Nvidia and Naver on June 8 formalized a roadmap to build gigawatt-scale AI factories across Asia, the Middle East, and Europe — a deal that makes the South Korean internet and cloud operator one of Nvidia's anchor tenants for the DSX platform strategy it unveiled at GTC Taipei 2026, and that directly counters the growing ecosystem threat posed by Google's custom TPU silicon. Under the agreement, announced during a live-streamed event at Naver's 1784 headquarters in Seongnam, Naver will deploy AI factory infrastructure starting at 55 megawatts at its Gak Sejong hyperscale data center in the first half of 2027, scale to 100 megawatts later that year and 200 megawatts in 2028, with a stated long-term target of 1 gigawatt — roughly four times Gak Sejong's current maximum capacity — and 25 trillion won in AI factory revenue by 2030. Naver shares closed up 9.20 percent on the Korea Exchange that day at 279,000 won.
The timing is not coincidental. Google debuted its eighth-generation TPU silicon at Google Cloud Next 2026 in April, naming Anthropic, Meta, and OpenAI as anchor customers for its TPU-as-a-service business. Nvidia's response, across a four-day Seoul visit by CEO Jensen Huang, was to sign DSX-based gigawatt commitments with Korean operators who have both the infrastructure credentials and the regional regulatory positioning to offer sovereign AI products that Google's US-headquartered cloud cannot match on data-residency terms. Naver, as Korea's largest domestic cloud operator with established government and enterprise relationships, is the most credible of those regional anchors.
How Nvidia DSX Converts a Standard Data Center Into an AI Factory
The technical center of the Naver deal is the Nvidia DSX platform — a full-stack infrastructure reference design that Nvidia announced at GTC Taipei 2026 on June 1 and that formally distinguishes what the company now calls an "AI factory" from a general-purpose hyperscale data center.
A traditional data center stores, retrieves, and processes information, and its performance is measured in uptime and throughput. An AI factory is purpose-built to manufacture intelligence tokens — the units of output produced when an AI model processes a prompt — and its performance is measured in tokens per watt. That shift in metric is not semantic; it reorganizes the entire infrastructure stack. DSX codesigns compute, networking, storage, power delivery, and cooling as a unified system rather than as separately procured components integrated after the fact, and it exposes six coordinated layers:
DSX Reference Design provides generation-specific, validated factory architectures from chip to grid, covering compute cluster design, networking topology, storage integration, and civil and structural facility specifications. Building on it means Naver does not engineer the factory architecture independently.
DSX Sim is a digital twin simulation engine that allows Naver to model the full factory — including GPU behavior, cooling performance, power distribution, and network contention — before physical construction begins. At the density of a 55MW AI facility, infrastructure carries hundreds of millions of dollars in capital cost; simulating day-one performance before committing that capital addresses the most common deployment failure mode in large GPU clusters.
DSX MaxLPS is the software layer that dynamically manages power at the GPU, rack, and workload level to maximize token throughput within a fixed megawatt budget. Nvidia has stated that DSX MaxLPS enables AI factory operators to provision up to 40 percent more GPUs within the same power envelope compared to standard GPU cluster deployments — a material difference in how many Vera Rubin or Blackwell GPUs can be active simultaneously and, directly, in the revenue-per-megawatt economics of the operation.
DSX OS is an open-source operating layer providing lifecycle management, multi-tenant operations, runtime consistency, health monitoring, and automated remediation across the full factory fleet. Naver, which has operated its own GPU supercomputing clusters since before most Korean cloud operators existed, brings existing operational expertise that DSX OS integrates rather than replaces.
DSX Flex connects the AI factory to grid signals — load shedding, demand-response events, real-time pricing — and dynamically adjusts workloads in response. At gigawatt scale, a data center is a meaningful grid participant; DSX Flex positions Naver as an entity that can absorb renewable energy when it is available and reduce draw during grid stress, which matters for securing power agreements in Korea's constrained grid environment.
DSX Exchange provides the communication layer integrating IT signals, operational technology, and AI factory management agents — the plumbing that allows all five other layers to act as a single coordinated system rather than a collection of siloed subsystems.
