Korea–Nvidia

Nvidia and Korea's government announced roughly 260,000 GPUs for the state, Samsung, SK, Hyundai and NAVER in October 2025, a plan to 2030 per the ministry; only the state's buys have figures.

A deal between Korean government, Nvidia, Samsung, SK Telecom, Hyundai Motor Group, NAVER

Why it matters

South Korea and Nvidia announced a plan for roughly 260,000 GPUs. Only the government's share has documented purchases and funding, with 13,136 GPUs bought in 2025 and 9,704 awarded in 2026. The companies' 210,000 have no contract, price or delivery date on record.

Tap a figure for the line in its source.

Nvidia gives

  • disputedGPUs for public use, bought through Korean cloud companies — 52,000 in some government documents → Korean government
  • disputedGPUs for a semiconductor 'AI factory' — 'up to' 50,000 in Nvidia's English blog; no price disclosed → Samsung
  • disputedGPUs for SK Group's AI factory — 'up to' 50,000 in Nvidia's English blog, which says the factory can host up to 60,000; no price disclosed → SK Telecom
  • Blackwell GPUs, part of a physical-AI collaboration with Nvidia and the government that Nvidia's release puts at about $3 billion → Hyundai Motor Group
  • disputedGPUs for NAVER Cloud — up to 60,000 in Nvidia's Korean text; no price disclosed → NAVER

Korean government gives NAVER

  • pays NAVER Cloud and four other cloud companies to buy, install and run the GPUs, and lets them keep a share for their own use: ₩1.46 trillion for 13,136 in 2025, ₩2.08 trillion for 9,704 awarded in 2026

Oct 29, 2025 reported by Bloomberg a day before Nvidia announced itupdated Nov 18, 2026announced disclosed by the parties

What happened

At the APEC summit in October 2025, Nvidia and South Korea's government announced a plan to add roughly 260,000 Nvidia GPUs, or graphics processing units. The plan calls for about 50,000 for the government, 50,000 each for Samsung, SK Group and Hyundai, and 60,000 for NAVER's cloud arm. The announcement gave no end date. The ministry later described it as a plan through 2030. No price, purchase agreement or delivery schedule for the plan is public. Nvidia's own statements disagree on whether each number is a minimum or a maximum. The government's own purchases can be checked. It pays Korean cloud companies to buy and run GPUs and lets them keep some of the chips. The program bought 13,136 in 2025, and deliveries had begun by the summit. Nvidia then presented them as the plan's first wave. The government awarded a further 9,704 in June 2026. By July 2026, the ministry counted about 35,000 toward its 50,000, using B200-equivalents. That measure weights each chip by its speed relative to Nvidia's B200. The ministry counted every purchase in full, including the roughly 6,000 GPUs the cloud companies keep.

Sources 25 documents — 11 government statements · 7 other documents · 4 press reports · 2 company releases · 1 SEC filing

Research trail Found 260 documents and read 234 of them while checking this deal and related ones.

  1. company releasenvidianews.nvidia.com
    States 2 figures: initial deployment · $3B hyundai investment
  2. nabo.go.kr · government statementnabo.go.kr
    States 2 figures: procurement 2025 · government use 2025
  3. Korea (government) · government statementkorea.kr
    States allocation split
  4. Korea (government) · government statementkorea.kr
    States horizon
  5. Korea (government) · government statementkorea.kr
    States first batch
  6. eiec.kdi.re.kr · other documenteiec.kdi.re.kr
    States procurement 2026
  7. Korea's science ministry · government statementmsit.go.kr
    States government plan
  8. Korea (government) · government statementkorea.kr
    States progress jul 2026
  9. government statementmsit.go.kr
    Cited by the record; no single figure is quoted from it.
  10. company releaseblogs.nvidia.com
    Cited by the record; no single figure is quoted from it.
  11. other documentblogs.nvidia.co.kr
    Cited by the record; no single figure is quoted from it.
  12. SEC filingsec.gov
    Cited by the record; no single figure is quoted from it.
  13. press report (archived copy)theedgemalaysia.com
    Cited by the record; no single figure is quoted from it.
  14. government statementkorea.kr
    Cited by the record; no single figure is quoted from it.
  15. government statementm.korea.kr
    Cited by the record; no single figure is quoted from it.
  16. government statementevaluation.go.kr
    Cited by the record; no single figure is quoted from it.
  17. government statementnars.go.kr
    Cited by the record; no single figure is quoted from it.
  18. press reportkoreaherald.com
    Cited by the record; no single figure is quoted from it.
  19. press reportasiae.co.kr
    Cited by the record; no single figure is quoted from it.
  20. press reportasiae.co.kr
    Cited by the record; no single figure is quoted from it.
  21. other documentbiz.sbs.co.kr
    Cited by the record; no single figure is quoted from it.
  22. other documentedaily.co.kr
    Cited by the record; no single figure is quoted from it.
  23. other documentkakaocorp.com
    Cited by the record; no single figure is quoted from it.
  24. other documente4ds.com
    Cited by the record; no single figure is quoted from it.
  25. other documentkedglobal.com
    Cited by the record; no single figure is quoted from it.

