Korea Says Nvidia Keeps 260,000-GPU Pledge After Jensen Huang Quits; Only 10,000 Confirmed in Use
18 hour ago / Read about 23 minute
Source:TechTimes

Brecht Corbeel/Unsplash

South Korea's science ministry said on Nov. 13 that Nvidia reaffirmed its 260,000-GPU plan one day after Nvidia announced Jensen Huang's resignation as CEO, but the only in-service figure Seoul has given is 10,000 chips, reported Nov. 9. Korean startups, researchers and companies waiting on state-funded compute need to know which of the ministry's numbers describe chips in racks and which describe chips on order.

Deputy Prime Minister and Science Minister Bae Kyung-hoon announced the reaffirmation in a Facebook post titled "Jensen Huang's resignation, but cooperation continues," according to Korean press reports. He said Nvidia reaffirmed the plan at the corporate level, that both sides agreed to continue existing projects without disruption, and that the government's own 50,000-GPU plan is on track. The reports do not quote a statement from Nvidia itself.

Ministry Counts 21,000 of 35,000 Direct-Purchase Chips as Secured, Not Delivered

The reassurance followed a meeting with Nvidia on pending issues, which Bae mentioned in his post, Korean press reported. The reports reviewed do not give the meeting's participants.

The headline numbers cover only the government's own slice of the program:

  • The government's plan covers up to 50,000 GPUs in total.
  • Of roughly 35,000 chips the ministry buys directly, about 21,000 are secured, or around 60%, in B200-equivalents.
  • The remaining 14,000 or so are expected by early 2027.
  • A separate 15,000 GPUs and NPUs are planned through the National AI Computing Center, a target its own roadmap dates to 2028, with 50,000 units across the national AI ecosystem by 2030.

The second and third lines add up: 21,000 plus 14,000 equals the 35,000 direct purchases. The fourth line is a multi-year target, not chips arriving now. The Korean report's wording says "secured," not "delivered" or "installed," and it does not say how many of the 21,000 are running in data centers.

What Does B200-Equivalent Mean for Korea's GPU Count?

A B200-equivalent is a normalized unit that lets chips from different Nvidia generations be added together. The yardstick is peak compute: one B200 delivers roughly 9 petaFLOPS at 8-bit precision. Newer parts count for more than one unit each.

The ministry's earlier tender shows how far a physical chip count and an equivalents count can diverge. Its 2026 "AI Highway" program will install 9,704 GPUs across three providers, Samsung SDS, Naver Cloud and Elice Group, in a mix of B300 and Vera Rubin parts, and TechTimes reported that as roughly 30% above a target set as 15,000 B200 units. By the same arithmetic, a count of 21,000 secured equivalents could correspond to fewer physical chips. Published reports do not give the physical-chip breakdown behind the 21,000.

Does Jensen Huang's Exit Change Korea's 260,000-GPU Deal?

Seoul's reassurance is the main public confirmation. Nvidia said on Nov. 12 that Huang resigned as president, CEO and board member, effective immediately, that CFO Colette Kress is interim CEO while remaining CFO, and that Harvey Jones becomes the board's first-ever chair. The filing gave no reason; Jones said the board expects "operations, customer commitments and product roadmap" to continue without interruption, language that names no country or customer. When Korea's vice minister made a similar assurance on Nov. 9, Nvidia did not immediately respond to Reuters' request for comment.

That earlier assurance came from Vice Minister Ryu Jemyung, who told a parliamentary budget committee that 10,000 GPUs funded in a 2025 supplementary budget had arrived and were in use, and that 2026 batches were under contract and expected to enter service this year. Neither account says whether those 10,000 chips sit inside the 21,000 secured figure.

The pledge itself began as a plan, not a published contract. Nvidia's October 2025 release said Korea's government and industry would add "over a quarter-million" GPUs, and Korean outlets reported a split of 50,000 for the government, 50,000 each for Samsung, SK Group and Hyundai Motor Group, and 60,000 for Naver Cloud. TechRepublic cited local reports valuing the plan at up to $10 billion over several years, while noting that funding details were not disclosed. Digital Daily reported when the deal was announced that pricing, timing and contract structure had not been officially confirmed.

