Appendix: evidence, arithmetic and sources

How to read the numbers

Each figure says what kind it is when it first appears. Reported figures come from the ministry’s releases or from press reports of the 2027 budget proposal, which the National Assembly has yet to pass; where reports disagree, the text says so. Measured figures are benchmarks or counts someone ran: InferenceX’s and NVIDIA’s runs of stand-in models on one B200, WiseApp’s panel, the token ledger of one coding session. Scenarios are measured figures multiplied out under stated assumptions, such as a fleet at full load all year. Assumptions are mine and are named as such. My view is labelled wherever it appears, and chapter 6 is all of it.

The tender’s rules come from NIPA’s amended call of 29 July 2026, its guide and its FAQ, read in the original. They are my reading, not a legal opinion, and the attachments weren’t re-authenticated for this revision. Money is in won first, then dollars at the European Central Bank’s rate of 15 October 2026: 1,346 won per dollar, a cross rate from the euro.^[1] API list prices were checked on 10 October 2026; GPT-6 Luna’s price, the cheapest of OpenAI’s current GPT-6 models, was re-checked on its own pricing page on 17 October 2026; the older gpt-5-nano is cheaper and isn’t on the site’s list of current models. ChatGPT’s plan prices were read on 18 October 2026 from OpenAI’s Codex pricing page, the one OpenAI page that answered automated fetches; openai.com, chatgpt.com and the help centre refused them on both days, so statements about the free tier’s limits rest on press reports and are labelled as reported. The research files were checked on 17 October 2026; the status of the programme was re-checked on 17 October 2026 against the reports of 14 October, none of which confirms that the closed test began on the 16th as planned.

The 2027 line

The line is called 전 국민 AI 서비스 보편적 활용지원, universal support for an AI service for all citizens. It is new in the 2027 budget proposal published on 1 September 2026.^[11]

Line	Won	Dollars	Per citizen	Share
Lease of about 2,000 B200s for a year	170 billion won	$126 million	3,281 won	68%
Support for the consortia (10 billion won each)	30 billion won	$22.3 million	579 won	12%
Specialised agents: finance, medicine, public services	50 billion won	$37.1 million	965 won	20%
Total	250 billion won	$186 million	4,824 won	100%

This year, the service runs on 512 B200s the government already holds: 256 to the top-ranked consortium and 128 to each of the other two, by the rule; the ministry didn’t publish the ranking.^[9]
After 2028. The minister said on 29 May 2026 that the service will be free to citizens until 2028, and that the plan is to keep it free after that, in talks with the operators. The operators’ support runs until 31 December 2030.^{[24, 28]}
Free means the basic features. Operators may charge for advanced and specialised features; essential ones stay free.^[24]
Population: 51.82 million residents on 1 November 2025, from the register-based census, Korean nationals and foreign residents together.^[23] The survey’s 44.5% of ever-users applied to it gives 23.1 million, in line with the ministry’s own count of about 23 million.^[27]
The free alternatives. ChatGPT’s free tier has had no cap on text messages since August 2026, as reported, with limits still on images, files and voice;^[72] Wrtn is free with rewards tied to ads;^[67] SK Telecom’s A. is free.^[78]

Capacity: the stand-in models

InferenceX, an open benchmark run by SemiAnalysis, measures how much a GPU writes at each speed per user. Its charts count input and output tokens together; for capacity I use output only, about half the total at these prompt sizes. Both points are on one B200 with NVIDIA’s TensorRT-LLM, with prompts and answers of about 1,000 tokens. The daily figures are the scenario, not measurements.^{[36, 56]}

Model	Speed per user	Input and output per GPU	Output per GPU	Output a day, 512 GPUs	Output a day, 2,512 GPUs	Measured
Llama 3.3 70B (dense), FP8	53.1	6,484	3,238	143 billion	703 billion	29 October 2025
gpt-oss-120b (mixture of experts), FP4	58.7	29,436	14,725	651 billion	3.2 trillion	18 May 2026