Why Nvidia Chose Naver Over a New-Build Partner
Jensen Huang addressed this directly at the 1784 media scrum, and his answer is technically specific: Naver was one of the earliest GPU-cluster SuperPOD customers in Asia, meaning the company has years of operational data on high-density Nvidia cluster behavior at scale. Building an AI factory on DSX with an operator who has never run a large GPU cluster is a substantially riskier proposition than building one with an operator who has run one for years. Naver's Gak Sejong campus, opened in November 2023, spans 290,000 square meters, can accommodate 600,000 servers, and was designed with a planned maximum capacity of 270 megawatts. The first phase of a 55MW AI factory expansion onto that campus is, in structural terms, an expansion of an existing proven facility rather than a greenfield build.
Naver also brings the model stack. Its HyperCLOVA X large language model — built on 6,500 times more Korean-language training data than GPT-3 — is being further advanced using Nvidia's Nemotron 3 Ultra open model. Naver is the first Korean company admitted to the Nvidia Nemotron Coalition, contributing to pretraining, post-training, and reinforcement-learning research. That technical depth means Naver's AI factory is not merely reselling GPU compute: it can offer customers sovereign Korean-language AI services that no foreign cloud operator's infrastructure can replicate with equivalent cultural and regulatory fidelity.
The Seoul World Model — a digital twin of Seoul's road environments built on Nvidia Cosmos and trained on 1.2 million panoramic images collected across the city — extends this advantage into physical AI. Naver's proprietary urban mapping data is not available to any external party, and the Seoul World Model sits at the center of the companies' joint physical AI roadmap including robotics and autonomous systems.
Capacity Roadmap: 55MW to 1GW on Four-Year Timeline
The published capacity roadmap, confirmed in Nvidia's June 7 press release and the June 8 joint announcement, is:
- H1 2027: 55MW goes online at Gak Sejong, Sejong City, South Korea
- H2 2027: 100MW phase
- 2028: 200MW phase
- Long-term target: 1GW across Asia, Middle East, and Europe
At 1GW, the combined infrastructure would be roughly four times the planned maximum capacity of Gak Sejong's original build. Nvidia's press release describes Naver as a "key global partner sharing both returns and risks" — language that signals a co-investment structure rather than a simple supply agreement, though the financial terms of the risk-sharing arrangement have not been publicly disclosed.
The 25 trillion won revenue target by 2030 — approximately $18 billion at current exchange rates — was cited in Korean financial press covering the announcement. Huang framed the deal's geographic scope during the Chzzk livestream: Nvidia chose Naver and plans to expand AI factories not only in Korea but across Asia. Naver VP Raj Mirpuri confirmed the sovereign AI positioning: the factories are intended to give government and enterprise customers a compliant alternative to US-headquartered hyperscalers for workloads requiring in-region data residency.
What Does Naver's Partnership Mean for Cloud Buyers in Asia?
For enterprises, governments, and AI developers in Asia, the Middle East, and Europe considering GPU cloud infrastructure, the Naver deal introduces a competitive option that did not exist at the same combination of scale, regulatory compliance, and Korean-language AI capability before June 8. Naver Cloud's AI factory will, when the first 55MW phase opens in 2027, offer the same Vera Rubin and Blackwell GPU infrastructure available on AWS or Azure — but within Korea's data-residency jurisdiction and bundled with HyperCLOVA X models optimized for Korean language and institutional context.
For buyers whose workloads require compliance with Korean or regional data sovereignty regulations — which prohibit certain data from leaving the jurisdiction — AWS Seoul Region and Azure Korea Central already exist, but neither offers the depth of Korean-language AI model integration that Naver Cloud can provide on its own infrastructure. The more direct competitive implication is for buyers evaluating regional cloud providers in markets where Naver is targeting expansion: the Middle East and Europe. In both regions, sovereign AI requirements are active policy priorities, and Naver's combination of a proven hyperscale data center operation, a non-US-headquartered ownership structure, and a DSX-certified factory platform positions it as a credible alternative to CoreWeave — the US-based specialized GPU cloud — for customers in those regions who prefer a non-US anchor.
