[Edaily Reporter Yun Junghoon ] Vessel AI, an artificial intelligence (AI) infrastructure company, will introduce NVIDIA’s next-generation AI computing system, “Vera Rubin NVL72,” to the South Korean market in the first half of next year. The company’s strategy is to accelerate its expansion into the global NeoCloud market by upgrading its existing Blackwell-based GPU clusters to the next-generation system and integrating its proprietary distributed GPU operation and inference optimization technologies.

At the “NeoCloud Strategy Presentation” held on the 15th at DreamPlus in Gangnam, Seoul, Vessel AI announced plans to build a Vera Rubin NVL72-based GPU cluster and launch full-scale services in the first half of 2027.

NVIDIA’s Vera Rubin NVL72 is a next-generation rack-scale computing system that connects 72 Rubin GPUs and 36 Vera CPUs via a high-speed interconnect and is equipped with HBM4 high-bandwidth memory. It is regarded as a key infrastructure for addressing the surging demand for inference and training driven by the proliferation of Agent AI agents.

Immediate Demand for GPUs Alone Reaches 30,000 Units… “Sells Out Within 1–2 Weeks of Securing Inventory”

The explosive demand in the AI market is driving Vessel AI to accelerate its business transformation and secure next-generation infrastructure. According to Vessel AI’s current estimates, customer demand for GPUs has already surpassed 30,000 units, with overseas demand accounting for 70–80% of that total.

Ahn Jae-man, CEO of Vessel AI, explained, “Even in Korea, there are quite a few companies requesting not just hundreds, but thousands of GPUs,” adding, “Compared to Korea, the U.S. market is at least tens of times larger and could be nearly 100 times larger.”

He added, “Demand for GPUs is so strong right now that once the schedule for securing new GPUs is finalized, sales are typically sold out within one to two weeks.”

Vessel AI is staking its future on expanding its infrastructure to meet this surge in demand. The company plans to significantly increase its GPU inventory—currently at about 7,000 units—to 10,000 by the end of this year, 50,000 by 2027, and 200,000 by 2030.

Fundraising to support these efforts is also proceeding smoothly. The company is currently securing 400 billion won in GPU financing and plans to raise an additional 2 trillion won next year in collaboration with global investment banks (IBs) and major domestic securities firms. Furthermore, it aims to secure a 500 MW domestic AI data center by 2028.

Technological Moats Such as ‘Fluid Computing’… “Win-Win Cooperation with Domestic Telecom Operators and Cloud Providers”

Vessel AI’s core competitive advantage lies not merely in securing hardware, but in its “Fluid Computing” technology, which operates distributed GPU resources as a single computing environment. It centrally manages GPUs scattered across different data centers and automatically allocates optimal resources for each task, thereby reducing GPU idle time and maximizing utilization rates.

Jeon Ji-hwan, Chief Technology Officer (CTO) of Vessel AI, stated, “GPUs are prone to failure; Meta’s Llama reportedly experiences a failure every 30 seconds while operating clusters of tens of thousands of GPUs for model training.” He added, “Using Fluid Computing technology, we can reduce the manual labor required for GPU operations by about 30% and proactively detect more than 80% of failure signs in advance.”

Achievements in inference optimization are also becoming evident. By applying its proprietary technology, the company has roughly doubled the cache hit rate of AI models and improved inference processing performance for certain workloads by up to nearly three times. It is also participating in Upstage’s “Solar Pro” AI model inference service, reliably processing approximately 145 billion tokens per day.

The company is also strengthening its partnerships with major domestic tech firms and telecommunications carriers. It is currently collaborating with SKTelecom(017670)on a project involving 1,000 NVIDIA B200 GPUs and is also building a B200-based cluster with NHN Cloud.

CEO Ahn placed particular emphasis on the company’s partnerships with domestic operators. “We are not necessarily in a competitive relationship with domestic telecom carriers or cloud companies,” he explained. “We collaborate with cloud service providers on data center and hardware operations, while Vessel AI leverages the large-scale GPU operations software capabilities we’ve accumulated over the past five to six years to divide roles and create synergy.”

He continued, “What customers are purchasing is not simply physical GPUs, but stable and fully integrated large-scale training and inference computing,” adding, “Based on our technological moat in fault prediction and resource optimization, we will leap forward to become a top-tier global neo-cloud provider representing Korea.”