We use cookies to understand how our site is used so we can improve it. Necessary cookies are always on. You can accept all, reject all, or choose what to allow. Learn more
Internet InfrastructureThe boring annuity two activists couldn't ignore.
Internet Infrastructure
The boring annuity two activists couldn't ignore.
Two national projects are converging on the same few thousand square metres of Korean server space. One is sovereign AI: a state-backed push to keep frontier compute, the data under it, and increasingly the chips inside it on home soil, channelled through a certification regime that quietly fences out foreign hyperscalers. The other is corporate-governance reform: a 'value-up' agenda, sharpened by an April-2026 ban on the split-listings that let Korean groups float their best subsidiaries away from parent shareholders. This is a study of where those two waves meet — the cloud-and-data-centre build-out and the unwinding of the holding-company discount — read through one company, Gabia, that sits squarely in the overlap.
Gabia runs the plumbing of the Korean internet — domains, web hosting, and the data centers underneath them. It is a high-margin recurring annuity caught inside a holding-group structure that Korea's governance-reform wave — and two activist funds — are now pressing to simplify.
South Korea is trying to build two things at once that have almost nothing to do with each other on paper, and everything to do with each other in practice. The first is a sovereign stack for artificial intelligence — the compute, the data centres and now the chips that a mid-sized, export-dependent country wants to keep inside its own borders rather than rent from a US hyperscaler. The second is a rewrite of how its public companies are owned, aimed at a peculiarly Korean habit: listing a group's best subsidiaries separately, on the same exchange, so that the parent's shareholders watch the value walk out the door. Both are national projects with budgets and statutes behind them. They land on the same scarce asset — certified, domestically operated server capacity — and that is where this market gets interesting.
Take the cloud layer first, because it sets the scale of everything underneath. Korea's cloud-computing market was worth on the order of US$10 billion in 2025 — a figure that has compounded for half a decade on the back of a government 'cloud-first' migration, subsidies that cover up to 80% of a small firm's adoption cost, and hyperscaler campuses landing on Korean soil. It is, by global standards, a market that is large enough to matter and still early enough to bend. The question is not whether it grows. It is who is permitted to capture the growth, and where inside the value chain the growth actually pays.
Korea cloud-computing market, 2025 base
≈US$10bn
Roughly US$9.95bn (Mordor) to US$12.4bn (Fortune Business Insights) depending on where the boundary is drawn around 'cloud.' Public cloud is about two-thirds of the spend; IaaS roughly 47% of it. Sources: Mordor Intelligence; Fortune Business Insights; Grand View Research.
How the market is built, and where the margin pools sit
A Korean cloud business is a stack of five layers, and they are not equally good businesses. At the top of the funnel are domains and SSL certificates — cheap, high-volume, low-margin, and quietly shrinking as the country-code domain base ages; their value is that they are the first thing a small business buys, the hook for everything above. Then web and managed hosting. Then cloud proper — infrastructure-as-a-service and desktop-as-a-service — where the public-cloud share keeps rising and where domestic providers fight the global platforms head-on. Then, at the foundation, the internet exchange and the data-centre colocation that physically house all of it. And bolted across the top, a nascent fifth layer: renting out AI accelerators by the hour.
The margin does not sit where the headlines are. It pools at the bottom, in the layer that is hardest to reproduce. Anyone with a brand and a billing system can resell hosting; almost no one can conjure a carrier-neutral internet exchange or a colocation hall in metropolitan Seoul, because both are gated by things money alone cannot fix quickly — interconnection density, certification, and above all electrical power. Seoul has effectively stopped issuing permits for new large-load data centres in the city; the national utility, KEPCO, is carrying debt north of US$140 billion, which constrains the grid upgrades that new capacity needs. Financial investors are backing roughly 90% of the planned 1.9 GW of new metro capacity through 2027, but the binding constraint is a substation, not a spreadsheet. Scarcity, in other words, is the asset. The colocation and exchange layer is where a structurally short market pays a rent that the resale layers above it never will.
