{"id":105879,"date":"2026-07-31T22:46:09","date_gmt":"2026-07-31T22:46:09","guid":{"rendered":"https:\/\/www.europesays.com\/korea\/105879\/"},"modified":"2026-07-31T22:46:09","modified_gmt":"2026-07-31T22:46:09","slug":"backend-ai-maker-lablup-wins-fastest-kosdaq-approval-of-2026-as-market-slips","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/korea\/105879\/","title":{"rendered":"Backend.AI Maker Lablup Wins Fastest KOSDAQ Approval of 2026 as Market Slips"},"content":{"rendered":"

On the same morning the KOSDAQ lost 2.7%, South Korean AI infrastructure software company Lablup Inc. received word that it had cleared the Korea Exchange’s preliminary listing review \u2014 the fastest such approval for any general company on the exchange in 2026<\/a>. The company filed its application on May 21; the two-month turnaround from filing to green light is the shortest recorded in the current listing cycle. NH Investment Securities is serving as lead underwriter.<\/p>\n

What Lablup is bringing to public markets is not a model or a chip. It is the software layer that determines whether the GPUs an organization has already bought will actually run at capacity \u2014 or sit idle. Its flagship platform, Backend.AI, manages AI workloads across GPU clusters using a patented virtualization technique that intercepts CUDA API calls in software, partitions physical GPUs into precise fractional slices, and routes competing workloads to shared hardware with less than 5% overhead<\/a>. The company claims that approach has lifted GPU utilization from roughly 25\u201330% to 85\u201395% in production deployments. That claim comes from Lablup’s own marketing materials, not an independent audit, but the underlying customer list is real: more than 120 organizations \u2014 including Samsung Electronics, LG Electronics, Shinhan Bank, the Bank of Korea, the Republic of Korea Navy, Samsung Medical Center, and the University of Southern California \u2014 are running Backend.AI in production today<\/a>.<\/p>\n

The listing approval arrives in the context of Korea’s deepest IPO drought in years. Only 17 companies completed listings across the KOSPI and KOSDAQ in the first half of 2026, down from 38 in the prior year, and total capital raised fell by nearly half<\/a>. The KOSDAQ, Korea’s technology-weighted secondary exchange, has been particularly hard-hit as institutional capital has rotated into semiconductor giants \u2014 SK Hynix, Samsung Electronics \u2014 at the expense of smaller listings. Against that backdrop, Lablup’s record-speed preliminary approval on a 2.7%-down day makes a pointed statement about where reviewers think the value in Korea’s AI story actually sits.<\/p>\n

How GPU Virtualization Actually Works \u2014 and Why MIG Hardware Cannot Do What Backend.AI Does<\/p>\n

The central technical claim that makes Lablup’s equity story worth examining is not “we do GPU orchestration.” That category is crowded, with NVIDIA’s Run:ai, Kubernetes-native GPU scheduling, and Slurm covering most of the enterprise HPC market. The claim worth examining is specifically about GPU virtualization: the ability to take one physical GPU and run multiple concurrent workloads on it simultaneously, with isolation guarantees, without dedicated hardware partitioning.<\/p>\n

NVIDIA’s solution to this problem is Multi-Instance GPU (MIG), introduced with the Ampere A100 in 2020<\/a>. MIG is a hardware-level feature that physically partitions a GPU’s streaming multiprocessors, L2 cache banks, memory controllers, and DRAM address buses into fully isolated instances \u2014 up to seven per A100, and comparable splits on H100 and B200. The hardware isolation is real and the performance guarantees are strong. The limitation is structural: MIG only exists on specific datacenter SKU generations. It does not work on AMD Instinct GPUs. It does not work on Intel Gaudi accelerators. It does not work on Korea’s domestic AI chips from FuriosaAI or Rebellions. And it does not work on NVIDIA’s own consumer RTX-series cards, which many academic institutions and smaller enterprises run.<\/p>\n

