South Korea’s National AI Computing Center has broken ground at the Solasido Data Center Park in Haenam, putting a long-delayed 15,000-accelerator public-private project on the construction path. The immediate payoff is practical: the Samsung SDS-led operator now has a route to deliver shared AI capacity to Korean universities, research institutes, startups and smaller companies that cannot independently finance frontier-scale GPU clusters.
But the August 3 ceremony also marks a clear policy retreat. The Ministry of Science and ICT originally required the future operator to accept public control, a put-option obligation, and a target for domestic AI semiconductors to account for 50% of installed hardware by 2030. No consortium bid in either of the first two rounds. Samsung SDS became the sole bidder only after Seoul reworked all three provisions.
The country did not abandon domestic AI chips. It abandoned a guaranteed purchase mandate for them, replacing it with a validation-and-voluntary-adoption model. That is a far more commercially credible structure for getting a data center built, but it puts the burden back on Korea’s NPU makers to demonstrate that their hardware can win real workloads against NVIDIA-led GPU deployments.
The original procurement was structured as a public-private special-purpose company, but its risks landed disproportionately with the private side. As the Ministry of Science and ICT acknowledged in its September 2025 implementation plan, the initial structure gave the public sector a 51% share, included a put-option obligation, and required domestic AI semiconductors to make up half of the installed base by 2030.
Those conditions were not merely unpopular. They created a mismatch between responsibility and control. A private operator would have been expected to finance, build, run, market and maintain a high-cost AI facility while lacking majority control and carrying a future obligation connected to the government’s capital. It would also have had to absorb the operational consequences of a chip mix dictated partly by industrial policy rather than by customer demand, software compatibility, availability, power efficiency, or benchmarked performance.
The procurement record also needs one clarification. The first call was not simply a May 2025 tender: the Ministry says the initial process ran from January 23 through May 30, followed by a second round from June 2 through June 13. Both ended with no applicants. The two failures were therefore the visible conclusion of a process that had been open for months, not a pair of brief market tests.
In September 2025, Seoul reset the terms. Public ownership was reduced to below 30%, the put option was removed, and the 50% domestic-semiconductor requirement was replaced by private-sector proposals for feasible adoption. The revised tender ran from September 8 through October 21, and Samsung SDS submitted the only bid. The Ministry selected its consortium as preferred bidder after technical, policy and financial reviews.
The first concrete lesson is straightforward: the project became investable only after the government stopped trying to use the operator as the enforcement mechanism for its semiconductor policy. That is not a failure of the center itself. It is an admission that a national AI utility cannot be built on terms private operators regard as structurally unfinanceable.
That capital split places the public share at 29%, consistent with the revised tender rules. The reported full investment target is ₩2.5 trillion, although recent local reporting ahead of the groundbreaking cited ₩2.4 trillion or ₩2.4065 trillion through 2030. The difference is modest relative to the scale of the program, but it is a reminder that the government and local project backers are not yet presenting one uniform final budget number.
The project’s governance deserves equally close attention. A sole bidder is not proof of improper conduct, particularly after two failed tenders made it politically and economically important to find an operator. But it means the final selection was a pass-or-fail review of one proposal rather than a competition among competing technical designs, financing plans, pricing models, or domestic-chip strategies.
The Ministry’s March selection notice said the details of the special-purpose company’s board, rights, obligations and operational arrangements still had to be specified with the government, Korea Development Bank, Industrial Bank of Korea and the consortium. The later implementation agreement moved the project forward, but the public record still does not provide the information prospective users will care about most: who receives priority access, how subsidized capacity will be allocated, what commercial pricing will be, which workloads qualify for public-interest discounts, or how much capacity will be reserved for the consortium’s own members.
Those omissions are more consequential than the ceremonial target of 15,000 chips. A national compute center can be technically impressive and still fail its public mission if scarce capacity is consumed by anchor tenants, opaque allocation rules, or prices that remain out of reach for the startups and university labs the project is supposed to help.
This is a materially different policy. Under the original model, domestic chip suppliers would have had an assured route into the national facility, whether or not their software stacks and performance profiles were mature enough for the workloads customers brought to the center. Under the revised model, domestic accelerators can enter through pilots and services, but they must earn production deployment.