The scarcity is physical. Colocation is roughly four-fifths of Korea's data-centre market, and the metropolitan supply that AI demand wants most is gated by a Seoul permit freeze and a grid the national utility cannot cheaply expand. Capacity, not code, is the constrained input.
What is driving demand — and where it goes by 2030
Three demand pulls are stacked on top of each other, and the newest is the largest. The base layer is ordinary digitisation: the state aims to move on the order of 10,000 government systems to cloud by 2030, and the subsidy regime keeps pulling small firms up the stack. On top of that sits hyperscaler capex landing physically in Korea — an AWS–SK partnership alone describes a roughly US$5 billion, 60,000-GPU campus. And on top of that, the thing that changes the slope: sovereign AI. In October 2025 a Korea–Nvidia alliance committed to bringing some 260,000 GPUs into the country by 2030, around US$9.8 billion (₩14 trillion) of silicon, with roughly 50,000 earmarked for government use — a national AI computing centre, due in 2027 at about one exaflop and ~15,000 GPUs, and sovereign foundation models. A separate ₩1.46 trillion supplementary budget seeds another 10,000-plus GPUs through domestic cloud operators. Every one of those chips has to sit in a certified, powered, domestically operated rack. That is the demand, made concrete.
Sovereign-AI compute commitment, by 2030
≈260,000 GPUs
About US$9.8bn (₩14tn) under the Oct-2025 Korea–Nvidia alliance; ~50,000 GPUs to government, the balance to Samsung, SK, Hyundai and Naver. A widely repeated '~₩100tn' national AI figure and a 52,000-by-2028 interim are directional and not independently confirmed here; the ₩14tn alliance, the ₩2tn AI computing centre and the ₩1.46tn supplementary budget are. Sources: KED Global; Korea Herald; OECD.AI.
Sized forward, the independent forecasts agree on the shape and argue about the number. The convergent read for Korean cloud is a compound rate around 24–26%, taking a ~US$10 billion market in 2025 to roughly US$31–38 billion by 2030–31 — Grand View lands near US$31 billion by 2030; Mordor's trajectory reaches about US$38 billion by 2031. Treat that as a direction, not a decimal: the spread is wide because 'cloud' is drawn differently by each house, and the data-centre layer underneath is worse, differing by nearly threefold depending on what you count. The trajectory is the reliable part. The precision is not.
Korea's cloud market is forecast to roughly triple-to-quadruple by the early 2030s — and the compounding rate holds remarkably steady at around a quarter a year, which is the part the forecasters agree on.Market size in US$bn (bars), plotted at sourced points only. The line is the convergent compound rate over the whole 2025–31 horizon (~25%), drawn flat on purpose — it is a period CAGR, not a year-by-year growth rate, so the single 2025→2026 step (≈20% on the rounded US$10bn→US$12bn points) is deliberately lower than the line; the flat line shows the pace is steady rather than front-loaded. All figures approximate and rounded; the size band spans ~US$31bn by 2030 (Grand View) to ~US$38bn by 2031 (Mordor). Sources: Mordor Intelligence; Grand View Research; Fortune Business Insights.
The structural shift: certification, domestic silicon, and the unwinding of the discount
Two structural forces are reshaping who wins this market, and they are the two waves named at the start. The first is regulatory and runs through a certification called CSAP, administered by KISA and the science ministry. CSAP gates public-sector and finance-grade cloud across three tiers, and even the lowest tier now requires operations staff physically in Korea — a de-facto data-residency rule that functions as a moat for domestic providers. A foreign hyperscaler cleared the top IaaS tier for the first time only in late 2024; the system is opening, but slowly, and on Korea's terms. Layered on top is a domestic-chip preference: the government's 2026 high-performance-computing program runs an explicit domestic-NPU track, and a homegrown accelerator maker, Rebellions — Korea's first AI-chip unicorn after merging with Sapeon — is betting that the inference market, served on an open PyTorch/Triton/vLLM stack, can route around Nvidia's CUDA lock-in. That bet is unproven at scale, and worth flagging as such. But the policy intent is unambiguous: keep the compute, the data and the silicon domestic.
