Everything About AI
NextFin News - Elice Inc., the South Korean data-center builder that priced the year's largest Kosdaq initial public offering at the top of its range, fell as much as 32% in its trading debut on Tuesday, a brutal reversal for a deal that had drawn more than 400 times subscription in the retail tranche. Shares slid to a low of 61,400 won, from an offer price of 90,500 won, erasing roughly 300 billion won of market value within hours of the opening bell and handing debut buyers a loss of about 29,100 won per share. The move signals that the market for AI infrastructure listings has shifted from "any data-center story will do" to a harsher test of profitability, power access and execution credibility.
The timing is the story within the story. Elice listed less than two weeks after Firmus Grid Ltd., a data-center operator backed by Nvidia Corp., pulled its own planned IPO amid investor concerns over borrowing costs, competition and uncertain returns on massive AI-related capital spending. Together, the two episodes mark a turning point in how public markets are pricing the neocloud buildout — and they arrive just as the category's listed leaders are themselves under pressure.
The Deal That Looked Unbreakable
On paper, Elice's offering had every marker of a can't-miss debut. The company raised 201 billion won ($151 million), the biggest Kosdaq IPO of 2026, and priced at 90,500 won per share — the top of a marketed range of 70,400 to 90,500 won. At that price the post-listing market capitalization exceeded 1 trillion won, making Elice the first Kosdaq debut of the year to cross that threshold. Demand appeared overwhelming: the retail tranche was subscribed 408-to-1, with roughly 290,000 applications and about 10.3 trillion won in deposits, while the institutional bookbuilding ran 388.8-to-1.
That 10.3 trillion won in retail deposits is a striking figure in its own right. It represents roughly 50 times the capital the company actually raised — a measure of how deeply the AI infrastructure narrative has penetrated South Korea's retail investor base. But an oversubscribed book is not the same as a stable aftermarket, and Tuesday's session proved the distinction.
The subscription data also contained a warning that the market read quickly. Only about 13.7% of the institutional applied volume committed to a post-listing lock-up of at least 15 days. In plain terms, the professional buyers who drove the book were not signaling a multi-year conviction hold; many were positioned for a quick exit. When a debut stock opens weak, that thin lock-up coverage removes the natural floor that long-only holders usually provide. The math is unforgiving: with 388.8-to-1 institutional demand but barely one-eighth of that volume agreeing to hold, the tradable float on day one was far larger than the headline subscription ratio suggested.
The pricing itself deserves scrutiny. Elice's underwriters, Mirae Asset Securities and Samsung Securities, set the offer at 90,500 won — the very top of the 70,400-to-90,500 won range. Pricing at the ceiling after a 400-times retail subscription is a familiar pattern in hot IPO markets: it maximizes proceeds for the issuer but leaves zero cushion for the aftermarket. The midpoint of the range was 80,450 won; had the deal priced there, Tuesday's 61,400 won low would have been a 24% drop instead of 32%. The top-of-range decision transferred the entire valuation risk from the company to debut investors.
Why the Market Turned So Fast
Three forces converged on Tuesday morning. First, the Firmus Grid withdrawal reset primary-market sentiment for data-center names. Investors who had been willing to pay for capacity promises became more attentive to what is actually built, powered and contracted. Second, Elice's valuation left almost no room for error: pricing at the top of the range after a 400-times retail subscription baked a near-flawless execution story into the opening print. Third, the weak lock-up signal meant the aftermarket was effectively left to retail holders just as the Firmus overhang hit.
The company's fundamentals are real but small relative to the valuation. In the first half of 2026, Elice posted revenue of 34.4 billion won and operating profit of 10.3 billion won, with operating profit already exceeding the 5.2 billion won earned for all of 2025. That is an operating margin of roughly 30% — impressive for an infrastructure builder — but it is a run rate measured in tens of billions of won against a market value above 1 trillion won. At the offer price, the company traded at roughly 29 times its half-year revenue — about 15 times on an annualized basis — a multiple that assumes years of uninterrupted hypergrowth.
"As global competition in AI technology intensifies, how quickly you build data centers and how efficiently you operate them matters enormously," said Kim Jae-won, chief executive of Elice, at an IPO briefing in Seoul on Sept. 30. "Using the KOSDAQ listing as a springboard, we will build regional AI data centers to ease domestic power bottlenecks and increase exports to become a global company."
The admission embedded in that quote — that South Korea faces domestic power bottlenecks — is the crux of the investment case and the risk. Elice's expansion plan depends on converting secured power intake into revenue-producing capacity on schedule, and the CEO's own framing concedes that power, not chips, is now the binding constraint.