That is tougher for companies such as FuriosaAI and Rebellions, because the primary barrier is not simply silicon performance. Enterprise adoption depends on compiler maturity, framework support, driver stability, observability tools, model-serving software, cluster schedulers, engineering support, and reproducible results on workloads customers actually run. A chip that looks compelling in a narrow inference test can still be difficult to deploy across a shared, multi-tenant national facility.
It is also the more defensible arrangement for a facility that will serve diverse users. Universities may need experimental access to non-NVIDIA systems. Startups may want lower-cost inference capacity. But an operator cannot reasonably promise shared production capacity on accelerators whose operational behavior has not been validated at scale.
The submitted account’s specific hardware roadmap — early NVIDIA B300 systems followed by Vera Rubin systems — should be treated cautiously. The Ministry’s official notices and the public project reports reviewed for this article promise “advanced AI chips” and more than 15,000 GPUs or accelerators by 2028, but they do not identify NVIDIA, B300, Vera Rubin, or a final hardware procurement schedule. Nor do they publish a binding domestic-NPU share. Those decisions appear to remain with the operator and later procurement process.
For IT buyers, that distinction matters. The center has a capacity target, not a confirmed bill of materials. The difference determines compatibility planning, model portability, power and cooling design, and whether future users should expect a homogeneous GPU environment or a mixed-accelerator platform.
There is, however, another inconsistency in the published project details. Yonhap reported gross floor area of 16,978 square meters for the data-center, operations and ancillary buildings. Local reporting in Gwangju Ilbo cited roughly 33,740 square meters. The public reports do not explain whether the larger figure includes a later expansion phase, additional buildings, or a different measurement definition. Until KOACC publishes a detailed final site plan, neither figure should be treated as the unqualified final footprint.
The power target also frames the real delivery risk. The initial 40 MW build is substantial but not extravagant by modern AI-campus standards, particularly if the center adopts dense next-generation accelerators and associated networking. Getting to 80 MW will require not only construction but firm electrical capacity, transmission and substation coordination, cooling infrastructure, and a schedule that aligns equipment deliveries with commissioning. The groundbreaking proves that the project has crossed its procurement barrier; it does not prove that every dependency needed for 2028 operation has been secured.
The Ministry has said the center will provide computing infrastructure in cloud form and support research and AI services across industry and academia. That makes availability, queue policy, tenant isolation, data governance and support tooling as important as the accelerator count. A university researcher who receives raw GPU hours but cannot obtain reliable storage performance, a supported software image, secure data access, or predictable scheduling has not gained a usable national AI platform.
The policy retreat makes that timetable more realistic. It removes a procurement condition that stopped the project twice, lets Samsung SDS and its partners choose hardware based on operational needs, and preserves a route for domestic NPUs to gain evidence in a national-scale environment. The center can now succeed as infrastructure even if Korean chips do not immediately capture half its installed capacity.
What Seoul has given up is the certainty that a flagship national compute facility will create demand for domestic accelerators at a pre-set scale. The R&D and NPU zones may generate the validation data that Korean suppliers need; they may also reveal performance, software, or support gaps that prevent broader adoption. Either outcome is more useful than an unworkable quota, but it is less protective.
Construction has started. The next meaningful public milestone is not another announcement of AI ambition, but a published operating model: confirmed accelerator procurement, allocation rules, user pricing, domestic-NPU trial criteria, and a credible 2028 commissioning schedule.
The country did not abandon domestic AI chips. It abandoned a guaranteed purchase mandate for them, replacing it with a validation-and-voluntary-adoption model. That is a far more commercially credible structure for getting a data center built, but it puts the burden back on Korea’s NPU makers to demonstrate that their hardware can win real workloads against NVIDIA-led GPU deployments.
The project that nobody would operate
The original procurement was structured as a public-private special-purpose company, but its risks landed disproportionately with the private side. As the Ministry of Science and ICT acknowledged in its September 2025 implementation plan, the initial structure gave the public sector a 51% share, included a put-option obligation, and required domestic AI semiconductors to make up half of the installed base by 2030.Those conditions were not merely unpopular. They created a mismatch between responsibility and control. A private operator would have been expected to finance, build, run, market and maintain a high-cost AI facility while lacking majority control and carrying a future obligation connected to the government’s capital. It would also have had to absorb the operational consequences of a chip mix dictated partly by industrial policy rather than by customer demand, software compatibility, availability, power efficiency, or benchmarked performance.