The record's own summary of its sourcing: Nvidia release + English and Korean blogs (Oct 30–31 2025) and Q3 FY2026 results release; MSIT English releases (Sep 2025, Oct 2025), working-group release (Nov 2025), briefings (Dec 2025, Jul 2026) and award release (Jun 2026, via KDI); NABO settlement analysis (Jul 2026); Presidential Office briefing (Oct 2025); power authorities' briefing (Dec 2025); NARS (2026); evaluation.go.kr; Korea Herald, Asiae, SBS, Edaily, e4ds, KED Global, Bloomberg via The Edge (Wayback); Kakao (read 2026-10) · as of Jul 2026

The evidence 10 sourced figures · 10 disputes

The record's headline Korea–Nvidia: roughly 260,000 GPUs announced at APEC (Oct 2025) — ~50,000 for the government, 50,000 each for Samsung, SK and Hyundai, 60,000 for NAVER — a plan to 2030 with no contract or price on file; the government's own share is the only part with counted units: 13,136 bought in 2025, 9,704 awarded in 2026

Size
~260,000 GPUs announced ('over a quarter-million' to 'Over 260,000'), no price · government share: 13,136 bought (2025) + 9,704 awarded (2026)
When
Oct 30–31 2025 (APEC, Gyeongju) → Nov 2025 working group → Dec 2025 first public tranche → Jun 2026 second purchase award → plan to 2030
Status
announced (a plan to 2030) · government share: 13,136 bought in 2025, 9,704 awarded Jun 2026, ~35,000 B200-equivalents counted as secured by Jul 2026, the operators' kept share included · the companies' 210,000: no delivery figure in any document
Structure
a state-brokered allocation: one announcement gives five allottees a GPU count, and an MSIT–company working group follows the rollout; the government buys its share by paying cloud operators to buy, install and run the GPUs, which keep part of them for commercial use in place of operating fees; the companies buy theirs under no disclosed contract
Programme
Korea–Nvidia GPU plan
Parties
Korean government broker (MSIT's GPU working group) + buyer of the public share through cloud operators · Nvidia supplier · Samsung allottee — a semiconductor AI factory · SK Telecom SK Group's allottee — an AI factory and an industrial AI cloud · Hyundai Motor Group allottee — a physical-AI factory in an ~$3B collaboration · NAVER allottee (NAVER Cloud) + operator of government GPUs
Record
disclosed / primary source · as of Jul 2026

Flows 5 as the record states them

  • Korean government → NAVERgpu-procurementNAVER Cloud's part of the government's purchases — 3,056 H200 in 2025 (2,296 for government use, 760 kept for its own) and 1,008 Vera Rubin + 3,112 B300 in the 2026 award; the government pays, the operator installs and runs them. The rest of the 2025 round's ₩1.46 trillion went to NHN Cloud and Kakao, and the 2026 round's ₩2.08 trillion is shared with Samsung SDS and Elice — none of them on this map
  • Samsung → Nvidiachip-spendover 50,000 GPUs for a semiconductor AI factory ('up to' that many in Nvidia's English blog) — no price, contract or schedule disclosed
  • SK Telecom → Nvidiachip-spendover 50,000 GPUs for SK Group's AI factory ('up to' that many in Nvidia's English blog, which has the factory 'capable of hosting up to 60,000'), with an industrial AI cloud on RTX PRO 6000 GPUs — no price disclosed; whether any of it overlaps sk-nvidia's Jul 2026 plans is not stated
  • Hyundai Motor Group → Nvidiachip-spend50,000 Blackwell GPUs, inside an 'approximately $3 billion' physical-AI collaboration — no price for the GPUs disclosed
  • NAVER → Nvidiachip-spendover 60,000 GPUs for NAVER Cloud (up to 60,000 in Nvidia's Korean blog) — no price disclosed; the 200 MW AI factory's ~100,000 GPUs are nv-naver's, and no document on file says whether the two overlap