Huang has also taken positions on US export policy. He opposed the GAIN AI Act, which would have required chipmakers to offer US buyers first call on advanced chips before exporting them abroad, and called it more harmful to the United States than the earlier AI Diffusion rule. The White House named him to the President's Council of Advisors on Science and Technology on March 25, 2026.

Seoul Calls It Sovereign AI; Government Adviser Warned of Nvidia Dependence

Bae, founding president of LG AI Research who led development of its Exaone model before his nomination as science minister in June 2025, frames the Nvidia relationship as a base for physical AI, Korean-designed chips and joint research. He listed five areas of cooperation: linking Korean AI chips to Nvidia's NVLink interconnect, building AI computing infrastructure, joint physical-AI research, talent development, and bringing Nvidia's GTC developer conference to Korea.

Lee Seung-hyun, a director at the Digital Platform Government Committee, called the 260,000-GPU package a positive step toward Korea's goal of becoming a top-three AI power, in remarks reported in November 2025. He cautioned that relying on Nvidia's ecosystem for software, simulation and model training could leave Korea unable to design its own "nervous system" over the long term.

The NVLink item has a concrete anchor. On Nov. 9, Nvidia, Rebellions and SK Telecom announced a plan to connect Rebellions' next-generation R200 inference chip to Nvidia systems through NVLink Fusion, backed by the science ministry, with R200 systems expected in 2028. ServeTheHome reported that adopters must still include Nvidia silicon in each node, so the link does not remove the dependence Lee described.

Three Markers Show Whether Korea's Pledge Holds After Nvidia Reports Nov. 17

Nvidia is scheduled to report third-quarter fiscal 2027 results on Tuesday, Nov. 17, at about 4:20 p.m. ET, its first report since Kress became interim CEO. Kress has been CFO since 2013 and now holds both roles. Nvidia's announcement did not name a permanent CEO. Three markers will show whether the Korean pledge holds:

  1. Delivery pace. Whether the remaining 14,000 or so government B200-equivalents arrive by early 2027, and how many of the 21,000 already secured are installed.
  2. Product mix. Whether Vera Rubin priority survives the transition. Newsis reported that Bae said Korea agreed at a June 8 meeting with Huang to receive Vera Rubin with top priority, and Herald Business reported that Bae also asked for an AI factory by year-end built on Vera Rubin NVL72 systems.
  3. Corporate tranches. The Samsung, SK, Hyundai and Naver allocations, which reports of the Nov. 13 update did not mention.

The resignation came amid a board inquiry. Nvidia's board opened a probe on Nov. 8 into Huang's relationship with a former executive at a Singapore-based chip company whom the House chairs' letter described as a publicly reported Chinese intelligence asset; Huang acknowledged the relationship, TechCrunch reported, and Beijing called the allegations groundless. Nvidia's board said it had no reason to believe US national security or company secrets were compromised, the resignation announcement gave no reason, and the espionage claims remain unproven allegations. Three House committee chairs, Brian Mast, John Moolenaar and Rick Crawford, asked Huang in a letter dated Nov. 4 to attend a closed-door briefing on Nov. 18.

For Korean AI teams planning around state compute, the practical test is installed capacity, not secured equivalents: ask which chips are physically in service, in which generation, and under what contract, because the reaffirmation was announced by Seoul after one meeting.

Frequently Asked Questions

What does "secured" mean in Korea's GPU count?

Korean reports of the update use "secured," not "delivered" or "installed," and count in B200-equivalents. The only in-service figure Seoul has given is 10,000 GPUs from a 2025 supplementary budget, stated on Nov. 9.

What is a B200-equivalent?

It is a unit that converts newer Nvidia chips into the compute of one B200, which delivers about 9 petaFLOPS at 8-bit precision. A B300 or Vera Rubin chip counts as more than one unit, so equivalents can exceed the number of physical chips.

Who is running Nvidia now?

Colette Kress is interim CEO and keeps her CFO role, and Harvey Jones is the board's first-ever chair.

Does the NVLink Fusion plan make Korean chips independent of Nvidia?

Not at the interconnect layer. ServeTheHome reported that adopters must still include Nvidia silicon in each node.