Executive Summary
NVIDIA's founder, Jensen Huang, has left the company. The Korean government met twice the following day and said that the supply of 260,000 advanced GPUs is a company plan rather than a personal one. Instead of passing that reassurance along, this article opens the number 260,000. Inside it are two yardsticks, and one of the two has not been settled.
NVIDIA's announcement wrote 260,000 as a bare count. No sentence anywhere names the unit. When the government describes its own share, by contrast, the tag is unmistakable: these are figures converted to B200 compute. Yet the government has never confirmed it will buy the B200 that serves as the basis, and the 2026 purchases did not include one. What arrived that year was B300 and Vera Rubin, and the 2025 purchase record holds several thousand H200s, which belong to the Hopper generation rather than Blackwell. National Assembly audit material records the same government volume as three numbers at once. Within a single ledger there are three yardsticks.
Everything to this point is what the announcement and the audit material record. What follows is the reading Pebblous brings to it. This article does not argue that the government is lying. Adding up equipment from mixed generations in one line requires conversion, and conversion is not a flaw. The objection is a single one: the converted number and the unconverted number go by the same name. That is why the reaffirmation reaches only the government's share, while the rest rests on plans each company announced rather than on agreements. The resolve to build an infrastructure that one person's departure cannot shake is sound. But measuring whether something is shaking requires one yardstick first. Figures here are as of November 15, 2026, Korea time, and the story is still moving.
13,100
Government GPUs actually running in the field
The same inventory is about 32,000 counted physically and about 35,000 counted in B200-equivalent terms. All three numbers sit side by side in National Assembly audit material
9,704 → 19,000
2026 physical purchase and its B200 equivalent
Restate 7,688 B300s and 2,016 Vera Rubins in B200 compute and the count roughly doubles
80.8%
Share with no published record of delivery
Of the 260,000, some 210,000 belong to private companies. At the audit the Ministry of Science and ICT called that volume a set of plans each firm announced
72.1%
Projects that used under half their allocated GPUs
2,441 of 3,385 GPU-supported projects between 2022 and 2025. Acquisition and utilization are different columns
No Reason in the Filing, Two Meetings in Seoul the Next Day
Jensen Huang has left NVIDIA. The Form 8-K the company filed with the U.S. Securities and Exchange Commission records November 11, 2026 as the day the board accepted the resignation, effective immediately. The filing itself was submitted, and the press release "NVIDIA Announces CEO Transition" published, the following day, November 12. Coverage that dates the resignation to November 12 is working from the announcement date, and this article keeps the two days apart.
The same filing carries three personnel items: Huang's resignation, the appointment of chief financial officer Colette Kress as interim chief executive, and the appointment of Harvey Jones as chair of the board. This is the first time NVIDIA has had a separate board chair, and the 8-K is signed by Kress as interim CEO. The filing gives no reason for the resignation.
With that field left blank, the background can only be traced through coverage, and this article confines it to one paragraph. NVIDIA's board opened an independent review on November 8, 2026, after reports that Huang had been the target of foreign intelligence activity, and announced the resignation on November 12. Ahead of that, multiple outlets reported that Huang had acknowledged an extramarital relationship in an internal message. The U.S. Congress has asked the former chief executive to appear. Since the filing states no reason, this article asserts none, and does not return to the subject.
Markets answered within a day. NVIDIA closed at $166.19 on November 12, down 4.34 percent, and at $162.85 on November 13, down another 2.01 percent. Market capitalization fell below $4 trillion for the first time since July 2025. The share price, though, sits some distance from the argument here. What this article watches is the pair of meetings held in Korea the next morning and afternoon.
Korea's response fell entirely on November 13, 2026, across two gatherings of different character. In the morning the Ministry of Science and ICT convened an emergency review meeting, chaired by Second Vice Minister Ryu Je-myung with NVIDIA Korea in attendance. At 2 p.m. the National AI Strategy Committee held a separate emergency review of its own. Standing Vice Chair Ha Jung-woo and Shin Jin-woo, who chairs the committee's technology innovation and infrastructure subcommittee, were among nine members present, joined by the ministry's AI computing infrastructure team leader and its semiconductor policy director. The two meetings differ in both convener and agenda. The first examined GPU supply; the second looked at widening the range of computing options, domestic AI chips included.
Bae Kyung-hoon, Deputy Prime Minister and Minister of Science and ICT, chaired neither. The minister took questions from reporters after a plenary session of the National Assembly's science and ICT committee that same day, and posted on Facebook under the heading "Jensen Huang has resigned, but the cooperation continues." The minister's remarks came out of those two settings rather than from either meeting's conclusions. Keeping track of who said what, and where, matters again later when this article turns to quotations. The order of the three days reads as follows.