The Firmus Grid Shadow
To understand why Elice broke on debut, the Firmus Grid episode is essential context. Less than two weeks before Elice's listing, Firmus Grid — a data-center operator with Nvidia Corp. backing — withdrew its planned IPO. The withdrawal came after investors pressed the company on three points: rising borrowing costs for capital-intensive projects, intensifying competition for AI tenants, and uncertain long-term returns on the massive spending required to build data-center capacity.
The Firmus case exposed a fault line in the neocloud IPO narrative. During 2025 and early 2026, investors rewarded capacity announcements almost reflexively: a signed land deal, a power allocation, a GPU procurement commitment — each was treated as a step toward a valuable asset. Firmus Grid's withdrawal signaled that the market now wants proof of construction, proof of tenancy and proof of returns, not just announcements. Elice, despite its stronger profitability profile, was caught in the same sentiment shift.
The contrast between the two companies is instructive. Elice came to market with actual revenue, a 30% operating margin and an operating 10MW facility. Firmus Grid was seen as more speculative, with a larger share of its value tied to unbuilt capacity. If the market is punishing Elice despite its superior fundamentals, it suggests the repricing is broad — affecting the entire data-center IPO complex, not just the weakest names.
Cyclical or Structural: Deciding the Call
The central analytical question is whether Tuesday's drop is a cyclical cooling of IPO appetite or a structural repricing of the neocloud model. The answer is both, operating on different time horizons — and conflating them is the most common error investors make in moments like this.
The cyclical leg is straightforward. Primary-market sentiment for data-center names shifted sharply after Firmus Grid's withdrawal, and sentiment swings of this kind typically mean-revert once the next batch of contract announcements and earnings arrives. If Elice delivers its 40MW southwest site on schedule and reports strong quarterly revenue growth, retail enthusiasm — which clearly exists, given the 408-to-1 subscription — can return quickly. Cyclical moves are about timing, not thesis. History offers a template: IPOs that gap down more than 20% on debut frequently find a floor within weeks if the underlying business continues to grow, because the initial flush exhausts the motivated sellers.
The structural leg runs deeper and is more consequential. The neocloud trade was built on a simple premise: scarcity of AI compute means any operator with GPUs and a power connection can command premium pricing. That premise is eroding as the category matures, and the evidence is visible in the public comparables. CoreWeave, the U.S. category leader, went public in March 2025 and its shares have fallen sharply from their 52-week high near $152, trading around $61 in mid-October for a market capitalization of roughly $34 billion. Nscale, the European challenger, filed for a New York listing in September 2026 and is meeting investor resistance: some prospective buyers are urging a valuation in the $16 billion to $19 billion range, well below the roughly $30 billion the company has targeted, citing operational slippage and a chilled market for neocloud listings.
The frontier is moving from "do you have capacity?" to "can you deliver power, cooling and financing at the lowest cost?" — a contest where scale, balance-sheet depth and sovereign relationships matter more than first-mover status. In that contest, Elice is both advantaged and exposed. Its modular construction approach lowers upfront capital cost, and its selection alongside NAVER and Samsung SDS for a 2 trillion won Ministry of Science and ICT GPU project gives it policy standing. But it lacks the balance-sheet depth of the hyperscalers and the sovereign backing that larger national champions can summon.
Elice's answer to the cost contest is its prefabricated modular data center, the first of its kind deployed in Korea, which the company says cuts investment cost by about 40% compared with conventional facilities. Modular construction works by shifting work from the building site to the factory: power modules, cooling skids and server halls arrive as pre-tested units that bolt together on site. The economic payoff is twofold — lower capital cost per megawatt and faster time to revenue, because a facility can be commissioned in months rather than the years a conventional build requires. That speed matters when customers are racing to deploy AI models and every quarter of delay risks losing a tenant to a competitor.
Currently Elice runs a 10-megawatt center in greater Seoul, has secured about 130MW of power intake, and targets a 40MW site in the southwest by the end of 2026. Converting 130MW of secured intake into operating capacity is the single most important execution task facing the company. At industry-typical power densities, 130MW could support tens of thousands of GPUs — enough to scale revenue several times over if fully utilized. But secured intake is not the same as energized capacity, and South Korea's grid constraints mean the timeline carries real risk.
That policy moat is real, but it is not the same as a pricing moat. Being chosen for a government GPU project guarantees a customer; it does not guarantee margins if power costs rise, if construction is delayed, or if hyperscalers decide to compete directly in the same regional markets.
The Second-Order Read: The Bottleneck Has Moved
The first-order lesson from Tuesday is that even hot IPOs can break on day one. The second-order lesson is more consequential: the binding constraint in the AI infrastructure buildout is shifting from chips to electricity.