Stand-ins. The tender’s FAQ sorts models into three types: domestic and meeting the sovereign standard, domestic but not meeting it, and foreign. A Korean model built by further training a public one is the second type, and only the first counts toward the 50%.^[22]
Full load means every GPU busy every second of the year. Maintenance, redundancy, one set of GPUs per model and the busiest hour all take their share, so the scenario is a ceiling on output, not a floor on service.
The Llama data is a year old. InferenceX hasn’t re-run the model since, so later software gains aren’t in it. That doesn’t make the scenario a lower bound for a different model.
Per-GPU figures are the benchmark’s totals divided by the GPUs in the run. Some points use several GPUs per copy of the model (the table below), so “per GPU” is a normalisation, not a one-GPU deployment.
Chat demand is chapter 3’s estimate. WiseApp’s minutes are averaged over every monthly user, days without use included, so no separate ratio of daily to monthly users is needed.
The ministry’s estimate of about 10 million users a year for the leased GPUs leaves each about 56,000 output tokens a day by the formula.^[6]
In ChatGPT Pro terms. OpenAI’s Codex pricing page lists the ChatGPT plans at $0, $8 and $20 and, for Pro, $100, $200 or $500 a month; it publishes no token allowance for chat on any plan, and its own usage figures are for Codex.^[68] Chapter 2’s count is 143 billion output tokens a day ÷ 189,449, the measured session’s output, assumed to be one subscriber’s daily use: about 760,000 subscribers, $151 million a month at $200. The assumption is mine; a Pro subscriber’s real use isn’t published, a frontier model writes fewer tokens per GPU than the stand-in, and the subscription buys quality, reasoning, tools and a service level that the scenario doesn’t price.

Every point InferenceX reported for Llama 3.3 70B in FP8 on B200s, on its run of 29 October 2025. Each is the best of several configurations: how many GPUs share one copy of the model, and how many users each copy serves at once. The seconds count only the writing of a 400-token answer; reading the prompt and any queue come on top.

Speed per user	Input and output per GPU	Output per GPU	Seconds for an answer	GPUs per copy	Users at once
33.4	8,156	4,073	12	1	128
53.1	6,484	3,238	7.5	2	128
74.5	4,543	2,269	5.4	4	128
101.5	3,057	1,528	3.9	4	64
124.8	1,885	944	3.2	4	32
136.2	1,044	519	2.9	4	16
152.8	596	296	2.6	8	16
175.5	338	170	2.3	8	8
183.5	176	88	2.2	8	4

The models the consortia listed with the ministry, sized from their model cards. The sizes say how much memory each needs; they don’t rank the models’ throughput. A.X K2’s weights alone need four B200s at one byte per parameter, at 180 GB each.^{[42, 46]} Kakao lists its Kanana family; the table shows the newest open Kanana model, which may not be the one it serves. KT lists Mi:dm K Pro 2.5 and Solar Pro 4, whose sizes aren’t published.

Company	Model	Parameters	Used for each token
SK Telecom	A.X K2	688 billion	33 billion
LG AI Research	K-EXAONE 2.0	750 billion	37 billion
Upstage	Solar Open 2	250 billion	15 billion
Kakao	Kanana-2 30B-A3B	30 billion	3 billion

What the domestic tokenizers claim for Korean, each in its own maker’s test. The metrics and test texts differ, so the figures can’t be averaged. On English they are no better: A.X K1’s tokenizer needs 792 tokens for an English sample where GPT-4o’s needs 769.

Tokenizer	Compared with	On Korean text
A.X K1 and K2 (SK Telecom)	GPT-4o, Qwen3	527 tokens a sample, against 812 and 906
A.X 4.0 (SK Telecom)	GPT-4o	about 33% fewer tokens for the same input
EXAONE 3.0 (LG)	Llama 3.1	2.46 tokens a word, against 3.01
K-EXAONE (LG)	LG’s older EXAONE tokenizer	about 30% more bytes a token, across domains, not Korean alone
Kanana-2 (Kakao)	its predecessor, on Llama 3’s tokenizer	over 30% more efficient
Solar Open 2 and Solar Pro 4 (Upstage)	global models’ tokenizers	they need 1.2 to 1.9 times the tokens
Motif 3 (Motif)	GPT-4o	0.69 times the tokens, on Korean Wikipedia
Mi:dm 2.0 (KT)	GPT tokenizers	“higher token compression”, no figure

The price per GPU, in won with dollars after:

	Per GPU-hour	Per GPU-year
The state’s 2027 lease, per B200	9,703 won ($7.21)	85 million won ($63,000)
Lambda, on demand (8 B200s)	9,004 won ($6.69)	78.9 million won ($58,600)
Lambda, cluster of 256 or more	11,938 won ($8.87)	105 million won ($77,700)
InferenceX’s own cost assumption	2,337 won ($1.74)	20.5 million won ($15,200)

Rent against buy. An HGX B200 server with eight GPUs costs about $450,000, or $56,250 per GPU, according to Mercatus AI, so a year’s lease is 12% more than the GPU’s share of a server bought outright.^[45] The two aren’t the same product: the lease is probably hosted and managed, as this year’s cluster is, while a purchase adds hosting, power, financing, operations and a residual value, and carries the risk of obsolescence. A fair comparison needs matched service boundaries, and the lease’s aren’t published.
“About”. The state’s price assumes exactly 2,000 GPUs, where the budget says “about”.^[15] The 2027 budget also has a separate 388.5 billion won ($289 million) line for other GPU leases whose term isn’t published.^[7]
Lambda’s prices were re-checked on 17 October 2026: $6.69 per GPU-hour on demand.^[34] InferenceX’s own cost assumption is $1.74.^[35]

Electricity

Every GPU busy all year	512 GPUs, this year	2,512 GPUs, 2027 proposal
Power at the servers, nameplate	915 kW	4,490 kW
With cooling and losses (PUE 1.76)	1.61 MW	7.9 MW
Energy a year, every GPU busy all year	14.1 GWh	69.2 GWh
Share of Korea’s 2025 electricity sales	0.0026%	0.013%
Per citizen a year	272 Wh	1,336 Wh
Korean households with the same use a year (3,800 kWh each)	3,700	18,000
Cost a year at 181.9 won per kWh	2.57 billion won ($1.91 million)	12.6 billion won ($9.36 million)
Share of the proposed 2027 line	1.0%	5.0%

Power per GPU is NVIDIA’s maximum for a DGX B200 system, which holds eight GPUs with their processors, memory, network and fans: 1,788 W per GPU, a nameplate maximum, not a measured average.^[33] NVIDIA’s rating for the B200 chip on its own couldn’t be confirmed, so I don’t use it.
PUE, power usage effectiveness, is a data centre’s total power divided by its IT power. The Board of Audit’s 1.76 is an estimate for Korea’s main data centres; the Uptime Institute’s global average for 2025 is 1.54.^{[65, 66]} Where the state’s GPUs run, and at what PUE, isn’t published.
The price is KEPCO’s average industrial price for 2025, 181.9 won per kWh, all charges included; a data centre’s own contract differs.^[62] The whole 2027 fleet would be 0.013% of Korea’s 2025 electricity sales.^[63]
Per million output tokens at chapter 2’s Llama point: 49 won of electricity against 832 won of lease, and 11 won at the mixture-of-experts point.^[56]
The household average is KEPCO’s 2025 sales on the residential tariff, 88.5 TWh, divided by the 23.25 million households the register-based census counted on 1 November 2025, all households included: about 3,800 kWh a year, 317 kWh a month.^{[23, 63]} That is the method Seoul’s energy portal uses for its own figure, 296 kWh a month for 2024, which the build checks against.^[64] Household-years of electricity are an equivalence of energy with the fleet at nameplate maximum, not households supplied; the residential tariff covers homes, not everything a household uses.
The fleet and the year. An earlier version put this year’s 512 GPUs against the 2027 line alone. The editorial review of 17 October 2026 recomputed both cases, 14.1 and 69.2 GWh, 1.03% and 5.04%, and the build now checks them.

Workloads: the chat estimate and the agent session

The chat estimate, step by step: WiseApp Retail’s smartphone panel counted 30.75 million people using the top ten generative AI apps in September 2026, for 149.59 minutes each that month; use on a computer’s browser isn’t counted.^[69] One prompt a minute is my assumption, and it errs high: typing a question and reading a 400-token answer can take longer. OpenRouter’s study of its API traffic found completions averaging 400 tokens with prompts of over 6,000 tokens; that traffic is developers’ and agents’, so I use the answer length only.^[79]

The agent session, one Claude Code session with Claude Opus 5.5 on 6 October 2026, on Anthropic’s hardware. Its token counts are the comparison, not its cost; thinking is billed and counted as output.^[70]

Tokens	Count
Input sent fresh	184
Input written to the cache	295,177
Input read from the cache	20,196,773
Output, thinking included	189,449
of which thinking	130,104

The three tasks, measured and hypothetical, in chat answers of 400 tokens and chat user-days of about 2,000. The hypothetical rows assume one chat answer of output per call and say nothing about the input each call carries, which grows with every step. They are illustrations, not estimates of the consortia’s agents.