| Date | What happened |
|---|---|
| 2025-10-31 | At the APEC CEO Summit in Gyeongju, Jensen Huang announces a supply of 260,000 GPUs |
| 2025-11-27 | The Ministry of Science and ICT launches a GPU working group with Samsung Electronics, Hyundai Motor, SK Telecom and NAVER — Second Vice Minister Ryu Je-myung attends |
| 2026-11-09 | NVIDIA AI Day in Seoul — a three-way NVLink Fusion agreement between NVIDIA, Rebellions and SK Telecom |
| 2026-11-11 | NVIDIA's board accepts the resignation (per the body of the 8-K). The U.S. Congress asks the former CEO to appear |
| 2026-11-12 | The 8-K is filed and the press release published. Close of $166.19 (−4.34%) |
| 2026-11-13 morning | Ministry of Science and ICT emergency review — chaired by Second Vice Minister Ryu Je-myung, NVIDIA Korea in attendance |
| 2026-11-13, 2 p.m. | National AI Strategy Committee emergency review — on widening the use of domestic AI chips |
| 2026-11-14 | TechCrunch publishes a roundup of Asian government reactions |
The word that catches the eye in that table is "emergency." No new body was created. Ryu Je-myung, who chaired the morning meeting, is the same vice minister who had launched the GPU working group with Samsung Electronics, Hyundai Motor, SK Telecom and NAVER a year earlier, on November 27, 2025. The afternoon agenda of widening domestic AI chip adoption was not new that day either. It had already been made public on November 9, three days before the resignation was announced, and section 5 returns to what that announcement was. A good deal of what looks like a reaction to the event is an existing apparatus doing what it was already doing.
What the 260,000 Figure Actually Counts
The number was born in Gyeongju in October 2025. NVIDIA previewed a "Korea infrastructure and technology advancement AI initiative" in an online briefing on October 30, and Jensen Huang announced it in person at the APEC CEO Summit the following day. President Lee Jae-myung received Huang at the Gyeongju convention center that same day, with Samsung Electronics chairman Lee Jae-yong, SK chairman Chey Tae-won, Hyundai Motor Group chairman Chung Euisun and NAVER chairman Lee Hae-jin in attendance. That a head of state and the heads of four conglomerates were in the room for this announcement matters again later.
The announcement sets out five lines. Up to 50,000 each for the government, Samsung Electronics, the SK Group and Hyundai Motor Group, and 60,000 for NAVER Cloud. The five add to exactly 260,000. Two further points need nailing down here for the later sections to hold. First, the four leading lines all carry the words "up to." They are ceilings, not confirmed quantities. Second, the supply period runs to 2030 and the hardware is the Blackwell family, covering the GB200 Grace Blackwell for servers and the RTX PRO 6000 series for workstations.
| Recipient | Announced volume | Announced use |
|---|---|---|
| Government (Ministry of Science and ICT) | Up to 50,000 | Supplied through NHN Cloud, Kakao and NAVER Cloud |
| Samsung Electronics | Up to 50,000 | Semiconductor AI factory |
| SK Group | Up to 50,000 | Industrial AI services — SK Telecom's sovereign AI infrastructure on RTX PRO 6000 |
| Hyundai Motor Group | Up to 50,000 | Blackwell AI factory, physical AI agreement with the ministry and NVIDIA |
| NAVER Cloud | 60,000 | Written as a separate line |
| Total | 260,000 | Supplied through 2030 |
No price appears in the announcement. A figure of KRW 10 trillion to 14 trillion circulates, roughly $7 billion to $10 billion, but it is an estimate the press worked backward from an assumed $30,000 to $40,000 per GB200 unit, and neither NVIDIA nor the government has ever put a value on the full 260,000. The only item in this announcement carrying an official price tag is Hyundai Motor Group's share of a physical AI joint investment, about $3 billion. One piece of government context belongs here as well. The earlier target was 200,000 units secured by 2030, and the deadline on the government's own 50,000 was pulled forward from 2030 to 2028.
That is what the announcement says. Now for what it does not say. Open the English press release on NVIDIA's investor page, the Korea infrastructure announcement in the newsroom, and the domestic and international coverage of October 31, 2025, and no sentence names the unit behind 260,000. Everything is written as a plain count: "over 260,000 NVIDIA GPUs," or 260,000 units and up. The words conversion and equivalent never attach to it. When the government speaks about its own share, the situation reverses. The tag "a figure converted on the basis of B200 compute" sits plainly on 35,000, on 21,000 and on 14,000 alike.
This asymmetry is where the article starts. The accurate sentence is not "260,000 is a physical count" but "only one side declared its yardstick." The diagram below draws that state. The width of the five boxes follows the announced volumes, and the fill inside each box shows what kind of tag, if any, is attached to that number.
The temptation is to color the empty boxes in as physical counts, but no document anywhere colors them that way. Saying instead that "260,000 is a converted number too" is an overreach of the same kind. The right move is to record what was not written.
Pebblous priced a national compute pledge a month ago as well, in a piece asking whether UK AI safety research running on NVIDIA's computers can stay independent. The snag there sat in the same place. The size of a commitment gets announced as a number, and what the number is a number of does not. The next section looks at what happens when that blank space moves inside the government's own books.
Two Numbers Called GPUs, One With Its Yardstick Written Down
Put "260,000 GPUs" and "50,000 GPUs" in the same paragraph and the subtraction writes itself: 210,000 for the private side. Both numbers end in the same word, and that word invites the same picture, of cards stacked in a warehouse. This section leaves the subtraction alone and asks a narrower question. Is that word measuring the same thing in both places?
