32% Down on Day One: Elice's KOSDAQ Crash Is Now the Tape Every Other AI-Data-Center IPO Has to Beat

Elice's retail tranche was oversubscribed roughly 408 to 1, with institutional book-building at about 388.84 to 1, and Korean retail investors parked about 10.26 trillion won in subscription deposits for what was the largest tech listing of the year on KOSDAQ and the second-biggest Korea IPO of 2026 behind only K Bank's US$337 million offering. The size of the bid told you how badly households wanted a Korean neocloud name; the size of the first-day drop told you the institutional side had already walked away from the price.

Elice had priced on October 6, before the second leg of the AI sell-off. By the time the stock started trading, KOSPI had just closed at 5,554.22 on October 15, down 7.57% on the session and 39% below the June 22 record of 9,114.55, with circuit-breaker halts triggered by Samsung Electronics and SK Hynix. The pricing date was the unlucky hinge of the trade — and no amount of oversubscription could make Korean retail enough to absorb what institutions were offloading.

By June 2026, Samsung Electronics and SK Hynix together represented more than half of KOSPI's market value. The two memory giants had become the lever for the country's AI exposure: Samsung's Q3 2026 operating-profit guidance of 107.4 trillion won (about US$80 billion, +783% year-on-year) and SK Hynix's Q2 2026 revenue of 79.3 trillion won (+257% year-on-year) on HBM scarcity are the real earnings engine. Korean neoclouds were supposed to be the next leg of that story.

Elice's prospectus says the cloud share of revenue went from about 7% in 2023 to 76% in H1 2026 — a pivot story, not a memory story. The company secured 130MW of intake (40MW south-west, 80MW south-east) and plans to deploy 2,560 NVIDIA B300 GPUs across four modular AI data centers. There is no fundamental question here that SK Hynix or Samsung has to answer; this is the layer below them, the integrator.

Elice's business is to wrap NVIDIA GPUs and Korean power in liquid-cooled, high-density modular data centers and rent the capacity back out. NVIDIA's June 7, 2026 deal with SK Hynix —260,000 chips destined for Korean AI factories with Samsung and Hyundai — confirmed that the upstream end is locked up. But integrators with no captive memory franchise of their own are the ones the market is starting to question.

What makes the Elice debut worse than a one-off embarrassment is the queue behind it. Within the past month, four other AI-infrastructure listings have either pulled, been delayed, or been forced to negotiate down:

The pattern is uniform: every one of these issuers priced into an AI-demand peak and is now being asked to discount either valuation, timing, or both. DayOne's prospectus is the cleanest window: H1 2026 revenue of US$512 million versus a US$77.2 million net loss (compared with US$151.5 million of revenue and a US$12.6 million loss a year earlier) — three-times growth, but a wider absolute loss. The model works on growth; it does not yet work on profit, which is exactly the gap investors are using to walk down the bid.

There is a second-order Elice risk that lives in the KOSPI leverage unwind. By June 19, 2026, ordinary Korean investors held roughly9.15 trillion won in SK Hynix-linked leveraged products and 5.22 trillion won in Samsung-linked products, against a backdrop in which Korean stock ownership rose from about 6 million people in 2019 to more than 14.5 million by the end of 2025. The retail bid for Elice (10.26 trillion won in deposits) is the same force that drove the leveraged funds: a populous, leveraged young, and apparently disconnected from underlying fundamentals.

Regulators already responded to the July unwind by raising the minimum cash subscription for single-stock leveraged funds from 10 million to 30 million won, halting new launches and tightening controls when the funds diverge from NAV. That fix did not extend to IPO subscription deposits, where 10.26 trillion won of small-check money was already committed before the listing.

Underneath the IPO wreckage, the underlying infrastructure spending hasn't stopped. Naver is building a sovereign AI factory with NVIDIA and Brookfield at gigawatt scale (starting at 55MW); NHN Cloud has 7,656 B200s in Yangpyeong, with 6,120 of them earmarked for national AI compute; Samsung SDS is putting427 billion won into a Korean AI data center; and South Korean conglomerates have publicly committed at least US$880 billion to AI chips and data centers. The Korean AI data-center market is forecast to reach roughly US$16.2 billion by 2031, up from US$5.04 billion in 2025, per industry trackers.

The economic structure is the same: somebody has to integrate the silicon, the power, the cooling and the rack. The question the market is now asking is how much multiple that integration layer deserves versus the silicon layer. Elice was the proxy vote, and the proxy vote was No.

Short-term (days to quarters): Elice's secondary close on October 21 is the first read on whether Korean retail steps in and holds the line. DayOne's first roadshow meetings into a November window are the second. Nscale's S-1 conversations about a US$16–19B valuation versus a US$35B target are the third. A clean bounce in any of the three would let the bears say this was an idiosyncratic Korea tape. A second bad print — especially DayOne pricing below the marketed range — would extend the multiple compression into the US queue.

Long-term (1–3 years): if the buildout numbers hold (US$880 billion of Korean AI commitments; US$103 billion of Nscale contracts; DayOne's US$34 billion backlog), the integration layer re-rates when GPU-funded revenues actually start converting to cash. Until then, leverage and growth-as-financing Doomsday loops are what investors will pay attention to. The Korean AI data-center market is forecast to roughly triple by 2031, but the public-market route into that growth just got more expensive.