Resolved figures 10 each with the verbatim line from its document

  • allocation-split260,000
    the rounded ledger, as MSIT recorded it when it launched a GPU working group with Nvidia Korea, Samsung Electronics, Hyundai, SKT and NAVER (Nov 28 2025): government 50,000; Samsung, Hyundai and SKT 50,000 each; NAVER 60,000 — 260,000 GPUs it describes as secured on the occasion of APEC ('APEC 등을 계기로 확보한'). Whether each share is a floor or a cap, whether the government's is 50,000 or 52,000, and how the total itself is worded all stay contested (allocation-qualifiers, government-share, headline-total)
    “류제명 제2차관은 엔비디아코리아·삼성전자·현대차·SKT·네이버와 GPU 워킹그룹을 발족하고, APEC 등을 계기로 확보한 첨단 GPU 26만 장의 이행 계획과 향후 전략을 논의했습니다. 정부- 5만장 삼성전자·현대차·SKT- 5만장 네이버- 6만장”
    MSIT GPU working-group release (korea.kr, Nov 28 2025) · korea.kr
  • horizon2030
    a plan running to 2030, in MSIT's answer at its Dec 12 2025 work-plan briefing — not a delivery schedule: the same answer says each company is working out a yearly supply plan with the ministry in a task force, and no schedule is published. NABO ('민관이 함께 2030년까지 26만장의 GPU를 확보한다는 계획') and the Korea Herald ('a broader 260,000-GPU supply plan that runs through 2030') say the same
    “GPU 26만 장은 2030년까지의 계획입니다.”
    MSIT 2026 work-plan briefing transcript (korea.kr, Dec 12 2025) · korea.kr
  • initial-deployment13,000
    Nvidia's release counts the government's 2025 purchase as the plan's start: MSIT's program 'will grow over the next several years from an initial deployment of 13,000' GPUs at NAVER Cloud, NHN Cloud and Kakao — the batch NABO counts as 13,136 (procurement-2025), of which 3,056 H200 + 2,424 B200 had arrived by the end of October 2025. The same paragraph has MSIT deploying 'up to 50,000' where the release's summary says 'over 50,000' (allocation-qualifiers); Asiae's English edition calls the batch 'separate from the 260,000' (government-batch-inside)
    “The AI infrastructure deployment will grow over the next several years from an initial deployment of 13,000 NVIDIA Blackwell and other GPUs by NVIDIA Cloud Partner NAVER Cloud, together with NHN Cloud and Kakao Corp., to expand computing infrastructure on the nation’s sovereign clouds”
    Nvidia release (Oct 30 2025) · nvidianews.nvidia.com
  • procurement-202513,136
    13,136 Nvidia GPUs — 10,080 B200 + 3,056 H200 — bought with the 2025 first supplementary budget of ₩1.46 trillion, 3,136 more than the 10,000 the budget had assumed; by operator (NABO's table, from MSIT data): NAVER Cloud 3,056 H200, NHN Cloud 7,656 B200, Kakao 2,424 B200. The government pays in full; each operator receives the GPUs and installs them in its own data center. The H200s and Kakao's B200s had arrived by the end of October 2025, NHN's in 2026; users were supplied from March–April 2026. MSIT and Nvidia round it to 13,000
    “구체적으로 사업자가 확보한 GPU는 총 13,136장 규모로서 예산 편성 보다 3,136장 많은 규모이며, 구체적으로는 엔비디아(NVIDIA) B200 10,080장, H200 3,056장 규모이다.”
    NABO, settlement analysis of MSIT's 2025 accounts (Jul 2026) · nabo.go.kr
  • government-use-202510,456
    10,456 of the 13,136 are the government's to use (8,160 B200 + 2,296 H200); the operators keep the other 2,680 — 20.4% on average — for commercial use, in return for bearing five years of operating costs (floor space, power, network, maintenance). 10,056 is the same pool less 400 held as spares against failures, and MSIT's '1만 장' (10,000) rounds it.
    “10,056장(정부활용분 10,456장 중 장애대응을 위한 예비물량 400장 제외)”
    NABO, settlement analysis of MSIT's 2025 accounts (Jul 2026) · nabo.go.kr
  • first-batch2,296
    the first public tranche, opened to companies, universities and research institutes on Dec 22 2025 (applications until Jan 28 2026): 2,296 H200 + 2,040 B200 — 4,336 in all, the Korea Herald's total — out of the 10,000 the government will offer as a cloud service. A NABO footnote prints the B200s as '255대(2,050장)' (first-batch-b200)
    “그중에서 먼저 구축할 H200 2296장, B200 2040장은 산·학·연의 시급한 AI 혁신수요에 우선 배분한다.”
    MSIT via korea.kr (Dec 22 2025) · korea.kr
  • procurement-20269,704