3.1The Government's 50,000 Is Converted All the Way Down
One split has to come first. Two separate ledgers share the name "the government's 50,000." One is the line from NVIDIA's announcement seen in section 2: up to 50,000 supplied through NHN Cloud, Kakao and NAVER Cloud by 2030, with nothing on that line about a unit either. The other is the line in the government's own plan: 50,000 secured by 2028, with the tag "converted to B200 compute" sitting plainly on it. The deadlines are two years apart to begin with. Whether the two 50,000s are the same 50,000 cannot be established from any public document. Which means the sentence "of the 50,000 pledged, 21,000 have arrived" may be laying progress from one ledger on top of a promise made in the other. What gets taken apart below is the second line, the 50,000 in the government plan.
That 50,000 runs along two branches: 35,000 the government secures directly, and 15,000 for the National AI Computing Center. The first branch divides again. As of November 13, 2026, 21,000 had been secured, and the remaining 14,000 arrive in early 2027, the point coverage refers to as "early next year." All four of these numbers, 50,000 and 35,000 and 21,000 and 14,000, are figures converted on the basis of B200 compute. Several outlets carried the same tag.
Four deadlines circulate separately as well. Read together they make the pace look faster or slower than it is. Here are the four.
| What | Deadline |
|---|---|
| NVIDIA's supply of 260,000 | Through 2030 |
| Government secures 50,000 | By 2028 — pulled forward from 2030 |
| Remaining 14,000 of the directly secured share | Early 2027 |
| National AI Computing Center | Operating in 2028, opening with 15,000 and expanding to 50,000 by 2030 |
3.2One Ledger Records the Same Inventory as Three Numbers
So far this is one tag on one line. Then the inside of that ledger came open at the National Assembly's annual audit in October 2026. Choi Hyung-doo, a member of the Assembly's science and ICT committee, obtained material from the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA). The same government inventory had been tallied three ways at once. Counted as physical equipment it comes to about 32,000; converted to B200 compute, about 35,000; counted as what is actually running in the field, 13,100.
What makes this material valuable is that not one step of inference enters it. This is not an outsider setting two numbers against each other and calling them inconsistent. One ministry, describing its own material, produced three numbers together. None of the three is wrong. Each is the correct answer to a different question. The trouble is that all three are called GPUs. Which number goes into a report moves the same program's score by as much as a factor of 2.7.
These three are numbers from inside the government's 50,000. They should not be read out across the full 260,000. For the private share no such tally has been published at all.
3.3The B200-Equivalent Formula Is Recoverable From One Purchase
How is a B200 equivalent calculated, then? No document stating the official formula could be found. The government did once describe a year's purchase with the physical units and the converted value side by side. What it bought that year was 7,688 B300s and 2,016 Vera Rubins, 9,704 units in all, and the explanation was that this "corresponds to roughly 19,000 B200s" and delivers about 30 percent more performance than the original target of 15,000 B200s. One line holding both the physical count and the converted value is a line that can be worked backward.
Before working it backward, one number has to be set aside. The "original target of 15,000 B200s" in that explanation is not the 15,000 for the National AI Computing Center seen in section 3.1. The first is the acquisition target a single year's program had set; the second is the opening inventory of a center that opens its doors in 2028. Two ledgers carry a number of the same size, and on both it is called 15,000 B200s. If the two are read over each other, a one-year target slides into the center's allocation. The problem this article is examining turns up here a second time, inside the source material.
Take the vendor-published FP4 dense compute figures as 9 petaflops for the B200, 13.5 for the B300 and 35 for Vera Rubin, and the arithmetic runs (7,688 × 13.5 + 2,016 × 35) ÷ 9 ≈ 19,372 units. That is within 2 percent of the "about 19,000" the government stated. Treating this as roughly the right formula is reasonable enough. The calculation is ours, though, and no official confirmation backs it.
Running the arithmetic also turns up something less comfortable. The performance figures fed into it wobble by source. Public material puts the B300 anywhere from 11 to 15 petaflops and Vera Rubin from 35 to 50. Set the B300 at 15 instead of 13.5 and the same calculation yields 20,653. Same equipment, same method, one line of a spec sheet changed, and the result moves by nearly 1,300 units. A converted figure is an assumed value rather than a measured one, and where the assumption is written down has not been made public.
3.4The Reference Chip Is Not What the 2026 Program Bought
One step further in. If every number is converted to B200 terms, the B200 should at least be a fixed basis. Here is how the Ministry of Science and ICT described the center's 15,000 at the August 2026 groundbreaking briefing.
"It has not been confirmed that the B200 will be adopted as such; the governing principle is to adopt whichever advanced GPUs and AI chips are most suitable at that point."
Translated from the Korean original.
The chip used as the measuring stick has never been confirmed as a purchase. What arrived under the KRW 2.08 trillion main-budget program for 2026, about $1.4 billion, was B300 and Vera Rubin, and not one B200. The 13,000 units that arrived under the KRW 1.46 trillion supplementary-budget program for 2025, about $1.0 billion, include 10,080 B200s alongside 3,056 H200s. The H200 is not Blackwell but the preceding Hopper generation. Vera Rubin belongs to Rubin, the generation after.
This is the firmest point in the article, because it rests on model names rather than inference. The 260,000 NVIDIA announced in October 2025 was Blackwell. Samsung's share is written as "Blackwell GPUs," Hyundai Motor Group's as "50,000 NVIDIA Blackwell GPUs," and the SK and NAVER shares as "RTX PRO 6000 Blackwell." Several thousand Hopper-generation units therefore fall outside what that pledge can hold. The government's 50,000 ledger is not a subset of the 260,000 announcement but a separate list that overlaps it somewhere, and where the overlap lies has not been published.