Backend.AI’s fGPU (fractional GPU) takes a different path. Rather than requiring hardware partitioning, it intercepts CUDA API calls inside containers at the software layer<\/a> \u2014 a technique sometimes called API forwarding or shim-layer interception. The interception layer sits between the application and the GPU driver, enforcing memory allocation limits and compute quotas per container. The result is logical partitioning that does not depend on hardware support: a single H100 can be split into ten fractional units of 8 GB each, each running a separate workload in isolation. Because the interception happens at the API layer, the same mechanism works across NVIDIA hardware of all generations, AMD Instinct GPUs, Intel Gaudi accelerators, and the domestic Korean NPUs \u2014 FuriosaAI RNGD and Rebellions ATOM+ \u2014 all managed through a Hardware Abstraction Layer (HAL) that presents a unified interface regardless of the underlying chip, as documented in Lablup’s FuriosaAI RNGD whitepaper<\/a>.<\/p>\n

This vendor-agnostic architecture is not merely a differentiator for Lablup’s sales pitch. It has direct implications for Korean AI infrastructure policy. The country’s national AI computing programs \u2014 from the \u20a92.08 trillion (~$1.44 billion) cluster program managed by Samsung SDS, Naver Cloud, and Elice Group, to the FuriosaAI RNGD deployments on Samsung Cloud Platform \u2014 explicitly mix NVIDIA GPUs with domestic Korean NPUs<\/a>. A management layer that handles both from a single control plane without requiring hardware-specific partitioning tools for each is structurally different from a Kubernetes plugin that works well only on NVIDIA hardware.<\/p>\n

The less-than-5% overhead figure Lablup cites for fGPU deserves a note. Software API interception adds latency compared to bare-metal GPU access, and the precise overhead depends on workload memory access patterns and data transfer sizes. The sub-5% figure is Lablup’s own benchmark claim; no third-party independent audit of this specific number has been published. For workloads that are primarily compute-bound rather than memory-bandwidth-bound \u2014 the majority of LLM training and inference tasks \u2014 the overhead is generally low for API interception approaches, but the specific figure remains self-reported.<\/p>\n

Sokovan: What Replaces Slurm in a National AI Cluster<\/p>\n

On top of the GPU virtualization layer sits Sokovan, Lablup’s proprietary workload orchestrator. The relevant comparison for understanding Sokovan is not Kubernetes but Slurm \u2014 the Simple Linux Utility for Resource Management<\/a>, which is the standard job scheduler at most supercomputers and HPC research clusters worldwide. Slurm is designed for batch job submission in homogeneous HPC environments: a researcher queues a job, requests a number of nodes, and waits for resources to become available.<\/p>\n

Sokovan was designed around the different demands of AI workloads: mixed batch training jobs alongside real-time inference serving, gang scheduling requirements for multi-node distributed training (where all nodes in a job must start simultaneously or not at all), and dynamic reallocation when researchers finish early or abandon sessions. The architecture uses a manager-agent dual-layer system<\/a> \u2014 the manager layer enforces scheduling policy, the agent layer handles per-node execution \u2014 and applies Dominant Resource Fairness scheduling to prevent any single team or project from monopolizing shared GPU capacity.<\/p>\n

In practice, the national AI context amplifies Sokovan’s importance. Lablup joined the Upstage consortium for the Korean government’s Sovereign AI Foundation Model project as infrastructure partner<\/a>. The consortium \u2014 the only one composed entirely of startups to make it through the Phase 1 evaluation \u2014 used Backend.AI to orchestrate hundreds of GPUs across distributed nodes while training a 102-billion-parameter model. Lablup’s materials state that Backend.AI achieved 47% faster failure recovery during those training runs compared to prior approaches.<\/p>\n

The 120-Plus Sites That Form the Equity Story<\/p>\n

The customer base Lablup is bringing to public markets spans five verticals that are not typical software startup territory. Defense (Republic of Korea Navy, LIG Defense & Aerospace), banking (Shinhan Bank, Bank of Korea), healthcare (Samsung Medical Center, Chonnam National University Hospital, Health Insurance Review and Assessment Service), and government (Gyeonggi Province) represent categories where air-gapped \u2014 fully offline \u2014 GPU cluster management is not optional. Backend.AI’s support for air-gapped deployment<\/a>, where installation, model deployment, and software updates all operate without internet connectivity, is a technical requirement for these sectors rather than a feature preference.<\/p>\n