    the 2026 round (₩2.08 trillion): 9,704 GPUs — 2,016 Vera Rubin + 7,688 B300 — to be secured and built by NAVER Cloud (1,008 Vera Rubin + 3,112 B300), Samsung SDS (1,008 + 2,016 B300) and Elice (2,560 B300), chosen on Jun 8 2026 from five bidders. 2,016 Vera Rubin + 4,360 B300 (6,376, our sum) are for government use and 3,328 B300 for the operators' own. The budget had planned 15,000 B200; MSIT cites rising memory and storage prices and counts the award as about 19,000 B200-equivalents, about 30% above that target ('약 30%를 상회하는 성능' — procurement-2026-equivalent). A plan to secure, not a delivery
    “이번에 선정된 CSP 3社와 협력하여 확보할 GPU는 베라루빈 2,016장과 B300 7,688장 총 9,704장으로, 네이버클라우드(주)는 베라루빈 1,008장과 B300 3,112장, 삼성에스디에스(주)는 베라루빈 1,008장과 B300 2,016장, (주)엘리스그룹은 B300 2,560장을 확보·구축할 계획이다.”
    MSIT release (Jun 8 2026), KDI's verbatim repost · eiec.kdi.re.kr
  • government-plan28,000
    the government's itemized plan, Sep 8 2025 — seven weeks before the summit: 'more than 50,000 GPUs through public–private collaboration' — 28,000 by procurement in 2025–26, 9,000 in Supercomputer No. 6 by 1H 2026 and more than 15,000 at the National AI Computing Center by 2028; the parts sum to the 52,000 other government documents carry (government-share). Against it, the documents show 13,136 + 9,704 = 22,840 GPUs from the two procurement rounds (our sum) and 8,496 in Supercomputer No. 6 (NABO)
    “First, to address urgent demand from industry, academia, and research institutes, the government will secure 28,000 GPUs through procurement programs in 2025–2026 and an additional 9,000 GPUs by the first half of 2026 through the Supercomputer No. 6 construction project.”
    MSIT English release (Sep 8 2025) · msit.go.kr
  • progress-jul-2026210,000
    MSIT's mid-2026 ledger (Jul 16 2026 pre-briefing; the answer is its AI-infrastructure policy director's): of the 260,000 target the companies plan 210,000 and the government 50,000, of which about 35,000 are secured — 'B200 기준으로 환산', converted to B200-equivalents because the chip generations differ — and about 11,600 B200-equivalents, the 2025 purchase, in use. SBS's account of the same report: ~11,600 (2025 purchase) + ~4,200 (Supercomputer No. 6's ~9,000 GPUs) + ~19,400 (the ~10,000 of 2026) = ~35,200. The 31,336 physical GPUs on the documents (13,136 + 9,704 + 8,496, our sum) read as ~35,000 on the new yardstick. That count includes the operators' own share. MSIT's 11,600 is the quantity bought with last year's supplementary budget ('작년 추경으로 산 물량'; SBS: the ~13,000 GPUs brought in, converted), which the government's 10,456 alone could reach only if an H200, the older chip, counted as 1.5 B200s; the ≈19,000–19,400 rates all 9,704, where the government's 6,376 come to 14,380 on MSIT's ratings (procurement-2026-equivalent; both our arithmetic). So 6,008 of the GPUs behind the ~35,000 (2,680 + 3,328, our sum) are the ones the operators keep for commercial use in return for bearing the operating costs (government-use-2025, procurement-2026). No delivery figure for the companies' 210,000 appears in any document on file
    “26만 장 목표 중에서 지금 사실 민간이 지금 21만 장을 계획하고 있고요. 정부는 5만 장을 계획하고 있는데 그 5만 장 중에서 지금 오늘 성과로서 지금 브리핑문에도 나왔지만 3만 5,000장 정도, B200 기준으로.”
    MSIT pre-briefing transcript (korea.kr, Jul 16 2026) · korea.kr
  • hyundai-investment$3B
    the only allottee with a dollar figure: Hyundai and Nvidia, working with the government on a national physical-AI cluster, put their collaboration at 'an approximately $3 billion investment' — an NVIDIA AI Technology Center, a Hyundai Physical AI Application Center and regional AI data centers — beside 50,000 Blackwell GPUs. Who invests how much is not said, and it is not a price for the GPUs
    “In support of the Korean government’s initiative to build a national physical AI cluster, Hyundai Motor Group and NVIDIA will work closely with government stakeholders to accelerate ecosystem development. This will result in an approximately $3 billion investment to advance the physical AI landscape in Korea.”
    Nvidia release (Oct 30 2025) · nvidianews.nvidia.com