The procurement record also needs one clarification. The first call was not simply a May 2025 tender: the Ministry says the initial process ran from January 23 through May 30, followed by a second round from June 2 through June 13. Both ended with no applicants. The two failures were therefore the visible conclusion of a process that had been open for months, not a pair of brief market tests.
In September 2025, Seoul reset the terms. Public ownership was reduced to below 30%, the put option was removed, and the 50% domestic-semiconductor requirement was replaced by private-sector proposals for feasible adoption. The revised tender ran from September 8 through October 21, and Samsung SDS submitted the only bid. The Ministry selected its consortium as preferred bidder after technical, policy and financial reviews.
The first concrete lesson is straightforward: the project became investable only after the government stopped trying to use the operator as the enforcement mechanism for its semiconductor policy. That is not a failure of the center itself. It is an admission that a national AI utility cannot be built on terms private operators regard as structurally unfinanceable.
Samsung SDS won a sole-bidder contract, not a competitive contest
Samsung SDS is leading a consortium that includes Naver Cloud, Samsung C&T, Kakao, Samsung Electronics, Clush, KT, the South Jeolla provincial government, and Southwest Coast Enterprise City Development. The group proposed the Solasido site in Haenam, South Jeolla Province. Seoul Economic Daily reported in May that an implementation agreement had been signed after public and private parties confirmed ₩400 billion in initial investment: ₩116 billion from public sources and ₩284 billion from the private sector.That capital split places the public share at 29%, consistent with the revised tender rules. The reported full investment target is ₩2.5 trillion, although recent local reporting ahead of the groundbreaking cited ₩2.4 trillion or ₩2.4065 trillion through 2030. The difference is modest relative to the scale of the program, but it is a reminder that the government and local project backers are not yet presenting one uniform final budget number.
The project’s governance deserves equally close attention. A sole bidder is not proof of improper conduct, particularly after two failed tenders made it politically and economically important to find an operator. But it means the final selection was a pass-or-fail review of one proposal rather than a competition among competing technical designs, financing plans, pricing models, or domestic-chip strategies.
The Ministry’s March selection notice said the details of the special-purpose company’s board, rights, obligations and operational arrangements still had to be specified with the government, Korea Development Bank, Industrial Bank of Korea and the consortium. The later implementation agreement moved the project forward, but the public record still does not provide the information prospective users will care about most: who receives priority access, how subsidized capacity will be allocated, what commercial pricing will be, which workloads qualify for public-interest discounts, or how much capacity will be reserved for the consortium’s own members.
Those omissions are more consequential than the ceremonial target of 15,000 chips. A national compute center can be technically impressive and still fail its public mission if scarce capacity is consumed by anchor tenants, opaque allocation rules, or prices that remain out of reach for the startups and university labs the project is supposed to help.
The domestic-chip policy has shifted from quota to proof
Seoul’s revised procurement did not remove domestic NPUs from the National AI Computing Center. It removed the mandatory 50% deployment target. The Ministry’s implementation plan instead calls for phased adoption and demonstrations of domestic AI semiconductors, while allowing the private partner to propose a workable route. Seoul Economic Daily later reported that the project was considering an R&D zone for domestic-chip design, prototyping and validation, plus an NPU zone intended to pilot pre-commercial hardware before it is used in services.This is a materially different policy. Under the original model, domestic chip suppliers would have had an assured route into the national facility, whether or not their software stacks and performance profiles were mature enough for the workloads customers brought to the center. Under the revised model, domestic accelerators can enter through pilots and services, but they must earn production deployment.
That is tougher for companies such as FuriosaAI and Rebellions, because the primary barrier is not simply silicon performance. Enterprise adoption depends on compiler maturity, framework support, driver stability, observability tools, model-serving software, cluster schedulers, engineering support, and reproducible results on workloads customers actually run. A chip that looks compelling in a narrow inference test can still be difficult to deploy across a shared, multi-tenant national facility.