The second force is ownership. For decades, Korean groups have grown by 'split-listing' — carving out a prized division and floating it separately while keeping control, which lets the controller expand the group on little capital but leaks value away from the parent's own shareholders. The reform wave is dismantling the practice from two directions. A June-2025 Commercial Act amendment redirected directors' fiduciary duty toward shareholders, added cumulative voting and mandated treasury-share cancellation; then, in April 2026, the financial regulator banned new split-listings outright and required parent-shareholder approval for subsidiary IPOs. The empirical case behind the reform is stark: across 261 dual-listing cases from 2000 to 2024, the parent's stock fell about 11% in the six months after a subsidiary went public. The template for what comes next is Japan, where Tokyo's governance push set off a wave of parent–subsidiary re-absorptions — buy-ins, mergers, take-privates — to collapse exactly this structure.
The policy tailwind, dated
Apr 2026
Financial-regulator ban on split-listings (물적분할 spin-off listings) plus parent-shareholder approval for subsidiary IPOs, following the June-2025 Commercial Act amendment that made directors' duty run to shareholders. Empirical basis: ~−11% parent stock six months post-subsidiary-IPO across 261 cases, 2000–2024. Sources: KED Global; Janus Henderson; ITIF.
The competitive structure: a low-concentration field
Despite the policy steering, this is a fragmented market, not a cartel — its concentration is low at every layer, which is why archetype matters more than market share. The competitors sort into recognisable kinds, and where a firm sits determines whether it earns scarcity rent or merely resells someone else's.
The global hyperscalers — AWS, Microsoft Azure (first to clear top-tier CSAP IaaS, in late 2024) and Google — with scale and tooling no domestic firm matches, but throttled by the residency and certification regime in exactly the public and finance workloads that matter most.
The domestic full-stack clouds — Naver Cloud (around 60,000 clients), KT Cloud, NHN Cloud, Kakao Enterprise, plus the systems-integrator arms Samsung SDS and LG CNS — which carry the sovereign-AI and public-sector workloads the rules reserve for in-country operators.
The scarcity-rent infrastructure layer — the carrier-neutral internet exchanges and colocation operators who own the constrained physical asset. Roughly half the country's data-centre IT load already sits with LG Uplus, KT Cloud and SK Broadband; independent market trackers name Korea's neutral exchange operator alongside Equinix and Telehouse in the same colocation tier.
The domestic-silicon challengers — Rebellions and the NPU-as-a-service layer beginning to form around them — a nascent option on the sovereign-chip thesis rather than a present-day profit pool.
The case in point: one company in the overlap
Which brings the two waves to a single address. Gabia is useful here not because it is the biggest player in any one layer, but because it is the rare company that sits in the cross-hairs of both — a full-stack internet-infrastructure group that owns a piece of the scarcity layer and is, simultaneously, a textbook specimen of the split-listed structure the reform wave is built to unwind. It runs the unglamorous plumbing: a leading Korean domain registrar at the top of the funnel — holder of the KISA .KR registry-agency franchise — then hosting and SSL, then cloud (IaaS and DaaS), and underneath it all the country's only carrier-neutral internet exchange. The group reported ₩335.7 billion of consolidated revenue for FY2025, up 19% year on year and compounding around 14% over five years (company-reported, via DART). The plumbing works.
Gabia runs the recurring layer of the Korean internet — domains at the top of the funnel, hosting and cloud in the middle, a carrier-neutral internet exchange at the base. It is one company straddling both national projects: the sovereign-AI build-out and the governance-reform unwind.