Contested figures 17 not reconciled — every position side by side, with its source and tier

headline-total4 positions, side by side
'Over 260,000' — the subheadline of Nvidia's newsroom release (Oct 30 2025: 'Korea Government, Computing and Manufacturing Leaders Adding Over 260,000 NVIDIA GPUs'); MSIT's English release the same day: Korea 'will secure more than 260,000 of NVIDIA's latest GPUs', 'approximately 50,000' of them public and 'more than 200,000' private Nvidia release + MSIT English release (Oct 30 2025)
'260,000', unqualified — Nvidia's English blog (Oct 31 2025: 'Samsung, SK Group, Hyundai, NAVER and Korea's government to deploy 260,000 NVIDIA GPUs'), MSIT's working group and its Dec 2025 briefing ('GPU 26만 장'); 'about 260,000' — the Korea Herald (Feb 19 2026), whose own split (52,000 + 3 × 50,000 + 60,000) adds to 262,000 Nvidia English blog; MSIT; Korea Herald
over 250,000 — Nvidia's Korean blog (Oct 31 2025: '25만 개 이상의 NVIDIA GPU를 추가 도입'); the first paragraphs of Nvidia's newsroom release ('over a quarter-million NVIDIA GPUs') and English blog ('more than a quarter-million NVIDIA GPUs'), below subheadlines that give the higher figure; and the Q3 FY2026 results release Nvidia furnished to the SEC ('over a quarter-million NVIDIA GPUs'): the lowest floor, and the only wording Nvidia has put before the SEC Nvidia Korean blog; Nvidia release + English blog (first paragraphs); Nvidia Q3 FY2026 results release (furnished)
'as many as 200,000 high-performing GPUs by 2030' — Bloomberg's preview the day before, from people familiar with the matter; also the first report that Huang would 'unveil new contracts to supply' Samsung and Hyundai, with SK named too Bloomberg via The Edge (Oct 29 2025, Wayback)
government-share2 positions, side by side
50,000 — MSIT's working group ('정부- 5만장', Nov 2025) and its Jul 2026 briefing ('정부는 5만 장을 계획'); 'approximately 50,000' in MSIT's English release; Nvidia: 'over 50,000' (release summary), 'up to 50,000' (release body), 'as many as 50,000' (English blog) MSIT (Oct 2025, Nov 2025, Jul 2026); Nvidia release + English blog
52,000 — the government's performance-evaluation page ('정부 5.2만장, 민간 21만장 ... (엔비디아 발표 등 기준)'), NABO ('2028년까지 5.2만장 이상의 GPU를 확보할 계획') and the Korea Herald ('52,000 for the government'): the Sep 2025 plan's 28,000 + 9,000 + 15,000 (government-plan). With the companies' 210,000 it makes 262,000, and the page cites Nvidia's announcement, which gives the government 'over 50,000' evaluation.go.kr; NABO (Jul 2026); Korea Herald (Feb 2026)
allocation-qualifiers2 positions, side by side
floors — Nvidia's release summary: government 'over 50,000', Samsung 'over 50,000', SK Group 'over 50,000', NAVER Cloud 'over 60,000' (Hyundai a flat 50,000); Nvidia's Korean blog: Samsung and SK '5만 개 이상' Nvidia release + Korean blog (Oct 2025)
caps — the same release's body: MSIT plans 'up to 50,000'; Nvidia's English blog: the government 'as many as 50,000', Samsung, SK and Hyundai 'up to 50,000 NVIDIA GPUs apiece', SK's AI factory 'capable of hosting up to 60,000 GPUs' (NAVER 'more than 60,000'); Nvidia's Korean blog: NAVER Cloud '최대 6만 개' (up to 60,000). One announcer, both readings — no document says whether 260,000 is a floor or a ceiling Nvidia release + English and Korean blogs (Oct 2025)
government-batch-inside2 positions, side by side
inside the 260,000 — Nvidia's release (the program 'will grow ... from an initial deployment of 13,000'); at MSIT's Dec 12 2025 briefing, asked about the '1만 3,000여 장' of the 260,000 Nvidia promised, the answer counts the government's 13,000 this year and 37,000 by next year inside the 2030 plan; NABO counts the 2025 purchase inside the government's 52,000 Nvidia release; MSIT briefing (Dec 12 2025); NABO