The model breakdowns for the two programs are program-wide figures. Strip out the cloud providers' own volumes and the government's usable share is smaller, by one set of estimates. Whichever estimate is taken, the presence of Hopper-generation equipment does not change.
3.5Korea Is Not the Only Place a Yardstick Shifts
Anyone who reads to this point and files it under "Korean officials fudging numbers" has missed half the article. NVIDIA did the same thing in its own accounts. From the third quarter of fiscal 2026 the company changed the basis for its regional revenue breakdown from customer billing location to customer headquarters location. The "Singapore" line, which had carried around 20 percent of total revenue in the first half of fiscal 2026, disappeared from the table. The revenue did not disappear; which column it belongs in did. Since the basis changed, that line is no longer disclosed.
Estimates of what the same 260,000 would draw in power run from 260 megawatts to 1 gigawatt. No official figure exists, and every published number is worked backward from a different assumption about kilowatts per rack. Rebuttals are on the record too. Kim Sun-kyo, a research fellow at the Korea Institute of S&T Evaluation and Planning, called the estimates excessive, and Lee Heon-seok, a policy director at Energy Justice Actions, pointed out the error of folding overseas data center demand into the same calculation. When one number is left unsettled, every calculation built on it scatters.
Two earlier pieces on the Pebblous blog took this problem up in advance. One asked whether computing power can be bought and sold the way electricity is, and the answer turned on defining a deliverable grade before compute can be traded at all. The other looked at ING's auction of data center loans, where the snag was that the useful life attached to one and the same chip ran from three years to eight depending on the document. Numbers that share a name and not a yardstick show up in every industry.
One point needs nailing down. Conversion is not a flaw. There is no other way to add equipment from different generations into a single line, which is why conversion is routine in public administration. The objection in this section is not that a conversion was made but that the converted number and the unconverted number are both called GPUs. And when two such numbers appear in one paragraph, each of them needs a name. That is also why this article never sets 9,704 against 21,000. The first is one year of physical purchasing and the second is a cumulative converted holding, so any sentence pitting them against each other is wrong on its face.
The Assurance Reaches the 50,000 and Stops There
What exactly did the November 13 morning meeting reaffirm? The question has to come first, because an assurance has a range. Here is the explanation from a Ministry of Science and ICT official, as carried in the coverage.
"Having put questions to NVIDIA Korea about GPU supply, we confirmed that matters relating to cooperation with Korea, including the provision of 260,000 GPUs, have proceeded at the level of the company rather than of an individual, and are expected to proceed without disruption."
Translated from the Korean original.
Who is speaking matters here. The speaker is the Ministry of Science and ICT, and the basis given is having put questions to NVIDIA Korea. That is a different thing from an official statement issued by NVIDIA's head office toward Korea. Render this as "NVIDIA said" and the source moves up a rung. Another explanation from the same meeting has the same subject: that the government's own GPU procurement is likewise proceeding as planned through the cloud providers. The line Deputy Prime Minister Bae Kyung-hoon posted on Facebook the same day, that NVIDIA had reaffirmed its plan to proceed at the company level as previously announced, travels the same route.
A public sentence from the NVIDIA side exists separately. In the press release issued with the 8-K, board chair Harvey Jones said the board "has full confidence in Colette and the leadership team," and that the company expects "operations, customer commitments and product roadmap to continue without interruption." Same thrust as what the Korean government relayed, and with one further thing in common. Neither side discloses the terms of any contract. What the reaffirmation carries, then, reaches as far as "the pledge stands." What arrives and when is not inside it.
4.1Only 19 Percent of the 260,000 Is Within Reach of a Check
What the government can verify on its own ends at its 50,000. The other 210,000 belongs to private companies, and neither delivery schedules nor progress has been published. This is not an outside assessment but something established at the National Assembly audit. Choi Hyung-doo noted that the 260,000 is a number the government has used as evidence of national AI competitiveness, while 81 percent of it consists of corporate plans whose timing the government cannot even confirm. The ministry answered that the volume in question had not been the subject of an agreement or a written commitment with the government, and that each operator had announced its own plan.
Choosing the accurate sentence matters at this point. Writing that the private volume is behind schedule would be inventing a fact that is not in evidence. There is no material showing a delay and none showing the absence of one. The sentence that holds is that no public route exists for checking. And that share comes to 210,000, 80.8 percent of the whole, the number rounded to 81 percent at the audit.