The structural crux is visible in the segments. Gabia's two largest engines are roughly co-equal — the parent's own cloud-and-IT line and the internet-exchange/colocation arm — but the exchange business, the security arm and a Microsoft-licensing arm are each listed separately, with much of their economics owned by outside shareholders. That is the split-listed structure in miniature: headline scale that grows every year, sitting atop a parent whose own shareholders own less of the best parts than the consolidated top line implies. It is exactly the configuration the April-2026 reforms target, and exactly why governance investors have gathered on the register. The cohort below shows the operating shape, segment by segment.
Segment
FY24 (₩bn)
FY25 (₩bn)
FY25 %
YoY
Cloud & IT services
148.2
181.3
54%
+22%
IX / IDC colocation
91.2
97.6
29%
+7%
Security
42.7
47.4
14%
+11%
Other / elims
0.4
9.4
3%
—
Total
282.4
335.7
Two market currents then run straight through this one company. On the governance side, the structure stopped being theoretical in 2026: governance-focused funds accumulated past the founder bloc — a combined position in the high-30s percent against a founder-and-related holding in the mid-20s — and at the March-2026 annual meeting they seated two directors, taking the board to three external members of five. There was no filed alliance; the convergence is inferred from the vote arithmetic, and it remains board influence rather than realised capital action. But it is the local enactment of the national reform: the people now in the room are there because Korean law and policy have made the split-listed discount a board-level problem to solve rather than a permanent feature to tolerate.
Governance reform, enacted in one boardroom. At the March-2026 meeting, two outside directors were seated with roughly 61% of the vote apiece, moving the board to three external members of five — the structural unwind the value-up agenda is pushing across the market, here made specific.
On the sovereign-AI side, the same company added the market's newest layer. In April 2026 Gabia launched NPU-as-a-service — cloud inference running not on scarce Nvidia GPUs but on Rebellions' domestic ATOM-Max accelerator — and in May was selected into the government's 2026 high-performance-computing program on the domestic-chip track. There is no disclosed price card and the revenue is nascent; this is option value on the sovereign-silicon thesis, not a base-case earnings line, and its durability hinges on whether the domestic-NPU channel survives once free-trial programs end. But the positioning is precise: a certified, in-country operator carrying domestic silicon into a build-out engineered to favour exactly that.
The newest layer of the market, forming in real time. Renting domestic NPUs by the hour — here on Rebellions' ATOM-Max — is the sovereign-chip thesis turned into a product. It is government-seeded and unproven at scale: a real option on the build-out, not yet a profit pool.
“When a national project privileges domestic, certified infrastructure, the durable economics collect where the asset is genuinely scarce — a powered colocation hall, a neutral exchange, a slot on the approved-silicon list — not where the software is easiest to resell. Policy can manufacture demand. It cannot manufacture a substation.”
— Nathan Research Group, KOSDAQ Frontier Series N°02
Where the market goes from here
Project the two waves forward and they either compound or they stall, together. The cloud market keeps its ~25% trajectory toward roughly US$31–38 billion by 2030–31; the sovereign-AI program lands its 260,000 GPUs into certified domestic racks; and the governance reforms harden split-listing from a loophole into a liability, pushing holding-style groups toward the Japanese path of buy-ins and consolidations. In that world, advantage accrues to certified domestic operators who own scarce physical capacity and can credibly run domestic silicon. The downside cases are equally structural: the grid slips and the metro power that AI demand wants most arrives after 2027; the domestic-NPU software ecosystem fails to overcome CUDA inertia; or governance reform proves to be box-ticking — '174 value-up plans filed' that change little in practice. The following frames the outcomes at the market level, defined by what the market does and what an operator must execute — not by any valuation.
Path
What would have to be true in the market
What an operator must execute
Both waves compound
Grid delivers metro power; sovereign-AI GPUs land in certified racks; reform forces real structural change.