For most of the AI boom, the scarce input was GPUs. Nvidia's supply chain dictated who could build, and neocloud operators earned their margins by securing allocations ahead of the crowd. Now, as GPU supply gradually loosens and data-center pipelines swell, the scarce input is becoming grid capacity, permitted sites and cooling. Elice's own expansion plan — building regional AI data centers to "ease domestic power bottlenecks," in Kim's words — concedes the shift. So does the company's emphasis on modular construction, which trades lower per-unit cost for faster deployment. The CEO's language is telling: he did not say Elice would win by securing more chips, but by easing power bottlenecks.
This reframes the value chain for investors. The winners of the next phase may not be the pure-play operators with the flashiest GPU fleets, but the companies that control sites, grid interconnections and permitting. Utilities, independent power producers, and developers with shovel-ready sites and grid access are likely to capture an increasing share of the economics, while pure capacity operators face margin compression as competition intensifies. It also explains the paradox of a 400-times oversubscribed deal collapsing on debut: retail investors were buying the AI compute narrative, while institutions were underwriting the power story, and they priced accordingly. With only 13.7% of institutional volume locked for 15 days, the aftermarket had no ballast when the Firmus Grid news resurfaced.
The implication extends beyond Elice. For the pipeline of data-center IPOs considering listings in the coming quarters — including names targeting the second half of 2026 and early 2027 — Tuesday's session sets a new pricing discipline. Issuers that come to market with top-of-range valuations and weak lock-ups will likely meet the same fate. The market is no longer paying for narratives; it is paying for megawatts under contract, construction milestones met, and margins that survive the capex ramp.
The Counter-Thesis, Taken Seriously
The strongest argument against a gloomy read is that Elice is a small, early-stage company being punished for a normal first-day volatility pattern, and that its roughly 30% operating margin and sovereign-AI backing make it a long-term winner regardless of short-term price action. Korea Investment & Securities cited Elice as a beneficiary of expanding data-center construction, and the Ministry of Science and ICT project gives it a policy moat that smaller rivals lack. South Korea's retail market, which poured 10.3 trillion won into the subscription, has demonstrated staying power in other technology names and could absorb the initial volatility.
There is also a valuation-relativity argument. Compared with U.S. neocloud peers trading at many times revenue with negative operating margins, Elice is already profitable, growing, and priced at a fraction of the category's peak multiples. From that vantage point, a 32% first-day drop looks less like a fundamental rejection and more like the price of admission to a public market that is temporarily risk-off.
That case is not without merit, but it confuses a good company with a good stock at any price. A market capitalization above 1 trillion won on 34.4 billion won of half-year revenue implies a price-to-sales multiple that leaves no margin for execution delays, power shortages or a slowdown in AI capital expenditure. The counter-thesis holds only if Elice converts its 130MW of secured power intake into revenue-producing capacity on schedule and maintains margins as capex ramps. If the company stumbles on either, the debut plunge will look like the first leg of a longer repricing rather than a buying opportunity.
What to Watch: Scenarios by Time Horizon
Short term (weeks): Sentiment is fragile and driven by headlines. A stabilization above 75,000 won — roughly 17% below the offer price — would signal that the initial flush has found a floor. A break below 60,000 won would invite momentum selling and could drag other recently listed Korean technology names lower.
Medium term (quarters): Fundamentals take over. The key data points are quarterly revenue growth, operating margin, and progress on the 40MW southwest site. The falsifying signal for the bull case is specific: if Elice misses the end-2026 target for the 40MW site, or if operating margin compresses below 20% as capital spending ramps, the structural-bull thesis is broken. Conversely, if the company reports revenue growth above 50% year over year while holding margins above 25%, the debut drop will be reinterpreted as an overreaction.
Long term (years): The structural question is whether modular data centers can win meaningful share against hyperscalers and sovereign-backed champions. Elice's 40% cost advantage on modular construction is the hinge. If it proves durable at scale, the company can grow into its valuation. If hyperscalers match the economics, or if power allocation favors state champions, the multiple will compress further.
Base case: Elice delivers the 40MW site with modest slippage, revenue grows strongly but margins compress toward 25% as capex ramps, and the stock trades in a wide range as the market reassesses neocloud multiples. Upside case: the modular model scales faster than expected, the sovereign GPU project converts to multi-year contracts, and the stock recovers toward the offer price. Downside case: power bottlenecks delay deployment, margins fall below 20%, and the valuation rerates toward smaller regional infrastructure peers.
Elice's technology may be sound, and South Korea's AI infrastructure buildout is real. But Tuesday's market delivered a clear verdict that extends far beyond one stock: in the AI infrastructure buildout, the price of admission just went up, and the bottleneck has moved from chips to the grid.
Explore more exclusive insights at nextfin.ai.