separate — Asiae's English edition (Dec 1 2025), on the 13,000-unit batch bought with the ₩1.46 trillion supplementary budget: 'It is separate from the 260,000 units that Nvidia CEO Jensen Huang promised to supply'. On this reading the government's purchases add to the 260,000 rather than fill part of it Asiae English (Dec 1 2025)
legal-nature2 positions, side by side
supply contracts signed — the Presidential Office's summary of the closed session with Huang (Oct 31 2025): there was much talk that Nvidia, having signed supply contracts for more than 260,000 ('26만장 이상 공급 계약을 체결해서'), would set off Korea's AI industry. Reported speech; no contract is on file anywhere Presidential Office briefing (korea.kr, Oct 31 2025)
a plan, not a contract — MSIT: 'GPU 26만 장은 2030년까지의 계획입니다' (Dec 2025); a column in Asiae by Park Young-sun, a former minister of SMEs and Startups: the number does not lead straight to immediate supply or a confirmed contract (즉시 공급, 계약 확정) and is closer to an allocation ('할당량 성격에 가깝다'). No contract, price or delivery schedule is on file MSIT briefing (Dec 12 2025); Asiae column by Park Young-sun (Nov 25 2025)
pledge-valueone entry, in its own words
no price in any party's document; industry estimates only — '11 to 14 trillion won will be invested in the introduction of 260,000 GPUs alone' (e4ds), 'valued at up to 14 trillion won ($9.8 billion)' on industry estimates that assume 'full delivery by 2030' (KED Global): unit-price × count arithmetic, not a deal value e4ds; KED Global (Oct 31 2025)
power-requiredone entry, in its own words
~500 MW for all 260,000 by 2030 — MSIT's Dec 12 2025 briefing ('GPU 26만 장이 2030년까지 들어와도 이게 한 500MW 정도') vs ~0.7 GW once cooling is counted — the power authority at the joint winter-supply briefing (Dec 2025), on 80% utilisation and a 60/40 split between GPUs and cooling, under 1% of 2030 demand vs up to 1 GW of continuous power — the National Assembly Research Service's 2026 audit-issues report, which warns of shortfalls where the GPUs sit MSIT (Dec 12 2025); power authorities' briefing (Dec 2025); NARS (2026)
first-batch-b200one entry, in its own words
2,040 B200 — MSIT (Dec 22 2025) and the Korea Herald; Kakao reports 255 nodes, 'GPU 2,040장', built — vs '255대(2,050장)' in a NABO footnote, which contradicts its own server count (255 × 8 = 2,040, our arithmetic); no erratum MSIT via korea.kr; Kakao; NABO (Jul 2026)
procurement-2026-equivalentone entry, in its own words
≈19,000 B200-equivalents ('B200 약 1.9만장 규모') — MSIT (Jun 8 2026), about 30% above the 15,000-B200 target, rating B200 at 9 PF, B300 at 13.5 PF and Vera Rubin at 35 PF — vs ≈19,400 (SBS, Jul 16 2026, from MSIT's second-half work plan). On MSIT's ratings 2,016 × 35 + 7,688 × 13.5 = 174,348 PF ÷ 9 = 19,372 (our arithmetic), which rounds to either; neither replaces the 9,704 physical GPUs MSIT release (Jun 8 2026); SBS Biz (Jul 16 2026)
government-cumulativeone entry, in its own words
37,000 by end-2026 — MSIT's Dec 12 2025 briefing ('내년까지는 정부 차원에서 3.7만 장'), the pre-award plan's 13,000 + 15,000 + 9,000 (our arithmetic) vs 32,000 cumulative in 2026 — the government's performance page ('’26년 정부 첨단 GPU 3.2만장(누적)') vs ~20,000 by end-2026 and ~30,000 by end-2027 — the Deputy PM for the economy, Koo Yun-cheol, at a Jun 26 2026 meeting where President Lee called the pace too slow (Edaily). The documents' physical units: 13,136 + 9,704 + 8,496 = 31,336 (our sum); MSIT's own July count is ~35,000 B200-equivalents MSIT briefing (Dec 12 2025); evaluation.go.kr; Edaily (Jun 26 2026)

