Elice's $150M IPO Plummets 32% on Debut

Key Takeaways

• <ul><li>Elice raised 201.1 billion won ($150 million) in its Kosdaq IPO, the largest of 2026, despite oversubscribing over 400 times before shares fell 32% on debut.</li><li>CoreWeave and Nebius shares also declined following Elice's collapse, signaling broader concerns about AI infrastructure valuations.</li><li>Elice's first-half 2026 revenue grew 200% to 34.4 billion won, yet investor sentiment toward AI infrastructure has become disconnected from fundamentals.</li><li>SandboxAQ launched expanded Switch capabilities to help companies track AI agent costs, performance and security in production environments.</li><li>ABACUS framework achieved 99.67% macro accuracy using reusable programs, outperforming direct Kev at 64% and Qwen3.5-4B at 68.80%.</li><li>BOSS training-free policy achieved 60.5 macro-average accuracy with Llama-3.1-8B under a 4k token budget, outperforming competing methods across 22 of 24 test cells.</li><li>SOMA system reduces unnecessary AI reasoning by up to 36% by using a fast model to monitor a slower model for errors.</li><li>xMEMS began mass production of solid-state cooling chips XMC-2400 and XMC-4800, capable of lowering device temperatures by 10 to 25 degrees Celsius.</li><li>South Korea's revised AI Basic Act requires AI training data from government projects to be shared on the national AI Hub, with implementation set for April next year.</li><li>Gujarat plans three-tier AI training for five lakh higher education students covering basic literacy, discipline-specific applications and advanced employment skills.</li></ul>

Elice shares drop sharply in South Korean trading debut

Elice, a South Korean AI cloud-computing company, saw its shares fall as much as 36% from the IPO price in its Kosdaq market debut. The stock traded at 58,600 won in afternoon trading, down from the IPO price. The decline reflected investor caution toward artificial-intelligence stocks. The broader Kosdaq index was also down about 3%.

Elice shares plunge on debut amid AI investment doubts

Elice Group shares fell 28% in their market debut on Tuesday, closing at 65,300 won versus an IPO price of 90,500 won. The IPO raised 201.1 billion won ($150 million) despite strong demand, with institutional investors subscribing 388.84 times and retail investors 408.26 times. The company plans to use proceeds to build data center infrastructure in South Korea and expand into the United States, Japan and Singapore. Investor concerns about high borrowing costs and uncertain returns on AI infrastructure spending weighed on the debut.

South Korea stock market signals new AI trade warning

South Korea's Kospi index fell 4% as chipmakers Samsung and SK Hynix extended the AI-stock selloff. AI cloud firm Elice crashed 32% in its market debut despite overwhelming demand for its IPO. Elice raised about 201 billion won ($151 million) in the Kosdaq's largest IPO of the year. Institutional and retail subscriptions topped available shares by roughly 400 times. The company's IPO valued it more richly than similar AI cloud companies whose shares had already fallen 20% to 60% in the prior month.

Elice IPO collapse adds pressure to AI infrastructure stocks

Shares of CoreWeave and Nebius fell in premarket trading after Elice sank 32% on its Seoul market debut. Elice shares closed at 61,500 won, well below the IPO price of 90,500 won, despite institutional demand at 388.8 times and retail subscriptions over 400 times the allocation. The company raised about 201 billion won ($151 million). CoreWeave fell 4% and Nebius fell 4.8%. Analyst Daniel O'Regan said the debut signals deteriorating sentiment toward neocloud companies as investors question whether AI infrastructure returns will justify the capital and borrowing costs required.

AI IPO oversubscribed 400 times still fell 32% on debut

Elice, a Seoul-based GPU data center company, saw its IPO oversubscribed more than 400 times by retail investors who placed about 10.3 trillion won in deposits. The stock still fell as much as 32% on its October 20, 2026 debut, dropping to an intraday low of 61,400 won versus an IPO price of 90,500 won. The deal raised 201 billion won ($151 million), the largest Kosdaq IPO of 2026. First-half 2026 revenue rose 200% to 34.4 billion won with operating profit of 10.3 billion won. Analyst Albert Yong said sentiment toward AI infrastructure has weakened and become disconnected from fundamentals.

Korean AI IPO Oversubscribed 400 Times Still Crashed 32% on Debut

Korean retail investors deposited about 10.3 trillion won for Elice, a Seoul AI data center company, oversubscribing the IPO more than 400 times. The stock still fell 32% on its October 20, 2026 debut, dropping to 61,400 won from an IPO price of 90,500 won. The company raised 201 billion won, about $151 million, the largest Kosdaq IPO of 2026. Elice first-half 2026 revenue grew 200% to 34.4 billion won, but weak market sentiment toward AI infrastructure drove the decline.

Elice IPO Plunge Sparks Global Worries About AI Investment Boom

Elice fell nearly 32% from its offer price on its first day of listing on the KOSDAQ on October 20, 2026. Major foreign media including Bloomberg, Reuters, Barron's and CNBC reported the drop as a warning sign for global AI infrastructure investment. Elice closed at 61,500 won against an offer price of 90,500 won. Despite strong first-half revenue growth and improved profitability, the IPO failure raised concerns about high valuations and future fundraising for AI companies. U.S. data center stocks also faced downward pressure after the news.

ABACUS Uses Programs to Boost Judgment Model Reasoning

ABACUS is a framework that turns judgment models into powerful multistep reasoning tools using reusable programs. A generative teacher designs procedures offline while the judgment model resolves questions online. Nine independently built programs solved all 450 executions of 150 development inputs on three BIG-Bench Hard tasks. In deployment, these programs achieved 99.67% macro accuracy compared to 64% for direct Kev and 68.80% for Qwen3.5-4B. The programs also ran 8.43 times faster than Qwen's unrestricted response policy.

