Consider the line item every AI infrastructure deck quietly buries: the depreciation schedule. A GPU is economically dead in three to five years. The hall that houses it โ switchgear, cooling loop, substation feed โ amortizes across ten to fifteen. That asymmetry means an operator scaling aggressively reports revenue growing roughly linearly against a cost base that front-loads. Public equity prices that curve without mercy. Last quarter, a Korean issuer with an explicitly capital-intensive story found out how much mercy is on offer.
I went looking for the numbers. There were none. The coverage was a single paragraph: no ticker, no offer price, no bookrunner, no listing date, no subscription multiple, no first-week performance. The word "latest" carried the entire payload, implying a sequence of prior disappointments behind it. In a euphoric market, an unreported number is a withheld number. That absence is the most informative datum in the piece, and I will treat it as such. A release that omits its own price is telling you which of its claims it believes.
Method before claim, because the evidence base is thin enough that over-reading it would be malpractice. I graded the source before I graded the event: one usable fact โ a Korean company's IPO underperformed โ and zero identifying variables. No name, no sector, no board, no size, no pricing outcome. Timing indeterminate. A single-paragraph brief on a crypto vertical, no byline, no quote, no counterparty response. My prior is aggregated or auto-summarised copy, not primary reporting.
So the analysis splits. Track A: what this event type says about AI asset pricing โ analysable, and I will do it with rigour. Track B: anything about the individual issuer โ speculation, labelled as speculation. I will not dress a guess as a finding.
Track A has a genuine anchor. CoreWeave priced at $40 in March 2025 against a $47โ55 indicative range, landing near $19โ23 billion against a private-round target around $35 billion. A 35โ45 percent haircut, on the record โ and note that CoreWeave still traded. Discounting is not death. That is the first clean measurement of what happens when an AI infrastructure asset is dragged from a narrative multiple to an earnings multiple. Trust is math, not magic โ and the math arrived with a date on it.
The mechanism is not sentiment. It is a function mismatch.
Private AI rounds price on narrative: addressable market, a hyperscaler logo in the appendix, a plausible inference-margin path. Public markets price discounted free cash flow. Between those two functions sits an asset class whose cash profile is "enormous outflow now, uncertain inflow later." While the exit channel is open, that gap is a spread early investors harvest. When the channel narrows, it becomes a loss someone must mark. The Korean event is one instance of that marking.
Korea amplifies it. The local IPO market runs on retail subscription momentum; 2024โ2025 issuance contracted sharply with broken-price rates elevated. Institutional book-building leverage rose accordingly. Add the convention of an optimistic indicative band, and the failure mode is specific: subscription multiples collapse and the first session gaps down. Loud locally, and roughly half as informative globally. It reflects local liquidity, not global AI fundamentals.
Two readings fit that price action and they point opposite ways. One: the bookrunner overpriced a weak asset โ market functioning, discipline applied. Two: demand genuinely evaporated โ market cooling. The coverage never separated them, though one number would have: whether the greenshoe was exercised. Unexercised means the syndicate absorbed stock. That resolves the ambiguity. It was not reported.
The structural layer is where the real information sits. AI infrastructure is migrating off equity. One large datacenter runs to ten figures; one frontier training run to nine. That intensity is incompatible with public equity's demand for predictable free cash flow. So financing shifts to private credit, GPU-backed lending, vendor financing. Composability is a double-edged sword โ and here it is the composability of liabilities. Vendor financing makes your supplier your creditor. GPU-backed lending pledges collateral depreciating on a three-to-five-year clock against loans written longer. That is duration mismatch wearing a growth costume, and it transmits risk from equity holders to creditors who never signed up for AI capex. Every layer of that stack assumes the layer beneath it is solvent. That is composability's bill arriving.
I have audited contracts that did exactly this โ pledged an asset whose realisable value decayed faster than the liability it secured. In Solidity the failure is a liquidation cascade inside one block. In structured credit it is a covenant breach inside one quarter. Same arithmetic, slower clock.
Korea sharpens it. Export-control quotas on advanced accelerators, high power prices, constrained land โ unit compute cost sits above US peers. Meanwhile the state expands: AI G3 ambitions, public compute commitments, dedicated funds. Public capital bids while private capital walks. That divergence cannot persist. It resolves as a national champion propped to viability or a fiscal sunk cost.
The talent channel is the one nobody models. A broken listing zeroes option value. Engineers with families in Seongnam do not wait for attempt two; they move to Samsung, to SK, overseas. Capital can be re-raised at a lower price. A dispersed design team does not reassemble, and tape-out cadence slips a full generation. In accelerators, one generation behind is out.
Consolidation already visible is defensive, not opportunistic. Rebellions merged with KT's Sapeon in 2024 โ scale manufactured to make a listing feasible while the window wobbles. FuriosaAI reportedly declined Meta at roughly $800 million to stay independent. If the window has cooled, that decision ages badly and becomes the reference case every Korean founder gets measured against.
Note what is missing from the price: any split between capex proceeds and operating proceeds. An issuer raising for silicon has a different risk profile from one raising for headcount. One paragraph could not tell me, and neither, apparently, could the brief.
One correction the coverage invites and I refuse: do not map a Korean AI investment slowdown onto the HBM cycle. HBM demand is set by global hyperscalers. SK Hynix and Samsung's memory franchise is not levered to a domestic startup's listing. That is a category error wearing a supply-chain costume.
The blind spot is sample selection. The brief frames one issuer's miss as a "global AI trade" signal while omitting the comparison that would settle it: what did Nvidia, TSMC and SK Hynix do over the same window? If they held or rose, the correct inference is issuer-specific quality, not regime change. Framing one unverified datapoint as macro is narrative amplification โ the same error I spent 120 hours on Uniswap V1 learning to distrust. A single anomalous branch in the price calculation looked systemic until I traced the call path and found it isolated. You cannot generalise from a node to a graph without the edges.
There is a reading the coverage would never print. A broken IPO is pricing discipline, not demand death. Speculation audits the soul of value. For capital not yet deployed, a 35โ45 percent haircut on AI infrastructure is an entry, not an alarm. And note who benefits: Nvidia, because alternatives thin out; the chaebols, because targets get cheaper; ASIC partners, because hyperscalers no longer fear startup disruption and can consolidate on internal silicon.
The risk that actually matters is downstream and mechanical. Lock-ups expire in six or twelve months. That is the second shock, and it lands on a calendar rather than a headline. Architects build, auditors break โ the break is scheduled.
Watch the withdrawal queue, not the broken price. An issuer that prices badly and trades down is noise; the next one that pulls its filing is signal. Patterns emerge from chaos, not noise โ and right now the queue is the only clean instrument on the table. If filings keep coming, Korea mispriced one asset. If they stop, the exit channel has closed, and every AI balance sheet in the region just acquired a new discount rate. Everything else is commentary.