South Korea's national AI factory just tripled in planned scale. NAVER, NVIDIA, and Brookfield Asset Management announced on July 25 that they will expand NAVER's NVIDIA DSX AI factory at the GAK Sejong hyperscale data center in Sejong, South Korea, from its initial 55-megawatt footprint to 200 megawatts by 2028 — more than three times the original buildout committed just six weeks earlier. The total financing envelope is $10 billion, making it one of the largest sovereign AI infrastructure announcements outside the United States. The largest number in that envelope, Brookfield's stated commitment of up to $9 billion, is a nonbinding term sheet. (Exchange rate as of July 26, 2026: approximately 1,460 Korean won per US dollar; Korean won figures below are approximate USD conversions at that rate.)
Announcement Arrives on the Sidelines of a Presidential Summit
The deal did not emerge from a boardroom alone. South Korean President Lee Jae Myung was in San Francisco on July 24-25 to host an AI summit that gathered the heads of NVIDIA, OpenAI, Anthropic, and Broadcom alongside the chairmen of Samsung, SK Group, Hyundai Motor Group, and NAVER. President Lee unveiled a "San Francisco AI Declaration" outlining Korea's ambitions for technological cooperation with the United States, positioned around making the country an indispensable node in the global AI supply chain.
The NAVER deal was the most technically specific commitment to emerge from that summit. On the sidelines, SK Group and NVIDIA announced a separate initiative valued at more than $500 billion covering large-scale AI data centers and next-generation memory, and Samsung Electronics signed a memorandum of understanding with Broadcom covering memory chips, foundry services, and advanced packaging.
Capital Stack: What Is Committed and What Is Not
The financing architecture matters for understanding what has actually been agreed. Under the announced terms, Brookfield will act as the exclusive capital partner, funding up to $9 billion; NVIDIA will invest approximately $1 billion directly into NAVER Corp. as an equity stake; and NAVER will fund the remainder of the $10 billion project.
Each of those three commitments carries a condition. Brookfield's up-to-$9-billion figure is a nonbinding term sheet — an intention to finance, not contracted financing. NVIDIA's roughly $1 billion equity investment is conditional on NAVER successfully finalizing at least $9 billion of committed financing from other sources, separate from NVIDIA's own contribution. NAVER covers whatever gap remains after both partners contribute.
For Brookfield, the move fits within a larger program: the firm manages approximately $100 billion of assets across the AI infrastructure value chain globally, including data centers, compute, semiconductor manufacturing, and dedicated power generation, and has been investing in Korea since 2014, accumulating approximately $12 billion of assets there across infrastructure, real estate, and energy. Brookfield launched its dedicated Brookfield Artificial Intelligence Infrastructure Fund in November 2025 with NVIDIA and the Kuwait Investment Authority as anchor investors, targeting $10 billion in equity commitments and up to $100 billion in total AI infrastructure assets globally. The Korea commitment channels through that same strategic program.
Sikander Rashid, Brookfield's global head of AI infrastructure, framed the deal in national-strategic terms: "Access to trusted, sovereign and scalable AI infrastructure is becoming an increasingly important strategic priority for companies and countries."
What NVIDIA DSX Actually Does: The Engineering Underneath the Investment
The 200 megawatts of planned capacity will be built on NVIDIA's DSX platform, announced at GTC Taipei 2026 in June — and the distinction between a DSX AI factory and a general-purpose hyperscale data center is the engineering story this deal rests on.
A conventional hyperscale data center handles a variety of compute workloads and is designed for flexibility. An AI factory is designed for exactly one thing: producing tokens at the lowest possible cost per megawatt. The NVIDIA DSX platform achieves this through co-design across five layers simultaneously — chips, systems, software, power delivery, and cooling — rather than treating them as independent procurements.
The platform's two primary software components are DSX MaxLPS and DSX OS. DSX MaxLPS is a power and efficiency suite that combines 45-degrees-Celsius liquid cooling — significantly warmer than standard liquid-cooled data center designs — with rack-level optimization tools that allow operators to run up to 40 percent more GPUs within the same megawatt budget at their most energy-efficient operating point, with minimal impact on workload performance. DSX OS is an open-source modular software layer purpose-built for AI factory operations, covering lifecycle management, intelligent scheduling, health automation, resiliency, and multi-tenant infrastructure management.
The 45°C cooling threshold is not an incidental detail. Standard liquid-cooled data centers typically operate at 20-30°C on the coolant return side; running at 45°C enables direct use of cooling tower rejection rather than chillers, reducing capital cost and energy overhead. This is a specific engineering tradeoff that makes the AI factory's economics possible at the described power density.
The platform also includes DSX Sim for digital-twin simulation and validation before physical deployment, DSX Flex for dynamically adjusting workloads based on grid load and electricity pricing, and DSX Exchange for real-time coordination among compute, networking, energy, and cooling systems. NVIDIA claims a 1,000,000-fold increase in inference throughput per megawatt across six GPU generations. More specifically: the Blackwell Ultra GB300 NVL72 system delivers up to 50 times higher throughput per megawatt and 35 times lower token cost compared to the prior Hopper generation for certain workloads; the Vera Rubin platform goes further, achieving up to 10 times higher inference throughput per megawatt and approximately 10 times lower token cost than Blackwell for AI inference workloads.