Lease up scarce capacity and convert domestic-silicon optionality into a recurring line.
Infrastructure wins, AI option doesn't
Colocation and certification stay scarce, but domestic NPUs lose to CUDA inertia.
Defend the exchange/colocation rent; treat NPU-as-a-service as upside, not core.
The build-out stalls
Power slips past 2027 and reform proves cosmetic; demand outruns deliverable capacity.
Survive the capex cycle; compete on resale layers where rent is thin.
Illustrative market-and-capability scenarios — analytical constructs to frame the question, not forecasts of record. Each is defined by a market outcome and an execution requirement, anchored to the 2025–2030 build-out and reform timeline.
The lesson generalises past one company and past Korea. When a state decides that a strategic capability — compute, energy, defence electronics, critical minerals — should be sovereign, it rewrites the competitive map twice over: once through a certification or residency rule that decides who may serve the demand, and once through the physical bottleneck (power, fab capacity, a port) that decides who actually can. The durable economics settle at the scarce, gated layer, and they settle there regardless of how the software market above it churns. Korea's cloud-and-data-centre build-out, braided together with its governance-reform wave, is simply the clearest current example of both rules operating at once — which is why a single mid-cap registrar ends up sitting at the intersection of two national projects. Our full brief on the Korean sovereign-AI and data-centre market — the sizing, the value-chain economics, the certification regime and the reform timeline — is available to download with this article.
Working With Nathan Research Group
Partner With Nathan Research Group
Filings give you the segment splits, the certification grants, and the cap-table arithmetic; they do not tell you whether a colocation hall actually has the megawatts to lease up, how domestic-NPU inference performs against the workloads buyers really run, or how a holding-style board behaves once reform law changes its duties. Nathan Research Group reaches the operators, regulators and engineers who live inside those questions, compliantly.
Korea’s first dedicated expert network — Seoul, since 2013
Who we put in the room
Activist / corporate governance
Korean shareholder-activism practitioners on the Miri–Align convergence, board-capture mechanics, and what a double-listing unwind actually requires after the seats are won.
Infrastructure finance / IDC-colocation
Carrier-neutral data-center economics, Seoul-metro power and siting constraints, and the PE-backed hyperscale-AI build-out pressing on KINX retail-colo pricing.
Cloud / public-sector IT (CSAP)
CSAP certification tiers, the 2023 logical-separation reform, and how a sub-scale domestic CSP actually wins regulated public and finance workloads.
AI-HPC / domestic NPU
Rebellions ATOM-Max inference economics, NPU-versus-GPU total cost, and how durable the MSIT/NIPA domestic-chip channel is once the free-trial program ends.
Gabia consolidated revenue by segment, FY2024→FY2025 (company-reported, DART). Cloud & IT is the largest and fastest line; the internet-exchange/colocation and security arms sit in separately listed subsidiaries — the structure the governance-reform wave is built to collapse.
Domain registry / SMB web services
The KISA .KR registry-agency franchise, registrar retention and churn, and what the shrinking ccTLD base means for the cross-sell funnel into hosting and cloud.
Holding-group structure & governance reform
How Korean split-listing structures get simplified in practice, the Japanese parent-subsidiary re-absorption template, and what the value-up and governance-reform wave changes for a holding-style group.
How an engagement works
1Scope
We turn your thesis into a precise expert profile and question set, mapped to the decisions you need to close.
2Source & vet
We screen and compliance-clear each expert — relevance, recency, and the absence of conflicts — before any call.
3Convene & synthesize
We arrange interviews on your timeline and, where useful, deliver written synthesis tied back to your questions.
If your team is evaluating Gabia, the Korean cloud-and-hosting sector, or the broader internet-infrastructure sector, tell us the decision you’re trying to make.
A direct reply from a partner, not an intake form.
Based solely on public sources. Nathan Research Group does not request or facilitate material non-public information, and runs every engagement through a documented compliance protocol.
