Free AI for a nation · Part 2 of 7
What 512 GPUs can deliver, and what they draw
The promise is unlimited use; the supply is a fixed number of GPUs. What they can write is a scenario built from benchmarks of stand-in models, a daily average says nothing about the busiest hour, and electricity is the smaller bill.
The ministry promised a chatbot with no charge and no usage limit.[27] What it supplied this year is 512 GPUs.[26] A GPU writes only so many tokens a second, so “no usage limit” is a rule per person, not a total: the total is bounded, and whoever shows up shares it.
The supply is a scenario
How much a GPU writes depends on the model, and the service’s own models have no public benchmark. So I use two that do, measured on one Nvidia B200: Llama 3.3 70B and gpt-oss-120b, stand-ins for what the service will run.[36, 56]
At the Llama point I use as an anchor, each user gets 53 tokens a second, faster than anyone reads, and the GPU writes 3,238 output tokens a second in all. Multiplied out, that is about 143 billion output tokens a day for 512 GPUs, about 7 chat answers per citizen. That is a benchmark-derived scenario for this model, this precision, prompts and answers of about 1,000 tokens and every GPU busy every second, before any allowance for production. It is not the service’s capacity, and not necessarily a cautious bound: the consortia’s flagships are larger mixture-of-experts models whose published sizes don’t settle serving speed, and nothing has been measured for them. The model moves the number most; gpt-oss-120b writes 4.5 times as much per GPU.
For a sense of scale in paid terms: OpenAI sells ChatGPT Pro at $100, $200 or $500 a month and publishes no token allowance for it, so I assume one.[68] If a Pro subscriber wrote as much every day as the heaviest day this series has measured, one coding-agent session of 189,449 output tokens (chapter 3), the scenario’s output would cover about 760,000 such subscribers, $151 million a month at the $200 tier. Halve the assumed use and the count doubles. That is an equivalence of output volume, not of value: a Pro plan buys OpenAI’s largest models, their reasoning, tools and a service level, and a model of that size would write far fewer tokens per GPU than the 70-billion-parameter stand-in. What Pro subscribers use, nobody has measured.
Chat fits on a daily average; the busiest hour is unknown
At chapter 3’s estimate of 2,000 output tokens per monthly user a day, the operators’ combined end-2026 target of 15 million users would use 21% of the scenario’s output, and 30 million users 42%. Those are scenarios, not a forecast. A service is judged at its busiest hour, though: use peaks in waking hours, part of a fleet is always down or in reserve, each model needs its own GPUs, and long prompts cut output several times over (chapter 3). A daily average below the ceiling shows that chat is feasible on paper, not that the service will be adequate at peak, and no operator has published a service-level target.
The proposed 2027 lease adds about 2,000 GPUs for a year at 170 billion won ($126 million), 85 million won ($63,000) per GPU, 8% more than Lambda’s on-demand price for the same GPU.[11, 34] That is within the market’s range for a hosted B200, and not a rent-versus-buy verdict: a managed lease and a server bought outright are different products, and the lease’s service boundaries aren’t published (appendix).
Electricity is the smaller bill; power delivery is another question
For this fleet the electricity bill is small next to the lease. That is a statement about spending, and only about spending.
The scenario: NVIDIA’s nameplate maximum for a DGX B200 system, times the 1.76 by which Korea’s Board of Audit estimates the country’s data centres multiply their IT power for cooling and losses, every GPU busy all year.[33, 65] This year’s 512 GPUs then use 14.1 GWh a year; the 2,512 the 2027 line would finance use about 69.2 GWh. A nameplate maximum is not a measured average, so a real fleet uses less. At KEPCO’s 2025 average industrial price, this year’s fleet costs 2.57 billion won ($1.91 million) a year, 1.0% of the proposed 2027 line, which pays for a larger fleet; the 2,512 that line would finance cost 12.6 billion won ($9.36 million), 5.0% of it.[62] Cheaper power does make tokens cheaper, by at most the 6% that electricity adds to a GPU’s lease. In household terms: Korean homes used 88.5 TWh in 2025, about 3,800 kWh for each of the country’s 23.25 million households, so this year’s fleet at nameplate maximum would use as much in a year as about 3,700 households, and the 2,512 GPUs as much as about 18,000.[23, 63] That is an equivalence of energy, not households supplied, and a real fleet draws less than its nameplate.
What the scenario doesn’t show is whether the power can be delivered: 1.61 MW of energised racks, cooling and grid connection this year, several times that in 2027. The supplier hosting this year’s cluster isn’t named, no report says where the leased GPUs would run, and I haven’t assessed any of it. Nor is it published whether the lease includes electricity; if the 2027 GPUs come hosted, as this year’s do, the power is inside the lease price and must not be added on top (chapter 4).