The GAK Sejong factory will deploy Vera Rubin and Blackwell systems, making it among the first national AI facilities to deploy both in combination.
Why the 200 Megawatts Figure Understates the Energy Challenge
Two hundred megawatts of continuous draw is roughly the sustained power appetite of a small city — and it must be delivered as firm, around-the-clock power to a single campus in Sejong. Grid interconnection queues, transmission capacity upgrades, and the availability of baseload power have emerged globally as the binding constraint on AI data-center expansion, frequently outpacing chip supply as the limiting factor.
Korea has a dense and capable grid, which is part of why the GAK Sejong site is viable. But a five-fold increase to a stated gigawatt ambition implies power procurement and grid coordination that no term sheet settles. Electricity infrastructure moves on permitting and construction timelines that do not compress the way software roadmaps do. When national AI factory projects stall, they typically stall here — at the substation, not the server — and the 2028 date for 200 megawatts is the milestone most exposed to that risk.
South Korea's government has stated it will support timely electricity and water supply for AI data centers and publish information about available high-voltage grid interconnection capacity, reflecting an awareness that power access is the operational variable that could delay even well-financed plans.
NAVER's Role: Korean-Language AI Champion with Hardware Help
The infrastructure deal is coupled to a model development strategy that defines what the factory will actually run. NAVER is advancing its HyperCLOVA X large language models using NVIDIA Nemotron 3 Ultra open models as a foundation, combining them with NAVER's own Korean-language training data and post-training expertise. NAVER became the first Korean company admitted to the NVIDIA Nemotron Coalition, contributing to open model development across pretraining, post-training, and reinforcement learning.
NAVER also plans to launch an AI agent platform in Korea in the second half of 2026, powered by NVIDIA Agent Toolkit software including NVIDIA NemoClaw blueprints, and is developing a Seoul World Model — a spatial AI system built on NVIDIA Cosmos world foundation models, using NAVER's proprietary urban street-view and spatial modeling data from Seoul.
Where NAVER's models have clear documented strength is on Korean-language benchmarks. On the KoBALT-700 test suite — developed by Seoul National University's linguistics department to assess large language models' Korean comprehension — HyperCLOVA X Think scored 48.9, well ahead of LG AI Research's Exaone Deep at 33 and Alibaba's QwQ-32B at 32.4. That linguistic advantage is directly relevant to the factory's data sovereignty value proposition: a Korean company running Korean-optimized models on Korean-controlled infrastructure, under Korea's Personal Information Protection Act, is a meaningfully different product for Korean enterprises and government agencies than a US hyperscaler can offer.
NAVER recorded KRW 12.04 trillion (approximately $8.18 billion USD) in sales for 2025, providing the financial base to credibly anchor a national AI infrastructure program at this scale.
NVIDIA's Role: Chip Supplier, Equity Holder, Platform Vendor
NVIDIA captures value at three distinct points in this deal, and understanding that structure is the key to understanding what "sovereign" actually means in practice here.
First, NVIDIA supplies the hardware: Vera Rubin and Blackwell GPUs fill every rack in NAVER's factory. Second, NVIDIA holds equity in the operator: the roughly $1 billion equity stake in NAVER Corp. converts a customer relationship into a long-term financial interest. Third, NVIDIA provides the platform: DSX is the operating software and methodology stack the entire factory runs on.
That triple position is by design. NVIDIA has taken similar minority equity stakes in sovereign AI infrastructure partners in other nations, turning what would otherwise be a chip-supply agreement into a compounding relationship: it sells the silicon, holds equity in the buyer, and earns platform revenue on top. Each national factory NVIDIA seeds makes the next one easier to justify.
The implication for Korea's sovereignty claim is direct and acknowledged by independent analysts. Digital Today's analysis of South Korea's national AI infrastructure concluded that the program risks amounting to little more than placing NVIDIA chips on Korean territory. NAVER operates the facility and the data stays in Korea — those are genuine dimensions of control. But the hardware itself originates from a single US vendor that also holds an equity stake in the operator. Location sovereignty and hardware sovereignty are different things, and this deal delivers the first without the second.
The contrast with China's approach is instructive. Z.ai recently brought a comparable gigawatt-scale AI compute facility online running entirely on Chinese-made accelerators, with no NVIDIA hardware inside — a harder and more costly version of sovereignty forced by US export controls. Korea chose the opposite trade: buy the best available silicon, keep control of the site and the data, and accept hardware dependence as the cost. Neither is obviously correct. They are genuinely different bets on what sovereignty requires.
