The practical significance is clear: researchers who win allocations may be able to run training and inference jobs on current-generation accelerators without buying servers, securing datacenter power, or waiting through the global GPU supply chain. But the public record shows a narrower and less settled arrangement than the “national AI research” framing suggests. Samsung SDS is one of two selected suppliers, not the sole builder or operator of a national AI computing center, and the precise quantity of Blackwell GPUs researchers will receive has not been disclosed.
ZDNet Korea reported in July that Samsung SDS and Elice Group were selected as negotiating partners for the project, ahead of AWS Korea. Digital Daily separately reported that Samsung SDS ranked first in the technical evaluation, with Elice second, and that both completed negotiations under a multi-supplier structure. For researchers and IT administrators, that means the service is a brokered public cloud allocation program: the available hardware, queueing experience, support process, and usable software environment may differ materially between the two providers.
The procurement record sets a much more specific baseline
The Korea Association of ICT, or KAIT, published the original procurement documentation in March for “GPU Resource Supply and Maintenance” under the 2026 AI Research Computing Support Project. The document set an initial budget of 15.3 billion won, including VAT, and called for at least 960 high-performance GPUs to be provided to researchers for up to eight months.
That original tender also required the supplier to prepare the environment, assist with data uploads, and provide researcher training before resources were made available. It aimed for a May service start and listed the supplier contract period as running from signing through December 31, 2026.
The schedule did not hold. A later KETI notice shows that the supplier recruitment was reopened on May 4 and closed on May 20. Digital Daily reported that the first March solicitation received no bids, delaying the program, and that the second round reduced the budget from 15 billion won to 7.5 billion won.
That creates a material discrepancy in the current reporting. Chosunbiz describes the project as worth approximately 15.3 billion won and says the service period runs from July 2026 to March 2027. The original KAIT procurement record supports the 15.3 billion won figure, but it identifies December 31, 2026 as the end of the supplier contract. Digital Daily, meanwhile, says the rebid cut the available budget in half and retained the December 31 end date.
Neither Samsung SDS nor KAIT has publicly explained whether the 15.3 billion won figure remains the executed contract value, whether 7.5 billion won reflects the final award amount, or how a service period extending into March 2027 fits with the procurement record’s December deadline. Until those terms are published, readers should treat the larger number as the original program budget rather than a confirmed measure of Samsung SDS’s present contract.
B300 availability is real, but the allocation is not yet public
Samsung SDS’s claim that it can supply B300-based capacity is credible in the narrow sense. The company announced its Samsung Cloud Platform B300 GPU-as-a-Service offering on March 23, describing it as South Korea’s first B300 cloud service. Samsung SDS said then that its B300 systems provide 288GB of HBM3E memory and 8TB/s of memory bandwidth per GPU, specifications that make the hardware particularly relevant to memory-intensive large-language-model inference.
That commercial product announcement is not the same thing as proof that a defined number of B300 GPUs is already assigned to the public research project. Chosunbiz says the service portfolio ranges from H100 to B300-based clusters, but Samsung SDS has not published the number of H100s, B300s, nodes, or total GPU-hours reserved for participating research teams. It has also not said whether users will choose a GPU generation directly, receive systems according to workload eligibility, or be placed into shared queues.
This distinction matters more than the product label. A Blackwell-capable cloud catalog does not guarantee Blackwell allocations for every approved project, particularly in a public program where the provider can pool resources and distribute them according to demand. Digital Daily reported that the final number of GPUs to be deployed was not fixed at the time of contracting and would depend on researcher demand and budget consumption.
The tender framework also makes clear that networking and cluster design are part of the award, rather than optional extras. Digital Daily reported that suppliers had to propose H100-class or better resources, GPU groupings in eight-GPU server units, multi-GPU clustering, and high-speed interconnect plans such as NVLink or InfiniBand. Those are essential requirements for distributed training, where the fabric between accelerators can determine whether a nominally powerful cluster delivers usable scaling.
Samsung SDS has not published the actual topology, interconnect bandwidth, storage performance, scheduler, container stack, supported frameworks, or job limits for the government project. For AI research groups, those omissions are more operationally important than the presence of B300 in a press release. A cluster’s real utility depends on whether a team can reserve a coherent multi-node allocation, bring its preferred PyTorch or JAX environment, move datasets quickly, and retain enough runtime to finish an experiment.
This is a cloud-access program, not a completed national supercomputer
The public material accompanying the announcement includes a rendering of the National AI Computing Center, but the 2026 AI Research Computing Support Project is a separate procurement for cloud GPU capacity and operations. The KAIT documentation describes a program designed to provide domestic researchers with cloud-based AI computing resources and an associated research environment. It does not describe the construction, ownership, or operation of a new national datacenter.
That difference affects what institutions should expect. This program lowers the barrier to using large GPU installations, but it does not give a university or government lab a permanent local cluster. Researchers receive time-limited access through a provider-managed environment, with the provider retaining control of hardware operations, upgrades, network configuration, and the terms under which resources can be reallocated.
It is also a competitive reset after AWS held the corresponding 2025 project. Electronic Times reported last year that AWS became the preferred bidder for the then-new AI research computing program shortly after obtaining South Korea’s cloud security certification. This year’s result gives Samsung SDS and Elice a public-sector AI infrastructure reference at a time when domestic cloud providers are competing to show they can provide more than conventional enterprise hosting.
Samsung SDS has been building that case through both NVIDIA and domestic accelerator services. In addition to its B300 cloud offering, it launched NPU-as-a-Service on July 20 using FuriosaAI’s second-generation RNGD accelerator. FuriosaAI and The Korea Times describe that platform as an inference-focused service, intended to offer an alternative to GPUs for deployed models rather than replace GPU clusters used for broad training workloads.
The government research program, however, remains centered on GPUs. That is the sensible choice for a broad academic-and-research allocation service: researchers need compatibility with existing training stacks, mature tooling, and the flexibility to run workloads beyond inference. The domestic NPU service may become relevant to future programs focused on efficient production inference, but it is not what this procurement is buying.
The missing details now determine whether the project delivers
Samsung SDS has won an important role in South Korea’s effort to make scarce AI compute available to research teams, and its B300 capability gives the project access to newer hardware than a typical H100-only deployment. Yet the useful measure of the program will be allocation transparency, not the appearance of the Blackwell name in the supplier’s inventory.
KAIT, the Ministry of Science and ICT, IITP, Samsung SDS, and Elice have yet to publish a consolidated account of the final budget, GPU split, researcher-selection timetable, supported environment, and service end date. The original tender envisioned at least 960 GPUs and up to eight months of access; the record now shows a delayed rebid, a possible budget reduction, and conflicting end dates.
Researchers planning projects around the program should therefore treat the service as available but avoid designing an experiment around a presumed B300 allocation, a fixed cluster size, or access through March 2027 until the agencies and suppliers publish the executed terms.