infrastructure delays.
The announcement comes at a pivotal moment for the industry, as the demand for generative AI compute continues to outpace the physical capacity to build and commission the massive data centers required to run it. By shifting the complexity of assembly from the construction site to the factory floor, NVIDIA aims to stabilize a global supply chain that is increasingly strained by shortages of specialized labor and the sheer physical scale of modern AI infrastructure.
The DSX Pod: A Factory-Integrated AI Solution
The hardware highlight of the keynote was the DSX Pod, which Huang described as a "factory-built integrated system" that brings the company’s full-stack platform into a transportable, modular format. The DSX Pod is essentially a containerized implementation of the Vera Rubin architecture, the successor to the company’s Blackwell generation.
According to technical specifications shared during the presentation, each DSX Pod begins with the Vera Rubin NVL72, a configuration consisting of seven liquid-cooled compute racks. These racks are interconnected by a staggering 17 kilometers of Spectrum X cabling to form a single, cohesive unit of scale-out computing. These pods are not merely designs but physical products that are integrated, fully cabled, and factory-tested before being shipped in transportable containers.
The engineering behind the pod reflects the evolving requirements of high-density AI workloads. The design utilizes dry cooling rather than traditional evaporative cooling, a move that aligns with increasing environmental regulations regarding water usage in data centers, particularly in Europe. To handle the immense power draw of the Vera Rubin chips, the pods utilize 800V DC power delivery, a choice NVIDIA says is essential for maintaining electrical efficiency at this scale.
While many of these components exist within the standard DSX reference architecture, the shift to a containerized, pre-validated model is a direct response to the deployment and supply chain challenges currently facing data center operators. Huang noted that the scale of production is now "unheard of" and "unimaginable," yet the industry remains tethered to traditional, slow construction methods.
Solving the "Electrician Bottleneck"
Huang was candid about the logistical hurdles that prevent AI capacity from coming online. He pointed out that even when the necessary GPUs and networking hardware are available, bringing these systems into a data center site often involves months of on-site assembly. The complexity of connecting 17,000 kilometers of wiring for a Spectrum X network alone is a daunting task that requires a specialized workforce.
This physical bottleneck was illustrated by the example of TeraWulf’s Lake Mariner site outside Buffalo, New York. At that facility, nearly half of the 1,800 tradespeople on site are electricians. Huang noted that the industry often runs into a shortage of qualified electricians long before it faces a shortage of GPUs or memory.
By moving the assembly, qualification, and "burn-in" processes into the supply chain, NVIDIA aims to ensure that when a pod arrives at a customer site, it is already fully functional. "The amount of cabling and on-site assembly is going to be greatly reduced," Huang said. "But, just as importantly, by moving this into the supply chain, the amount of idle computing that’s going to be sitting inside at your datacenters will reduce dramatically."
The economic argument for this speed is compelling. Huang noted that a modern 1-gigawatt (GW) data center represents a capital investment of approximately $50 billion. Operators of such facilities expect to generate roughly $40 billion in annual revenue from that infrastructure. Under those conditions, a three-month delay in deployment translates to $10 billion in missed revenue. "Even for me," Huang joked, referencing his estimated $200 billion net worth, "$10 billion worth of revenues will make anybody anxious."
The goal for the DSX Pod is to shrink the window from the arrival of hardware to the production of AI tokens from the current average of four to five months down to just one month. NVIDIA announced that it is working with a broad ecosystem of partners—including Schneider Electric, Vertiv, HPE, Lenovo, and Dell—to deliver these DSX Pod "AI factories" globally starting in early 2027.
Expanding the Global Manufacturing Footprint
Beyond the modular hardware, NVIDIA used GTC Europe to announce a significant expansion of its manufacturing ecosystem. The company is partnering with a group of major supply chain providers to increase global capacity for substrates, advanced packaging, testing, and optical engines.
Key partners in this expansion include Amkor Technology, IBIDEN, King Yuan Electronics (KYEC), Samsung Electro-Mechanics, Siliconware Precision Industries (SPIL), and Unimicron Technology Corporation. This move is designed to strengthen the resiliency and redundancy of NVIDIA’s supply chain, ensuring that the company can meet the projected demand for its full-stack AI platforms through the end of the decade.
A notable part of this strategy involves Samsung Electro-Mechanics, which is joining NVIDIA’s substrate supply chain to provide immediate capacity from its existing operations in South Korea. Furthermore, NVIDIA expects its total semiconductor testing capacity to double between 2026 and 2027, supported by major buildouts in the United States, Taiwan, and Singapore.
Optical connectivity is also receiving a massive boost. NVIDIA’s deal with SPIL involves expanding manufacturing capacity in Taiwan for optical engines used in co-packaged optics. The company is targeting a 15-fold increase in optical engine production by mid-2027, with plans to double that capacity again by the end of 2028. This aggressive expansion highlights the critical role that high-speed interconnects play in modern AI clusters, where data movement is often as much of a bottleneck as raw compute power.
Robotics and the Agile One SN
The move toward factory-integrated systems is also reflected in NVIDIA’s latest updates to its Isaac robotics platform. Huang unveiled a new reference humanoid robot, the Agile One SN, which is designed to be as "plug-and-play" as the DSX Pods.
Standing 1.4 meters tall, the Agile One SN features five-fingered hands equipped with tactile sensing for dexterous manipulation. It is powered by an onboard compute system capable of 2,070 FP4 teraflops and 128GB of unified memory, allowing it to process multimodal sensing data in real-time.
"This reference robot will be available in the first quarter of next year," Huang said. "And it will come pre-assembled with computer inside, software integrated, and it should be able to perform quite well right off the bat." The message was clear: whether it is a data center or a humanoid robot, NVIDIA is moving toward a model where hardware and software are inseparable, pre-validated, and ready for immediate deployment.
The "Death" of Moore’s Law and Environmental Friction
Throughout the keynote, Huang revisited a familiar theme: the end of Moore’s Law and the rise of accelerated computing. He argued that the traditional method of scaling performance through transistor density is no longer sufficient. Over the last decade, NVIDIA says it has improved AI performance by a factor of one million. By comparison, Huang noted that if the industry had relied solely on Moore’s Law, it would have seen only a 100x improvement over the same period.
This massive leap in performance is what NVIDIA is pitching to European companies as a solution to the continent’s energy and sovereignty concerns. Huang highlighted several European firms using DSX reference designs—specifically those utilizing liquid cooling and dry coolers—to reduce water and power consumption while helping to stabilize local power grids.
However, the vision of an AI-powered future is not without its critics. The early part of Huang’s speech was briefly disrupted by activists, reportedly from Greenpeace, who shouted about "environmental destruction." Reports also indicated that some protesters cited concerns regarding the alleged use of NVIDIA technology in Russian drones operating in Ukraine.
Huang remained composed during the interruption, asking security to "go easy" on the protesters. "All important issues," he acknowledged, before returning to his presentation. "But let me get through my keynote."
The protest underscored the growing pushback the data center industry faces across Europe, where environmental impact and digital sovereignty have become central political issues. By emphasizing the efficiency of the DSX Pod and the shift toward liquid cooling, NVIDIA is clearly attempting to position its "AI factories" as a more sustainable alternative to traditional data center builds.
As NVIDIA prepares to roll out the DSX Pods in 2027, the company is betting that the path to AI dominance lies not just in the speed of the chips, but in the speed at which those chips can be put to work. For an industry where a few months of delay can cost billions, the "factory-in-a-box" may become the new standard for the next phase of global compute expansion.