Scope note what this record leaves to other records

NOT this record: Nvidia's $1B share placement in NAVER and Brookfield's financing → nv-naver. SK Group–Nvidia's $500B-plus partnership and its AI factory (sk-nvidia): MSIT's Jul 2026 English release puts that factory's size at 'up to 2 GW' — a ceiling, which belongs to sk-nvidia's skt-ai-factory cell, not this record; and the $500 billion SK figure is not the $500 billion of Nvidia's compute-financing platforms with Brookfield and others. The national AI data-center build (₩550 trillion for 8.4 GW by 2029, ~₩1,000 trillion with a further 10 GW) → the krgovernment entity. The government share's other legs are carried only as counts: Supercomputer No. 6 (8,496 GPUs incl. GH200 on a ₩382.5 billion HPE contract; 600 vs 614 PF and ₩74.6 billion vs ₩78 billion of maintenance remain disputed) and the National AI Computing Center (more than 15,000 GPUs by 2028; its equity split reads 'to public 70%' in MSIT's English release and '공공 30% 미만, 민간 70% 초과' in NABO's Korean — not carried here). Not carried here: the sector split of the first 4,224 GPUs allocated, where MSIT's card and Edaily's NIPA data move 72 GPUs between industry and academia; the 2026 budget's ₩2.0831 trillion (NABO) vs MT's ₩2.0805 trillion; Korea's pre-pledge GPU base (45,000–65,000 in the press, no primary inventory); press power estimates (312 MW–1 GW on varying per-GPU and PUE assumptions); MSIT's 2027 budget proposal funds 'GPU 1만 장' for sovereign-model development (Sep 2026). Premise corrections: one reading puts Koo's '올해까지는 2만 장' at end-2025, but Edaily's article is dated Jun 26 2026 (government-cumulative). Not adopted: the Presidential Office's line that Nvidia signed supply contracts (reported speech, legal-nature); a dollar value for the pledge (pledge-value); DCD's 50,000-by-2030 figure for the National AI Computing Center (the national target attached to one site); DCD's NAVER deploying its 60,000 GPUs in partnership with LG AI Research, SK Telecom and NC AI (two lines of Nvidia's release run together).

More between these companies

  1. Jul 24, 2026 updated Nov 21
  2. Jul 24, 2026 updated Nov 21
  3. Oct 1, 2025 updated Nov 19

Everything between Nvidia and SK Telecom → · Everything between Nvidia and NAVER → · Everything between Samsung and SK Telecom →