Task	Status	Model calls	Output tokens	Chat answers	Chat user-days
Simple application	hypothetical	5	2,000	5	1
Multi-step booking	hypothetical	20	8,000	20	4
Heavy coding session	measured	92	189,449	474	95

Tasks the size of the session a day, with the 512 GPUs busy every second, from NVIDIA’s TensorRT-LLM tables for Llama 3.3 70B in FP4 on one B200, answers of about 1,000 tokens, no cache reuse: 512 × output tokens a second per GPU × 86,400 ÷ 189,449.^[36] An earlier version divided chapter 2’s FP8 anchor by FP4 ratios, which mixed two benchmarks; the review of 17 October 2026 recomputed the rows below and the build checks them. With a history as long as the session’s, a GPU also holds fewer conversations at once, so the rows are output arithmetic for one configuration, not a capacity.

Prompt	Output tokens a second per GPU	Tasks a day
about 1,000 tokens	6,842	1,600,000
about 8,000 tokens	1,362	320,000
about 33,000 tokens	274	64,000

The traffic rule. The tender’s FAQ, in its answer on abusive users, says the principle is a chatbot “without charges or usage limits”; that speed limits, limits on simultaneous connections and blocking abnormal automated mass calls are allowed; that they must be run so that citizens using the service normally don’t feel them; and that the test for any control is usage behaviour, not purpose.^{[16, 22]} Non-users’ reasons come from KISDI’s panel of smartphone users aged 18 to 72, where each reason is rated on its own, so the shares overlap.^[73]

Operating costs: reference prices

The size of the operating bill per B200 a year, as a share of the state’s lease price of 85 million won ($63,000), in the scenario where the lease excludes hosting. These are reference prices, not the operators’ costs, which aren’t published.

Per B200 a year	Won	Share of the lease
Space: 250,000 won per kW a month × 1.79 kW	5.36 million won	6%
Power: chapter 2’s electricity, every GPU busy all year	5.01 million won	6%
Space and power	10.4 million won	12%
Hosting, at SemiAnalysis’s share of capital plus hosting (H100, 2023)	17.9 million won	21%
Hosting, at SemiAnalysis’s hosting over capital alone	22.6 million won	27%
The state’s operating support, 2027 (proposal)	15 million won	17.6%

The rent is CBRE’s average colocation rent for the Seoul area in 2025, 250,000 won per kW, with no time unit given. Colocation is quoted per kW a month by convention, so I read it as monthly, and I assume the rent is per kW of IT load with electricity billed on top; CBRE states neither.^[90]
SemiAnalysis’s figures are for an HGX H100 server in the US in 2023: hosting $1,871 a month against $7,025 of capital. Hosting over capital plus hosting is 21%; hosting over capital alone is 27%. A lease price is nearer to capital than to the sum, and a B200 draws more power than an H100, so neither ratio transfers cleanly.^[92]
This year’s GPUs are delivered online only, as a managed cluster of 64 nodes run by the GPU supplier, and can’t be moved to an operator’s site. The call values each B200 at 6.6 million won a month, 79.2 million won a year, 93% of the 2027 lease price, hosting included.^[86] Everything else in 2026 is the operators’ cost, and foreign models’ API fees can’t be charged to the project at all.^{[22, 86, 89]}
The 2027 grant is 17.6% of the GPU line, or 15 million won per leased GPU. A ministry official told Digital Daily that the whole 2027 line covers the GPU lease and the costs of developing, managing, maintaining and upgrading the service; the itemisation comes after the final budget.^{[11, 82]}
A consortium’s third assumes the 2027 lease is split evenly, about 670 GPUs each; in practice GPUs may follow usage.
Kakao’s revenue per user: Talk Biz ads and subscriptions were 399.9 billion won ($297 million) in the second quarter of 2026; times four, divided by 49.6 million domestic KakaoTalk users, gives 32,230 won a year. One quarter isn’t a year, and it is a reference for a messenger’s whole business, not attainable AI revenue.^{[71, 85]}
Tokens are priced at the state’s lease, 832 won per million output tokens with the GPUs at full load; at lower utilization each token costs more. The agent user is an assumption: one task the size of this site’s build session every day, priced as if its prompts were short.