BOSS Improves LLM Reasoning Under Strict Token Budgets

BOSS is a training-free policy that helps large language models decide when to stop generating, create another candidate, or adjudicate existing answers under a hard token cap. It manages generation and adjudication allocation using one shared token ledger. Across two Llama-3.1 models and four reasoning benchmarks, BOSS achieved the highest accuracy in 22 of 24 model-task-budget cells. At Llama-3.1-8B with B=4k, it reached 60.5 macro-average accuracy, outperforming Adaptive-Consistency, TALE-EP and fixed-schedule League methods.

SandboxAQ Expands Switch to Monitor AI Agent Costs and Security

SandboxAQ launched expanded Switch capabilities on October 19, 2026 to help companies track AI agent costs, performance and security. Switch Trust discovers, scans, evaluates and monitors AI agents across connected environments. It detects vulnerabilities, tests agents with adversarial probes and protects them in production with guardrail policies. Switch ROI measures agent spending at the session and turn level, calculates value against cost and provides optimization recommendations. Mount Sinai West's MSW Lab is already using Switch Trust to build G Board for safe clinical AI applications.

New SOMA Tool Helps AI Models Self-Correct Errors

Researchers created a new system called SOMA to help AI reasoning models catch their own mistakes. This tool uses a fast System One model to watch a slower System Two model while it works. If the fast model sees an error, it can stop the process early or take over control. Testing shows this method reduces unnecessary thinking by up to 36 percent and improves accuracy on hard questions. The system works without needing extra training and helps AI avoid getting stuck on unanswerable problems.

Smart Routing System Improves AI Evidence Checks

A new software system uses benefit-targeted routing to decide when AI models need human help during evidence screening. Researchers tested this on 6,164 manufacturing cases and found it achieved high agreement rates with human reviewers. The system learns to balance the cost of extra checks against the benefit of finding correct answers. It also uses strict controls to manage data and budget costs without needing to retrain existing policies. A final audit of 200 pairs showed a weighted merge precision of 97.45 percent after eight author reviews.

Medical AI Tools May Prioritize Company Profits Over Patients

A new article questions who medical AI tools are truly working for as they become more common in healthcare. While these systems help doctors diagnose diseases and create treatments, they are often built by companies with their own financial goals. For example, IBM's Watson for Oncology is owned by IBM, which profits from selling the technology to hospitals. This raises concerns about whether the AI provides unbiased advice or if it favors the company's interests. Experts warn that patients and doctors need to ask tough questions about the ethics behind these powerful tools.

xMEMS Starts Making Solid-State Cooling Chips for AI Devices

The company xMEMS has begun mass production of new solid-state cooling chips called XMC-2400 and XMC-4800. These chips use tiny vibrating structures to move air without spinning fans or bearings. The XMC-2400 chip is designed for small devices like AI glasses and will appear in products by early 2027. The larger XMC-4800 chip is meant for smartphones and data center hardware that need more airflow. Testing shows these chips can lower temperatures by 10 to 25 degrees Celsius in various devices.

Shoppers Worry About AI Pricing Tools in Grocery Stores

Consumers are becoming anxious about retailers using AI to set prices for everyday items like groceries. New digital price tags are replacing paper labels in stores, and AI tools help businesses decide what to charge based on market data. Some people fear these systems might charge different prices to different shoppers based on personal information. While companies say they price products, not people, there is confusion about how much data is collected. Lawmakers in at least 30 states are now considering new laws to regulate these algorithmic pricing practices.

Gujarat plans AI training for five lakh higher education students

The Gujarat state education department has drafted a policy for three-tier AI training for five lakh students. The programme covers basic AI literacy, discipline-specific applications and advanced employment-oriented skills. It also aims to promote responsible AI use and protect student and institutional data with safeguards for different fields.

South Korea debates rules for disclosing AI training data from government projects

South Korea now requires AI training data from government-backed projects to be shared on the national AI Hub. Questions remain about whether companies must also disclose data they paid for themselves or mixed with private investment. The revised AI Basic Act passed the National Assembly on Oct. 1 and a Cabinet meeting on Oct. 13, with the law set to take effect in April next year. A key example is the dokpamo foundation model project, where 29 types of training data totaling about 35.44 million items were disclosed by teams including Naver Cloud, Upstage, SK Telecom and NC AI.

Sources

• Elice Shares Slide in Kosdaq Market Trading Debut
• Elice shares plunge 28% on debut as investor doubts over AI trade deepen
• South Korea's stock market is flashing another warning about the AI trade
• Elice’s IPO collapse puts AI infrastructure stocks under fresh pressure
• Retail Investors Oversubscribed This AI IPO 400 Times. It Fell 32% Anyway
• Retail Investors Oversubscribed This AI IPO 400 Times. It Fell 32% Anyway
• Elis debut slump fuels global AI investment jitters from South Korea
• ABACUS: Program-Guided Multistep Reasoning over Frozen Judgment Models
• BOSS: Token-Budgeted Online Sampling and Selection for LLM Reasoning
• SandboxAQ Expands Switch to Track AI Agent Costs, Performance and Security
• SOMA: Continuous Monitoring and Intervention for System Two Reasoning with System One Models
• Benefit-Targeted Routing and Bounded Revision for LLM-Based Evidence Screening in Manufacturing Supplier Discovery
• Medical AI's cost-benefit dilemma
• xMEMS Begins Mass Production of Solid-State Cooling Chips for AI Glasses and Data Center Hardware
• Consumer apprehension grows as retailers adopt AI-assisted pricing tools
• Gujarat proposes three-tier AI training for 5 lakh students
• Mandatory disclosure of AI training data from government-backed projects raises questions over detailed criteria