Korea's Broader AI Infrastructure Bet
The NAVER deal is a single piece of a sweeping national AI industrial policy that Korea's government formalized in late June 2026. Under a three-part "megaprojects" strategy, South Korea committed to an 8.4-gigawatt first phase of AI data-center capacity by 2029, to be built by SK Group (5 gigawatts), GS Group (2.4 gigawatts), and NAVER (1 gigawatt), with total private investment of KRW 550 trillion (approximately $377 billion USD). A second phase running from 2030 to 2035 would bring total capacity to 18.4 gigawatts at a cumulative investment projected to exceed KRW 1,000 trillion (approximately $685 billion USD).
The NAVER factory sits within NAVER's 1-gigawatt allocation of that broader program. At 200 megawatts, it represents 20 percent of NAVER's stated national commitment — meaning even the tripling announced this week, if completed on schedule, puts NAVER only a fifth of the way toward its portion of the national target.
What makes Korea's position strategically distinct from most sovereign AI programs is the hardware position it already occupies at the upstream end of the supply chain. Samsung and SK Hynix together supply roughly 80 percent of global high-bandwidth memory production — the chip architecture that feeds AI accelerators at both training and inference speed. A country that makes the memory that goes inside NVIDIA's GPUs occupies a structurally different position in the AI supply chain than a country that only deploys the finished product.
Jensen Huang, NVIDIA's CEO, framed the strategic logic: "AI factories are the infrastructure nations need to compete and innovate in the intelligence era. Together, NAVER, NVIDIA and Brookfield are building sovereign AI infrastructure at the scale needed to fuel Korea's startups and industries."
The Gap Between the Announcement and the Compute Pool
For Korean AI startups, the deal's promise is real but deferred. A dedicated national compute pool — access to production-scale AI infrastructure without owning the hardware — is exactly what an early-stage lab needs. The open questions are price and timing: a pool announced for 2028 does not help a team training a model in 2026, and compute capacity priced comparably to US hyperscalers delivers less competitive advantage than the national branding implies.
No public pricing for the Korean AI compute pool has been announced. That is the variable that will determine whether the factory functions as a genuine national resource that lowers AI development costs domestically, or as NVIDIA-priced capacity on Korean soil with a national flag attached. The 2028 delivery date for 200 megawatts is a concrete, checkable milestone. The conversion of Brookfield's nonbinding term sheet into signed financing is the financial event that transforms intention into infrastructure. And the actual startup pricing, when announced, will be the test of whether the deal's sovereign AI framing delivers for the companies it claims to serve.
Frequently Asked Questions
Is Brookfield's $9 billion commitment to Korea's AI factory actually signed?
Not yet. Brookfield's up-to-$9-billion contribution is a nonbinding term sheet — an intention to finance, not contracted capital. NVIDIA's roughly $1 billion equity stake in NAVER is the firmer commitment, though it is also conditional on NAVER securing at least $9 billion in financing from other sources. NAVER covers whatever gap remains. The deal's total $10 billion figure is the combination of these three still-contingent commitments.
What makes NVIDIA DSX an "AI factory" platform rather than just a data center?
A general-purpose hyperscale data center handles diverse compute workloads and prioritizes flexibility. An AI factory is co-designed from the start for a single purpose: producing AI inference tokens at the lowest possible cost per megawatt. NVIDIA DSX achieves this by unifying chip selection (Blackwell, Vera Rubin), power management (DSX MaxLPS with 45-degrees-Celsius liquid cooling), operational software (DSX OS), and facility design into one validated system. That 45°C cooling standard specifically enables up to 40 percent more GPUs per megawatt compared to conventional liquid-cooled designs. NVIDIA claims the result is up to 10 times higher inference throughput per megawatt for Vera Rubin versus the prior Blackwell generation for certain workloads.
Is South Korea's AI factory actually "sovereign" if it runs on NVIDIA hardware?
In location and operation, yes — the data resides in Korea under a Korean operator, and Korean enterprises and government agencies have access without routing data through foreign-controlled systems. In hardware, no. Every GPU in NAVER's factory is manufactured by NVIDIA, and NVIDIA holds an equity stake in NAVER as part of the deal. Analysts at Digital Today have noted the risk that the program amounts to placing NVIDIA chips on Korean territory rather than building genuine technological independence. The approach Korea has chosen — buy best-available hardware, control the site and data — is a deliberate strategic tradeoff, not an oversight. China took the opposite path under US export control pressure, building gigawatt-scale AI compute on domestically manufactured chips. Both models exist; only one is available to a country with full access to NVIDIA's export-controlled silicon.
How does Korea's 200-megawatt AI factory fit into the country's broader AI investment plans?
South Korea's government announced in late June 2026 a three-part national AI megaproject targeting 8.4 gigawatts of AI data-center capacity by 2029, backed by KRW 550 trillion (approximately $377 billion USD) in private investment from SK Group, GS Group, and NAVER. NAVER's portion of that commitment is 1 gigawatt. The 200-megawatt target announced on July 25 represents 20 percent of NAVER's stated national allocation — meaning completion of this facility, on its 2028 timeline, would bring NAVER one-fifth of the way to its portion of the national program. The government has committed to supporting electricity supply access and grid interconnection as part of enabling that buildout, reflecting an understanding that power availability is the binding constraint most likely to cause delays.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.