Progress inside the government's share is public, and here too the denominator has to be read alongside the figure. Audit material records deliveries of 2,608 units as of October 8, 2026, a share of 26.9 percent. The denominator there is the full 9,704 units of the 2026 program. The "roughly 40 percent" the Ministry of Science and ICT described a month later, on November 13, rests on a different denominator: the 7,688 B300s alone. Set the two side by side and read a 13-point gain in one month and the reading is wrong. The 2,016 Vera Rubins were scheduled to begin arriving in December, so at that point they sat outside the denominator entirely.
| As of | Delivered share | What the denominator is |
|---|---|---|
| 2026-10-08 | 26.9% (2,608 units) | The full 2026 program, 9,704 units |
| 2026-11-13 | About 40% | Only the 7,688 B300s within it |
4.2NVIDIA's Own Filings Carry No Line Named Korea
Why not open the seller's own accounts and check instead? No such column exists. The regional revenue breakdown NVIDIA files with the U.S. Securities and Exchange Commission runs to four entries, the United States, Taiwan, China and other, and Korea is buried whole in the last of them. No method exists for reading out of the filings how much of the 260,000 has actually shipped. And as section 3.5 showed, the composition of that table itself changed once, in the third quarter of fiscal 2026. The buyer's ledger carries a converted number, and the seller's ledger carries no line for the country.
Pebblous has examined the gap between a promised number and a filed number several times. One case was that the total Meta promised to spend does not appear in its filing. Another was that Firmus has two data centers running and five planned, 46 megawatts against 865. What is missing here sits on the seller's side of the transaction rather than the buyer's.
One thing could not be resolved from public documents. NVIDIA's announcement records the government's 50,000 as supplied through NHN Cloud, Kakao and NAVER Cloud, and records NAVER Cloud's 60,000 on a separate line. Meanwhile the 9,704 units the government took in under the 2026 program arrived split across three operators: 4,120 to NAVER Cloud, 3,024 to Samsung SDS and 2,560 to Elice, which sums to 9,704. Which of those two announcement lines the 4,120 taken in through NAVER Cloud belongs to is not distinguishable from any public document. The relationship between NAVER and NVIDIA was covered once before, in a piece on the chip seller taking an equity stake in the buyer, and that column was not resolved there either.
One last entry belongs in the same place, about the limits of this article. The original press releases issued by the Ministry of Science and ICT and the National AI Strategy Committee on November 13, and the original Facebook post by Deputy Prime Minister Bae Kyung-hoon, could not be opened. Every quotation above is taken from what the press reproduced, which is why none of them is introduced with "according to the press release." For a piece that insists on marking the boundary of what it verified, this column should not be left blank either.
Three Prescriptions for an Infrastructure That Holds Steady
Before taking the prescriptions apart, one acknowledgment. The angle of this article was raised first by someone else. At the afternoon meeting on November 13, Shin Jin-woo, the professor who chairs the National AI Strategy Committee's technology innovation and infrastructure subcommittee, said nearly the same thing first.
"What has happened confirms that national AI infrastructure needs a structure that does not wobble even with a particular company's management variables."
"Cooperation with NVIDIA should continue unchanged, but we have to raise the utilization rate and the operating capability of the GPUs we have secured, and widen the range of computing resources to include domestic AI chips."
Translated from the Korean original.
The diagnosis came from that side first, and so did the prescriptions: wider adoption of domestic AI chips, pairing with NVIDIA technology, and better utilization and operating capability. All three are reasonable. The job of this article is not to call them wrong but to check, in numbers, what each of them presupposes. All three presuppose measuring, and the yardstick is not yet a single one.
5.1Domestic NPUs Are Growing, and the Order of Magnitude Is Different
Start with the domestic AI chip side. The 2025 results of the Korean fabless firms building NPUs, or neural processing units, are known from coverage based on regulatory filings. Rebellions booked KRW 32.02 billion, about $22 million, and FuriosaAI KRW 5.74 billion, about $4 million. Six companies together come to KRW 50.4 billion, about $35 million. The growth itself is steep, ranging from 93 percent to 209 percent year on year.
A comparison placed beside it makes the order of magnitude visible. The government spent KRW 2.08 trillion, about $1.4 billion, on GPU acquisition in 2026 alone. The combined annual revenue of all six Korean fabless firms amounts to roughly 2.4 percent of that single year's budget. Why "switch to domestic" is not an easy prescription is contained in that ratio. Before asking whether substitution is possible, it is worth writing down what is possible at the current scale.
| Item | 2025 revenue |
|---|---|
| Rebellions | KRW 32.02bn ($22M) |
| FuriosaAI | KRW 5.74bn ($4.0M) |
| Six Korean AI-chip fabless firms combined | KRW 50.4bn ($35M) |
| Government GPU acquisition budget, 2026 | KRW 2.08tn ($1.4B) |
| Combined revenue as a share of that budget | About 2.4% |
How domestic chips get into public infrastructure is also a problem that has already been designed once. In July 2026 Pebblous covered the National AI Computing Center dropping its domestic NPU mandate and breaking ground anyway. Setting aside a dedicated zone inside the center for designing and prototyping domestic AI chips is part of that same design.
5.2Pairing Moves the Layer of Dependence Rather Than Removing It
The second prescription, pairing with NVIDIA technology, already has a shape. At NVIDIA AI Day in Seoul on November 9, 2026, three days before the resignation was announced, NVIDIA, Rebellions and SK Telecom signed a three-way NVLink Fusion agreement. The method is disaggregated inference. Inference is split into two stages, with NVIDIA GPUs taking the prefill at the front and Rebellions' R200 NPU taking the decoding at the back. The underlying rack architecture is NVIDIA MGX, and supply begins in 2028.
