₩3.85 trillion. One line item in a leaked business plan, all of it earmarked for 140 Vera Rubin NVL72 racks. That works out to ₩27.5 billion apiece.
The document is the "Global AI Frontier Lab" business plan, pulled by Democratic Party lawmaker Lee Jung-heon's office and reported by Money Today. It tells you the Korean science ministry is budgeting to buy the 10,080 Rubin GPUs that Deputy Prime Minister Bae Kyung-hoon said in June were needed to build a Mithos-class frontier model — a number he named precisely, and a number these 140 racks hit exactly.
Now set the per-rack figure next to the sell-side. Bernstein pegged the Vera Rubin NVL72 rack at $9.1 million in a June report — roughly ₩12.3 billion. Morgan Stanley put it near $7.8 million, about ₩10.5 billion.
Korea budgeted more than double the higher estimate. That gap is the story.
Korea budgeted more than double the higher sell-side estimate. That gap between manufacturing cost and price paid is the entire story.
Last week in this space, we traced a different number: Nvidia telling large customers that server prices would rise more than fifteen percent next year, with memory the stated cause. That was the supply side writing the invoice. This week is the receipt — a sovereign buyer that has already accepted the pass-through, penciled it in, and added a buffer on top.
The Korean government didn't hide its reasoning. It cited the possibility of price increases driven by GPU shortages, plus facility costs, as the basis for the budget. In plain terms: it is pre-paying for the memory-driven price hike we described last issue, and then paying more still for the certainty that the racks arrive.
The temptation is to read this as a rich government overpaying. That reading is wrong, and the corroborating buy-side commitments show why.
Look at Anthropic. In eleven months it has, according to a tally reported by The Information, signed compute contracts worth $517 billion — roughly ₩700 trillion — covering 14.8 gigawatts of capacity to be spent over the next decade.
Put that number in physical terms, as Seoul Economic Daily did: 14.8 gigawatts is the output of more than ten APR1400 nuclear reactors, the power draw of a metropolitan area of 26 million people.
Before October last year, Anthropic held one to two gigawatts. It has added at least 14.8 in under a year.
The single $100 billion-plus commitment to AWS is, by one reading, equal to about a fifth of AWS's entire contracted future revenue — the backlog Amazon has signed but not yet delivered. One customer, big enough to move the math on one of the world's largest cloud businesses.
None of this is priced on return on capital the way a hyperscaler prices a server refresh. Anthropic is racing an IPO clock — its S-1 is expected within weeks — and the asset investors are staring at is not margin, it is secured capacity. Compute you cannot get at any price is the thing that kills a frontier lab, so the lab pays to remove that risk entirely.
OpenAI is running the same play. It anchored Firmus's Malaysian AI factories in a multi-year deal that pushed the Sydney-based company's contracted capacity past 900 megawatts, and it is spending toward 30 gigawatts and up to $750 billion by 2030.
Notice where OpenAI is buying. Malaysia. And notice who is building for it.
In Hyderabad, TCS's HyperVault subsidiary and partners committed ₹70,000 crore — about $7.4 billion — to a one-gigawatt campus explicitly designed, in TCS's words, to serve frontier AI companies like OpenAI and hyperscalers. Down the coast, HCLTech is putting ₹14,257 crore — a sum that, per its own announcement, includes financial assistance from the Government of Odisha — into an AI data center in Odisha's Sovereign AI Park.
Read those two Indian projects together and the pattern is unmistakable: Asian capital and Asian land, built to serve Western labs. Firmus's co-CEO said the quiet part out loud — the moment Asia-Pacific "becomes a producer of intelligence, not just a consumer of it."
Here is the common thread across Seoul, San Francisco, Kuala Lumpur and Hyderabad. Every one of these buyers is committing for a decade, up front, for supply security rather than for the best unit price.
That single behavioral fact reorganizes the entire market. When the buyer pre-commits ten-year offtake to guarantee allocation, all the discipline migrates to the seller.
The buyer has already surrendered its leverage — you cannot walk away from a fixed appropriation, and you cannot walk away from an IPO you've promised investors is backed by 14.8 gigawatts. So the thing being priced stops being manufacturing cost and becomes allocation itself.
That is what Korea's ₩27.5 billion actually buys. Not a rack at cost-plus, but a guaranteed slot in Nvidia's queue at a moment when SK Group's chairman is telling CNBC the worst shortage arrives in 2027 and tying capacity expansion to long-term orders, not spot demand.
Note who is being disciplined and who is not. Samsung's board approved the largest shareholder return in Korean corporate history — ₩90 trillion to ₩110 trillion — in a year no one would call difficult. The memory makers are converting scarcity into dividends and buybacks rather than a capacity race, which is precisely the condition under which prices stay high.
The Korean government sits on the other side of that transaction. It is a sovereign buyer paying a premium to a supply chain that has decided not to race, funded from an appropriation that cannot reprice mid-year even as the underlying hardware does.
This is why the Korean line item is a benchmark and not an anomaly. A sovereign budgeting double the sell-side estimate, and stating shortage as the reason, is the market-clearing price revealing itself — the number a buyer will actually pay when certainty matters more than IRR.
The consequence is that chipflation stops being a cyclical squeeze you wait out. It becomes a structural feature of a market whose marginal buyer is, by design, insensitive to price.
A hyperscaler optimizing return on capital is the buyer who disciplines a market — the one who walks when the number gets silly. Replace that buyer with a sovereign defending national capability and a lab defending its IPO, and you have removed the only actor with an incentive to say no.
And the geography follows the willingness to pay. The racks, the gigawatts, the sovereign parks are landing in Korea, Malaysia and India — pulled there by exactly the buyers most willing to overpay for certainty.
The old question was how cheaply the world could manufacture a rack. The new one, the one Korea's leaked budget answers in a single line, is how much a buyer will pay to be sure the rack shows up. The answer, for now, is ₩27.5 billion — and rising.