The deal's ecosystem breadth also matters for developers. Naver's participation in the Nemotron Coalition means developers building Korean-language AI applications will have access to HyperCLOVA X advances incorporating Nvidia's open-model research. The planned AI Agent Platform for Korea, expected in the second half of 2026, is built on Nvidia NemoClaw blueprints and will run on the same infrastructure stack.
SK Telecom Signed Same DSX Deal Same Day: Korea as Nvidia's Sovereign-AI Anchor Region
The Naver deal is not the only gigawatt-scale DSX commitment Nvidia signed in Korea on June 7–8. SK Telecom announced on June 8 that it would pursue a parallel full-stack AI cloud cooperation with Nvidia, also at gigawatt scale, also built on the DSX platform, with SK Telecom's first AI factory planned to come online in 2027. The fact that two Korean operators signed structurally equivalent gigawatt DSX commitments on the same day — with a third Korean announcement (Hyundai's 50,000-GPU AI factory deploying Blackwell for automotive applications) also in the same week — makes clear that Nvidia's Korea strategy is not about individual deals but about establishing Korea as a regional anchor for the DSX sovereign-AI ecosystem.
The strategic logic from Nvidia's perspective is coherent. Google's TPU 8t and TPU 8i, announced at Google Cloud Next 2026, target frontier training and inference and have won Anthropic, Meta, and OpenAI as anchor customers. AWS Trainium continues to scale. Nvidia's data-center segment generated $193.7 billion of its $215.9 billion in total fiscal 2026 sales — a figure built almost entirely on GPU infrastructure demand. The risk is not that hyperscalers will abandon Nvidia GPUs immediately; it is that large customers will build mixed-portfolio strategies, as Anthropic has done explicitly with its tri-platform approach using Google TPU, AWS Trainium, and Nvidia GPUs in parallel. Nvidia's sovereign-AI DSX partner network — which now includes operators in Korea, the Middle East, Japan, and Europe — creates a class of regional cloud operators whose entire value proposition depends on the Nvidia GPU stack and who therefore have no incentive to diversify toward custom silicon. Korea's simultaneous DSX signings are the densest cluster in that network.
Frequently Asked Questions
What is the Nvidia DSX platform and why does the Naver deal use it?
Nvidia DSX is a full-stack AI factory design platform that coordinates compute hardware, networking, storage, power management software (DSX MaxLPS), a digital twin simulation layer (DSX Sim), and open-source lifecycle management software (DSX OS) as a single codesigned product. Building on DSX rather than assembling a GPU cluster independently allows Naver to deploy a validated, generation-specific architecture and use DSX MaxLPS to provision up to 40 percent more GPUs within the same power budget — a significant unit-economics advantage at the 55MW to 200MW scale the roadmap targets.
What is sovereign AI and why does it give Naver an advantage over AWS or Google Cloud?
Sovereign AI refers to AI infrastructure controlled by a specific nation or organization, using locally compliant data, models, and computing resources. For customers in Korea, the Middle East, or Europe with regulatory requirements that data not leave their jurisdiction, Naver Cloud offers HyperCLOVA X models trained on Korean-language data and operated within Korean infrastructure — a compliance and cultural-fit advantage that neither AWS nor Google Cloud, both headquartered in the United States, can fully replicate for Korean-language AI workloads.
How does this deal connect to Nvidia's competition with Google's TPU chips?
Google's TPU 8t and TPU 8i, announced in April 2026, have attracted Anthropic, Meta, and OpenAI as infrastructure customers. Nvidia's response includes building a network of regional sovereign-AI cloud operators — including Naver and SK Telecom in Korea — whose business model depends on the Nvidia GPU and DSX stack and who therefore have no incentive to adopt TPU-based alternatives. Regional operators serving sovereign-AI markets have a specific reason to stay on Nvidia hardware: Google's US-headquartered cloud cannot match their data-residency compliance positioning.
When will Naver's AI factory be operational?
The first 55MW phase at Gak Sejong in Sejong City, South Korea, is scheduled to open in the first half of 2027. A second phase targeting 100MW is planned for later in 2027, with a 200MW phase in 2028. The full 1-gigawatt target and the 25 trillion won revenue goal do not carry a fixed completion date in the public announcement.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