Separate out what becomes independent and what stays put and it reads like this. One stage of inference moves onto a domestic chip. The interconnect joining chip to chip, and the rack architecture housing them, remain NVIDIA's. This pairing removes nothing; it relocates the layer at which the dependence sits. It is also some distance from an industry consensus. Five days earlier, on November 4, FuriosaAI signed its own agreement with MangoBoost at an AMD event in Seoul, choosing AMD's ROCm stack and an Ethernet interconnect. The domestic NPU camp is not walking a single road.
5.3The Utilization Prescription Already Has Measured Numbers Attached
The third prescription reads as an abstraction but in fact points at specific figures. Material that Han Jun-ho of the Assembly's science and ICT committee obtained from NIPA was reported in October 2026. Of 3,385 GPU-supported projects under the high-performance computing support program between 2022 and 2025, 2,441 used less than half the resources allocated to them, or 72.1 percent. The sharper number comes next. GPUs the government secured directly and distributed to 215 research and development projects, 10,104 units in all, ran at 51.8 percent utilization in September 2026. That figure attaches not to support programs in general but to the government's own holdings.
The direction is improving. The same material puts monthly average utilization in 2026 at 26.8 percent in March and 51.5 percent in September, and the share of projects running under 20 percent utilization fell from 52.3 percent to 9.5 percent. The ministry's rebuttal belongs in the same place: research projects idle for stretches by their nature, and this is not idle capacity. What gets named as the structural cause is people rather than equipment, a shortage of specialists who can handle distributed training design and job scheduling.
Seats filled and no work being done is observed on clusters in other countries too. Pebblous once covered a Princeton University cluster measured at 0.1-second resolution over ten days, where roughly half the jobs were using under 5 percent of the GPU compute cores on average. On paper that time also counts as running. Raising the count and using the count are separate tasks, which is where this meets section 3. If acquisition and utilization are different columns, the two should not go by one name.
A signal pointing the other way deserves an entry as well. Hangang, the sixth national supercomputer at the Korea Institute of Science and Technology Information, carries 8,496 GPUs, and demand from 626 applying projects came to 2.4 times the supply. Average utilization on Nurion, the fifth machine, has run at 76 to 80 percent in recent years. These resources are research supercomputing infrastructure, a different pool from the government AI infrastructure described above. Put the two sets of numbers in one table and "there is too much" and "there is not enough" become true at once, which makes both unreadable.
Japan Gave One Line, Taiwan Gave One Line, Korea Held Two Meetings
Reactions across Asia came in different sizes. In a roundup published by TechCrunch on November 14, 2026, Japan's Ministry of Economy, Trade and Industry called NVIDIA "an important player in Japan's AI semiconductor industry" and said it would "closely monitor" the company. Taiwan's economy minister said the departure was "unlikely to significantly change the growth of the global AI industry," and the Taipei city government said NVIDIA's plan for an overseas headquarters was proceeding on schedule with its contractual obligations being met. China's foreign ministry repeated its standing position against "export controls that disrupt global supply chains." One line per ministry. Korea alone held two meetings on the same day.
6.1The Quiet Reaction Has 27,500 Units Behind It
It is easy to jump to a conclusion here: that Japan and Taiwan were quiet because they had less riding on it. The material says the opposite. Japan has committed ¥387.3 billion to its sovereign AI program Noetra in 2026, about $2.5 billion, and ¥1 trillion over five years, about $6.5 billion, with 44 companies taking part. The GPUs being procured are roughly 27,500 Vera Rubins. Nothing about that is a small stake. The form the reaction took was simply different.
6.2Four Countries' Numbers Do Not Belong in the Same Table
Line the national figures up and Korea looks overwhelming. Against Japan's 27,500, the gap is close to tenfold. And this is exactly where the section meets section 3. The four numbers differ in unit, in generation and in who made the promise. Korea's 260,000 is a count of Blackwell-family units with no unit declared; Japan's 27,500 is a count of Rubin units, one generation later; Taiwan's number is not a count of units at all but an investment figure; and the UK's 120,000 is not a volume a government is receiving but a deployment plan by private operators. The table below pulls those four columns apart.
| Country | Announced figure | Unit of that figure | Generation | Who promised |
|---|---|---|---|---|
| Korea | 260,000 | Not declared | Blackwell family | NVIDIA → government and four groups |
| Japan | About 27,500 | Units | Vera Rubin (Rubin) | Government sovereign AI procurement |
| Taiwan | Up to $150bn a year | Investment | Not applicable | Supply chain investment |
| United Kingdom | 120,000 | Units | Blackwell Ultra | Nscale, CoreWeave, Microsoft and other private operators |
Once the table is pulled apart this way, sentences of the "ranked Nth in the world" variety have less ground to stand on. Pebblous covered the UK case separately, and the snag sat in the same spot: the compute NVIDIA pledged to UK AI safety research was part of a private deployment plan rather than a volume the government was receiving, and the portion actually turning for public research had to be counted on its own. Ranking countries by quantity is ranking them on one of four columns.
6.3Why Korea Alone Met Twice
That leaves the size of the reaction. Two reasons appear. One is the scale of the pledge. The 260,000 is the largest of the four by quantity, and the government has cited it repeatedly as evidence of national AI competitiveness. The other is the form it took. This deal did not arrive quietly in a corporate press release. Jensen Huang announced it in person, in front of the president and the heads of four conglomerates. The evening before, there had been a gathering at a fried-chicken restaurant in Seoul's Samseong-dong with chairmen Lee Jae-yong and Chung Euisun, with chairman Chey Tae-won absent and met separately in Gyeongju, and on June 7, 2026 a second gathering followed at the same shop. It was the seventh known meeting in the seven months since late October 2025.
