Samsung SDS Plans Vera Rubin GPU Cloud for March 2027, 800 MW by 2031

Samsung SDS plans to offer GPU-as-a-service on NVIDIA's Vera Rubin GPUs from March 2027, CEO Lee Joon-hee said on the company's third-quarter earnings call on Oct. 29, according to The Korea Herald.

NVIDIA's B300 comes first. Lee said Samsung SDS will start B300 GPU-as-a-service (GPUaaS) in the west wing of its Dongtan data center in Hwaseong, Gyeonggi Province, by the end of 2026, after the Ministry of Science and ICT selected it to build and run GPU infrastructure. EDAILY reported that the Vera Rubin service will also launch at Dongtan. Samsung SDS, the Samsung group's IT services and cloud company, says it wants to be the first cloud provider in Asia to offer Vera Rubin. That is the company's own claim, and the reports read carry no NVIDIA confirmation.

From Dongtan to Haenam and 800 MW

Lee also set out a stopgap. From April 2027, part of the Dongtan west wing will host an interim service to absorb demand for AI computing until the national AI computing center in Haenam, South Jeolla Province, opens in 2028, The Korea Herald reported. Further out, Samsung SDS is aiming for more than 800 MW of AI infrastructure capacity by 2031, built three ways: direct investment, equity stakes, and a design-build-operate (DBO) model in which it designs, builds and runs data centers for customers.

That target makes a cloud provider a builder of grid-scale load. One piece is already lined up. EDAILY reported that Samsung SDS is the preferred bidder to design, build and operate a data center of more than 200 MW in the Seoul metropolitan area, led by Pacific Asset Management, with construction starting as early as the end of 2026 and operations phased in from 2029. By Voltgrid's arithmetic, a project of more than 200 MW would be roughly a quarter of the more-than-800 MW goal. For scale, Naver plans to run 55 MW in the first half of 2027 and 200 MW in 2028 under its up to $9 billion Brookfield deal. Samsung SDS gave no power figure for the Dongtan halls.

The services sit inside a business whose profit is growing more slowly than its revenue. Third-quarter revenue rose 11.2% from a year earlier to 3.7725 trillion won and operating profit rose 3.7% to 241 billion won, Seoul Economic Daily reported. Cloud revenue grew 18% to 796.1 billion won, per EDAILY. The company is targeting more than 1 trillion won ($748 million) in operating profit in 2027, according to The Korea Herald.

Where the government GPUs fit

The Oct. 29 call did not say whose chips the March service will run on. As earlier context: on June 8 the ministry picked Naver Cloud, Samsung SDS and Elice Group for a 2.08 trillion won project to secure 9,704 GPUs, 2,016 Vera Rubin and 7,688 B300, Nocut News reported at the time. Samsung SDS's share was 1,008 Vera Rubin and 2,016 B300 units; Naver Cloud's was 1,008 Vera Rubin and 3,112 B300, and Elice took 2,560 B300. The ministry said then that B300 services would start within 2026 and Vera Rubin services would roll out in the first half of 2027. Whether the March service uses those allocated units or chips Samsung SDS buys itself has not been stated. Nocut News reported that the ministry earmarked all 2,016 Vera Rubin units for national AI model development and national AI projects, and set aside 3,328 B300 units for the companies' own cloud services, so the allocation does not on its face describe a commercial Rubin service.

Vera Rubin is NVIDIA's step after the Blackwell Ultra generation that includes the B300. NVIDIA lists 270 GB of HBM3E per GPU for the HGX B300, and 288 GB of HBM4, a newer high-bandwidth memory, per Rubin GPU in its Vera Rubin NVL72 rack, which holds 72 Rubin GPUs and 36 Vera CPUs.

The power is set per rack, not per chip. NVIDIA rates the NVL72 at 227 kW in its maximum-performance mode and 187 kW in MaxQ mode, so the rack count behind Samsung SDS's March service will decide the Dongtan load.

What is not confirmed yet

The Asia-first claim is open to challenge. Naver Cloud holds the same number of Vera Rubin units under the same program, and Google Cloud said its A5X instances on Vera Rubin NVL72 reach general availability in November, without naming regions. Samsung SDS has not disclosed the rack configuration, GPU count or power draw behind the March service. The first checkpoint is the B300 go-live at Dongtan, which Lee said would come by the end of this year.