Where the models come from

By their model cards and technical reports. The consortia listed SK Telecom: A.X K2, Solar Open 2, Motif 3, K-EXAONE 2.0; Kakao: Kanana, EXAONE, K-EXAONE 2.0; KT: Mi:dm K Pro 2.5, Solar Pro 4, Motif 3.^[18]

Model	Developer	Weights	Base model
A.X K2	SK Telecom	From scratch	None; its architecture follows DeepSeek-V3 (China)
Solar Open 2	Upstage	Grown from an earlier model	Solar Open 100B, Upstage (Korea)
Motif 3	Motif Technologies	From scratch	None
K-EXAONE 2.0	LG AI Research	Enlarged from an earlier model	K-EXAONE, LG (Korea)
Kanana-2	Kakao	“Self-developed”; from scratch not stated	Not published; it uses DeepSeek-V3’s model code (China)
EXAONE 4.5	LG AI Research	Built on an earlier model	EXAONE 4.0, LG (Korea)
Mi:dm K 2.5 Pro	KT	Enlarged from an existing model	Not named
Solar Pro 4	Upstage	Not published	Not published
A.X 4.0 (2025, not listed)	SK Telecom	Further trained on Korean data	Qwen2.5-72B, Alibaba (China)

Which versions. Kakao names only “Kanana” and “EXAONE”, so the table shows the newest of each. Kakao’s earlier Kanana 1.5 was trained from scratch on Llama 3’s architecture and tokenizer, without its weights.^[57] KT lists Mi:dm K Pro 2.5, which its technical report calls Mi:dm K 2.5 Pro.
The standard, as the ministry restated it in January 2026: design and pre-training by the team, not a derivative fine-tuned from a foreign model, and no licensing issue with another company’s model.^[29] In August 2026 the second round kept Upstage, SK Telecom and LG and cut Motif, with all four models judged independent; the final two are to be picked around February 2027, by one press report.^{[5, 19, 20]} Motif 3 still counts toward AI for All’s 50%, because the tender’s test is the standard, not a place in the programme.^[22]
The tender’s model rules, from its FAQ and guide: foreign models only for the functions needed; their API fees can’t be paid from the grant or the companies’ own contribution; the floors leave no 20% foreign allowance; a model outside the sovereign programme must show its pre-training record and disclose any outside weights; the 30% counts by corporate group; shares are counted in output tokens, hidden reasoning included.^{[22, 25]}
Memory: TechInsights’ teardown found SK hynix memory, eight stacks of eight dies each, in one HGX B200. Silicon Analysts estimates it at $3,250 of the $6,750 it costs to build a B200; both are estimates, and Nvidia publishes no list price.^{[49, 59]}
KT’s token factory is KT’s own description, starting with 32 Korean chips, including its claim that for every hundred GPUs it is given, the factory does the work of 121.^{[13, 37]}

The other 2027 lines

For context on where the hardware money goes. The figures are a proposal, as reported.

2027 line	Won	Dollars	Note
Vera Rubin: 140 NVL72 racks at 27.5 billion won	3.85 trillion won	$2.86 billion	lease or purchase: sources conflict
High-quality training data	800 billion won	$594 million	30 billion won in 2026
Other GPU leases (4,000 H100-class, 4,000 B200-class, and more)	388.5 billion won	$289 million	term not published
AI for All: lease of about 2,000 B200s	170 billion won	$126 million	one year

The figures conflict. Money Today puts the 2027 line for AI computing at 4.28 trillion won; Digital Daily and Etoday give 3.85 trillion won for a line of the same name, which is also the Vera Rubin total. Money Today calls the Vera Rubin racks a lease, Digital Daily a purchase.^{[7, 10, 83]}
In context: the ministry’s AI budget for 2027 is 9.42 trillion won ($7 billion); AI for All’s line is 2.7% of it. The Vera Rubin line is about 15 times it, 10,080 GPUs at 382 million won ($284,000) each with the rack’s processors included. The training-data line grows 26.7 times, to 800 billion won ($594 million).^{[10, 15, 83]}