Counting the meetings is not a matter of gossip. It is the evidence that explains why an assurance phrased as "the company, not the individual" was needed at all. And the person who gave that assurance acknowledged the individual's part in the same breath. Here is Deputy Prime Minister Bae Kyung-hoon, speaking to reporters on November 13 after the Assembly committee session.
"Jensen Huang has shown a great deal of friendship and personal engagement toward Korea over this period, but I expect no problem in the cooperation between the government and the companies."
Translated from the Korean original.
There is no intention of being arch about this. Setting the two statements side by side is enough. Cooperation is at the company level, and there has been personal friendship, and both sit inside one sentence. Both are probably true. That they sit together, however, also means that nobody has yet written down which parts of this plan were institutions and which were relationships.
Why This Matters to Pebblous
This is not a semiconductor story. It is a story about units. There is one question Pebblous asks first whenever a client's data is opened: what is the unit of this column, and when did it change? Finding mixed units and shifted definitions inside a dataset is what DataClinic does, and section 3 of this article is a record of putting that question to a single line of a national AI infrastructure plan. Total a column that holds two units and the total stops working as a number. GPUs or customer transactions, the situation is the same.
7.1Four Data-Quality Failures, All Present Here
The four defects that turn up in every metrics review are all present in this one event. First, one name covers two units. The 260,000 and the 35,000 are both GPUs. Because the name is the same, nobody asks, and because nobody asks, a sentence subtracting one from the other gets written without friction. Second, the basis itself is provisional. The government says the B200 that anchors the conversion has never been confirmed, what it actually buys is B300 and Vera Rubin, and the earlier program holds Hopper-generation equipment. Change the reference generation and every past figure has to be counted again.
Third, one inventory is recorded as three numbers. Which of 32,000 physical, 35,000 converted and 13,100 in use gets reported becomes the conclusion. That all three are true is what makes this defect so hard to catch. Fourth, a verifiable subset is used to assert the whole. Only 19 percent of the 260,000 can be checked at all, yet the announcement and the coverage both speak of the full 260,000 as one block. Of the four, this is the most common.
7.2Four Questions to Put to Your Own Organization
A line reading "we have secured N of AI resources" now appears in almost every company's reporting. Four things are worth asking, by whoever writes the line and whoever receives it. None of the four is specific to GPUs. They apply unchanged to headcount, to license quantities and to training-data record counts.
- What was counted — physical quantity, a performance conversion, or a monetary conversion?
- What is the conversion basis, and is it fixed? In the reverse calculation here, one line of a reference spec sheet moved the answer by nearly 1,300 units.
- How far into the total does our own ledger reach? Are we counting someone else's holdings as ours?
- Are acquisition and utilization separate columns? Are arrivals and uses being written into one? The 72.1 percent is what splitting that column produced.
Earlier Pebblous pieces have pressed on what a company promises against what its books record. The total Meta promised to spend is absent from its filing; what Firmus runs and what Firmus plans to build differ by close to twentyfold; the useful life of ING's collateral has no agreed figure. This article carries that line of work into a Korean national plan still in motion. One thing differs here: this ledger is not closed. Another 14,000 arrive in early 2027, and the center starts turning in 2028. Which yardstick records those numbers is still something that can be decided.
The figures and quotations in sections 1 through 6 are confirmed from NVIDIA filings and press releases, material submitted to the National Assembly audit, and the coverage that carried both. The original press releases from the Ministry of Science and ICT and the National AI Strategy Committee could not be opened, so press reproductions were used, and section 4.2 says so. Treating the three numbers as one problem of yardsticks, and section 7, are readings Pebblous has added. To state it once more: this article does not claim the government falsified a number. Conversion is a necessary procedure, and the objection is narrow: one word, GPUs, is doing duty for a measured count and a calculated one. Figures are as of November 15, 2026, Korea time, and parts of this will need rereading after NVIDIA's quarterly results on November 17, 2026.
References
Sources are grouped by kind. Items 1 through 5 are documents NVIDIA issued directly, and items 6 and 7 are the National Assembly submissions this article is built on. From item 8 onward the sources are press coverage, and where an original could not be opened and only the reporting was available, the body says so.
NVIDIA filings and announcements
- 1.NVIDIA Corporation, Form 8-K (resignation accepted November 11, 2026; filed November 12, 2026). sec.gov — the two dates in section 1, the interim CEO and board chair appointments, and the absence of a stated reason all rest on this.
- 2.NVIDIA Corporation, Ex-99.1, "NVIDIA Announces CEO Transition." sec.gov — the quotation from board chair Harvey Jones in section 4 is here.
- 3.NVIDIA Investor Relations, "NVIDIA Announces CEO Transition" (November 12, 2026). investor.nvidia.com
- 4.NVIDIA Newsroom, "NVIDIA, South Korea Government and Industrial Giants Build AI Infrastructure" (October 31, 2025). nvidianews.nvidia.com — the five-line allocation in section 2, the "up to" wording, the hardware, and the period running to 2030 come from here. The absence of any sentence naming the unit behind 260,000 was confirmed against this document.