It is also the more defensible arrangement for a facility that will serve diverse users. Universities may need experimental access to non-NVIDIA systems. Startups may want lower-cost inference capacity. But an operator cannot reasonably promise shared production capacity on accelerators whose operational behavior has not been validated at scale.
The submitted account’s specific hardware roadmap — early NVIDIA B300 systems followed by Vera Rubin systems — should be treated cautiously. The Ministry’s official notices and the public project reports reviewed for this article promise “advanced AI chips” and more than 15,000 GPUs or accelerators by 2028, but they do not identify NVIDIA, B300, Vera Rubin, or a final hardware procurement schedule. Nor do they publish a binding domestic-NPU share. Those decisions appear to remain with the operator and later procurement process.
For IT buyers, that distinction matters. The center has a capacity target, not a confirmed bill of materials. The difference determines compatibility planning, model portability, power and cooling design, and whether future users should expect a homogeneous GPU environment or a mixed-accelerator platform.
Haenam’s construction milestone is only the first delivery deadline
Yonhap reported before the ceremony that the initial facility will occupy a 48,996-square-meter site in Haenam’s San-myeon district, begin at 40 megawatts and later expand to 80 megawatts. The reported 2028 target is 15,000 GPUs, with expansion toward 50,000 by 2030 across the broader national program.There is, however, another inconsistency in the published project details. Yonhap reported gross floor area of 16,978 square meters for the data-center, operations and ancillary buildings. Local reporting in Gwangju Ilbo cited roughly 33,740 square meters. The public reports do not explain whether the larger figure includes a later expansion phase, additional buildings, or a different measurement definition. Until KOACC publishes a detailed final site plan, neither figure should be treated as the unqualified final footprint.
The power target also frames the real delivery risk. The initial 40 MW build is substantial but not extravagant by modern AI-campus standards, particularly if the center adopts dense next-generation accelerators and associated networking. Getting to 80 MW will require not only construction but firm electrical capacity, transmission and substation coordination, cooling infrastructure, and a schedule that aligns equipment deliveries with commissioning. The groundbreaking proves that the project has crossed its procurement barrier; it does not prove that every dependency needed for 2028 operation has been secured.
The Ministry has said the center will provide computing infrastructure in cloud form and support research and AI services across industry and academia. That makes availability, queue policy, tenant isolation, data governance and support tooling as important as the accelerator count. A university researcher who receives raw GPU hours but cannot obtain reliable storage performance, a supported software image, secure data access, or predictable scheduling has not gained a usable national AI platform.
Korea has bought time, not hardware sovereignty
South Korea’s strategy is coherent in one respect: it is building shared compute capacity while trying to create a testbed for local AI-semiconductor companies. The country cannot wait for a wholly domestic accelerator stack to mature before supplying researchers and companies with the compute required to train and serve modern models.The policy retreat makes that timetable more realistic. It removes a procurement condition that stopped the project twice, lets Samsung SDS and its partners choose hardware based on operational needs, and preserves a route for domestic NPUs to gain evidence in a national-scale environment. The center can now succeed as infrastructure even if Korean chips do not immediately capture half its installed capacity.
What Seoul has given up is the certainty that a flagship national compute facility will create demand for domestic accelerators at a pre-set scale. The R&D and NPU zones may generate the validation data that Korean suppliers need; they may also reveal performance, software, or support gaps that prevent broader adoption. Either outcome is more useful than an unworkable quota, but it is less protective.
Construction has started. The next meaningful public milestone is not another announcement of AI ambition, but a published operating model: confirmed accelerator procurement, allocation rules, user pricing, domestic-NPU trial criteria, and a credible 2028 commissioning schedule.
References
- Primary source: Tech Times
Published: 2026-08-03T11:27:23+00:00
Loading…
www.techtimes.com - Related coverage: en.sedaily.com
Loading…
en.sedaily.com - Related coverage: koreatimes.co.kr
Loading…
www.koreatimes.co.kr - Related coverage: korea.net
Loading…
www.korea.net - Related coverage: korea.kr
Loading…
www.korea.kr - Related coverage: korea.kr
Loading…
www.korea.kr