The bids and the scorecard

The six bids, from the ministry’s release of 18 August 2026 and its selection of 28 August. The winners brought 49 partner companies between them; the three losers brought 7.^{[17, 21]}

Lead	Partners	Result
SK Telecom	19	selected
KT	16	selected
Kakao	14	selected
ESTsoft	7	not selected
MKD	none (alone)	not selected
Xenon	none (alone)	not selected

Reach. KakaoTalk’s figure is Kakao’s own count of domestic monthly users in the second quarter of 2026; the apps are WiseApp’s smartphone panel for September 2026.^[69] Kakao’s own targets are 5 million monthly users by December 2026, 10 million in 2027 and 30 million by 2030.^[8]
Ecosystem, at selection. Of the 20 points, 10 went to the spill-over for Korea’s AI industry, including cooperation with Korean AI firms, startups and SMEs, and 10 to keeping the service running after support ends.^[89] The FAQ makes an ecosystem plan mandatory and allows a hub or marketplace for other firms’ services without requiring one. The winners’ plans, in the ministry’s releases: Kakao an agent marketplace linked to its PlayMCP platform, KT its token factory grown into an “open platform”, SK Telecom partnerships with startups; I found none launched for AI for All by mid-October.^{[12, 18]}
The alternatives in chapter 6 use Lambda’s on-demand B200 at $6.69 an hour and dollars at 1,346 won: the whole line would rent about 3,200 GPU-years at that price, or about 2,900 at the state’s own lease price, with nothing left for the grants or the specialised agents.^[11]
“A third”: the ministry’s call says about a third of citizens still don’t use AI. The survey’s own figures are the 33% who had never used an AI service and the 55.5% who had never used generative AI.^{[27, 77]} The survey’s release gave 7.9% paying for generative AI; NIA later re-based it to 17.8% without naming either base, which fits users as the new base: fewer than one in five users pay.^[75]
The state’s other GPUs: Korea Data Economy counts 35,200 B200-equivalent GPUs secured and 11,600 in service, one outlet, not confirmed elsewhere; “in service” is a deployment status, and whether those GPUs are busy isn’t reported.^[4]
The minister’s test, in his words at the kick-off on 4 September 2026: “국민이 서비스를 얼마나 쉽게 접하고 자주 찾는지, 그리고 그 과정에서 실제 삶에 얼마나 도움이 되는지가 무엇보다 중요하다”.^[12] Financial Post Korea, quoting an official, names subscribers, frequency of use and actual GPU use as the reallocation measures, with repeat visits expected to weigh most; that is its paraphrase, not a published formula.^[11]

Chapter 6’s scorecard, and what is measured today:

Measure	Today
Budget in GPU-years, 2027	2,512 (proposal; if the 512 stay in service all year)
Budget per monthly user, 2027	16,667 won ($12.38) at the end-2026 targets (proposal ÷ targets)
Tasks completed, by class, with a quality check	not published; completed tasks is a count the minister named, with no definition
Errors, reversals and rework by citizens or staff	not published
Net time saved, after checking the answer	not published
Inclusion: new users from the 55.5% who had never used generative AI, by age and region	not published; the national survey counts ever-use only
Additionality: benefit over existing public-service channels and the free alternatives	not published
Full cost per successful outcome: GPUs, operations, engineering, safety, support	not published; only the GPU lease and the consortium grants are in the line
GPU status, separately: delivered, commissioned, allocated, available, busy, useful work within a service level	not published; for the state’s other GPUs one outlet reports 33% in service, a deployment status, not utilization
Queue time, time to first token and answer speed at the busiest hour	not published
Model mix, in output tokens	at least 50% sovereign-standard, 30% other domestic (rule)
Electricity	a scenario: 12.6 billion won ($9.36 million) a year for 2,512 GPUs at nameplate maximum (chapter 2); nothing measured
Prompt data used, for what, with whose consent	not published

Sources

Every source the series uses, by subject, newest first. The small numbers in the chapters point here.

How to read the numbers

The 2027 line

Capacity: the stand-in models

Electricity

Workloads: the chat estimate and the agent session

Operating costs: reference prices

Where the models come from

The other 2027 lines

The bids and the scorecard

Sources

The programme, its rules and its budget

GPUs, models and their prices

Electricity

How people use AI

Money and the market