- 5.NVIDIA Corporation, Form 10-Q (fiscal 2026 first and second quarters). The basis for the regional revenue breakdown, the Singapore billing-location share, and the change of basis to customer headquarters location from the third quarter. Used in sections 3.5 and 4.2.
National Assembly audit submissions
- 6.Office of Choi Hyung-doo, National Assembly Science, ICT, Broadcasting and Communications Committee; material submitted by the Ministry of Science and ICT and NIPA (reported October 14, 2026; cross-checked in ZDNet Korea and SR Times) — the three tallies in section 3.2 (about 32,000 physical, about 35,000 B200-equivalent, 13,100 in use), the 2,608 units and 26.9 percent in section 4.1, and the ministry's answer on the private 81 percent all come from here.
- 7.Office of Han Jun-ho, National Assembly Science, ICT, Broadcasting and Communications Committee; material submitted by NIPA (reported October 15, 2026; ZDNet Korea, IT Daily, Asia Economy) — the 72.1 percent in section 5.3, the shift in monthly average utilization from 26.8 to 51.5 percent, and the 51.8 percent September utilization of the 10,104 directly secured units are here.
Press coverage
- 8.Electronic Times (ETNews), coverage of the Ministry of Science and ICT emergency review following Jensen Huang's resignation (November 13, 2026). etnews.com — the ministry official's explanation quoted in section 4 is here.
- 9.Money Today, related coverage the same day (November 13, 2026). mt.co.kr — used to cross-check the times and attendees of the two meetings in section 1 and the remarks by subcommittee chair Shin Jin-woo in section 5.
- 10.TechCrunch, "Asian governments react to Nvidia CEO Jensen Huang's exit" (November 14, 2026). techcrunch.com — the Japanese, Taiwanese and Chinese reactions in section 6 all pass through this piece.
- 11.Financial News (November 13, 2026) — the remarks by Deputy Prime Minister Bae Kyung-hoon in section 6.3, and the "converted to B200 compute" wording attached to the government figures, were confirmed here.
- 12.Munhwa Ilbo, Jodal Economy and Supple (2026) — the program details in sections 3.3 and 3.4. These cover the model composition of the 13,000 units under the 2025 supplementary budget program (10,080 B200 and 3,056 H200), the composition of the 9,704 units under the 2026 main budget program (7,688 B300 and 2,016 Vera Rubin) and their allocation by operator, and the ministry's statement that the B200 is not confirmed for the National AI Computing Center.
- 13.ZDNet Korea and Numbers (2026) — the Rebellions figure of KRW 32.02 billion, the FuriosaAI figure of KRW 5.74 billion, and the KRW 50.4 billion combined total for six Korean fabless firms in section 5.1 were confirmed in this coverage, which is based on filings with the Financial Supervisory Service.
- 14.Kyunghyang Shinmun, Munhwa News and Etoday (2026) — the range of power estimates in section 3.5 and the rebuttals from Kim Sun-kyo of the Korea Institute of S&T Evaluation and Planning and Lee Heon-seok of Energy Justice Actions. All of these are estimates; no official figure exists.
- 15.Kuki News and HelloDD (2026) — the basis for the 8,496 GPUs on the KISTI sixth supercomputer Hangang, the 2.4-times oversubscription, and the 76 to 80 percent average utilization on the fifth machine, Nurion, in section 5.3.
- 16.Nikkei-affiliated coverage (2026) — Japan's sovereign AI program Noetra in section 6.1, its procurement of roughly 27,500 Vera Rubins, and the scale of ¥387.3 billion in 2026 and ¥1 trillion over five years come from here.
- 17.Coverage of the NVIDIA, Rebellions and SK Telecom NVLink Fusion agreement (November 9, 2026) and of the FuriosaAI and MangoBoost agreement (November 4, 2026) — the basis for the division of labor and the supply timing in section 5.2, and for the statement that the two camps have diverged.
Adjacent pieces on the Pebblous blog
- 18.Can UK AI Safety Research on NVIDIA's Computers Stay Independent? — cited in sections 2 and 6.2. The immediately preceding record of putting a price on a national compute announcement.
- 19.Can Computing Power Be Traded Like Electricity? — the generalizing bridge in section 3.5, on defining a deliverable grade before compute can be traded.
- 20.ING Is Auctioning Twelve Data Center Loans, and the Collateral Has No Agreed Lifespan — referenced in one line in section 3.5, as a case where the period attached to one chip varies by document.
- 21.They Dropped the Mandate, and Broke Ground — the anchor for section 5.1, on the design that turned domestic NPU adoption from a requirement into a choice.
- 22.GPU Cluster Fully Booked, Compute Cores Mostly Idle — the sibling record to section 5.3. That piece measures at 0.1-second resolution; this one uses project-level tallies in Korea.
- 23.Meta Promised to Spend the Money — The Total Is Not in the Filing · Firmus Pulls Its IPO With Two Data Centers Running, Five Planned — referenced in one line each in sections 4.2 and 7. Both weigh a pledged figure against what the filings record.
- 24.Naver Builds the AI Factory, and Nvidia Owns a Piece of It — cited in section 4.2. It covers the structure of the recipient of 60,000 of the 260,000.