The international footprint is more modest. Alongside Korean clients, Lablup has deployments at the University of Southern California’s Center for Advanced Research Computing, where Backend.AI manages shared GPU access for a multi-department research community. The company also has a US office in San Jose and has attended SuperComputing conferences (SC25 in late 2025) and NVIDIA GTC, including a strategic collaboration with VAST Data on Korea’s Sovereign AI infrastructure<\/a>. The global GPU orchestration market is dominated by US-based players \u2014 Run:ai (acquired by NVIDIA in 2024), Kubernetes-native tools from major cloud providers, and legacy HPC tools like Slurm. Lablup’s path into international markets from a KOSDAQ base is a legitimate open question that will require the company’s prospectus to answer.<\/p>\n

Lablup also received certification under NVIDIA’s DGX-Ready Software program<\/a>, which validates that a software platform has been tested and confirmed compatible with NVIDIA DGX enterprise AI systems. The company claims it was the first Asia-Pacific platform to receive this designation, in 2021; that specific priority claim comes from Lablup’s marketing materials and has not been independently verified through NVIDIA’s certification registry.<\/p>\n

Korea’s IPO Market Is Having Its Worst Year in Six \u2014 That Context Matters<\/p>\n

Lablup’s preliminary approval was the fastest among general companies in 2026, but a “fastest” in a year of very few listings is context-dependent. The Korea Exchange issued only 30 listings total across KOSPI and KOSDAQ through July 31, 2026 \u2014 the fewest in six years, matching levels last seen during the pandemic in 2020<\/a>. Overlapping listing regulations, revised institutional lockup rules, and a KOSDAQ index down sharply from its highs have collectively deterred companies from listing<\/a>. Thirteen of the 17 companies that did list in the first half of 2026 were trading below their offer prices by mid-year.<\/p>\n

Lablup’s record is genuine \u2014 the Korea Exchange’s preliminary review process still involves independent technology evaluation by NICE D&B and Ecredible (both awarded A grades in Lablup’s case), plus a full financial and governance review by the underwriter. A two-month turnaround from May 21 filing to July 30 approval is objectively fast in a system that typically takes longer. But the market into which Lablup must sell shares is a 2.7%-down, circuit-breaker-prone KOSDAQ that has not been kind to recent technology listings.<\/p>\n

The earlier KOSDAQ AI software listing that drew significant retail investor attention \u2014 MakinaRocks<\/a>, which attracted roughly \u20a914 trillion (approximately $9.7 billion) in retail subscription deposits \u2014 was a different category of company: vertical industrial AI for manufacturing and defense, not horizontal AI infrastructure software. Whether the investor appetite MakinaRocks demonstrated for Korean AI software companies will extend to a company whose product is the compute management layer beneath other companies’ AI applications is a central open question for Lablup’s bookbuilding.<\/p>\n

What the IPO Proceeds Will Fund<\/p>\n

Lablup plans to deploy IPO proceeds across three priorities: R&D advancement, global market expansion, and infrastructure scaling<\/a>. The geographic expansion targets are specific: North America, Japan, Southeast Asia, Europe, and the Middle East.<\/p>\n

The global push is rational given where sovereign AI policy is going. Every government now building a national AI computing program \u2014 Japan’s announced investment of \u00a5100 trillion (~$609 billion) in physical AI, the UAE’s AI factory ambitions, the European AI Act’s compute sovereignty provisions \u2014 is simultaneously building a need for the management software that makes national GPU clusters economically viable. A platform that already handles multi-vendor hardware from a single control plane, that runs air-gapped in classified environments, and that has an NVIDIA DGX-Ready Software certification starts that international pitch from a stronger position than a pure software startup with no validated enterprise deployments.<\/p>\n

The path from preliminary approval to an actual KOSDAQ listing still runs through securities registration, institutional bookbuilding, and public subscription. None of those steps is guaranteed, and the KOSDAQ’s current pressure \u2014 the same 2.7% single-session drop that framed Wednesday’s announcement \u2014 will shape both the timing and the pricing of what comes next.<\/p>\n

“We received KOSDAQ preliminary listing approval in the shortest time among general companies this year,” Lablup CEO Jeongkyu Shin said in a statement on Wednesday<\/a>, “a recognition of our AI infrastructure technology and commercialization capabilities. We will prepare the remaining offering procedures without delay and pursue a successful KOSDAQ listing.”<\/p>\n

Currency conversions in this article are approximate, based on exchange rates as of July 31, 2026.<\/p>\n

Frequently Asked QuestionsWhat is Backend.AI and how does it differ from standard Kubernetes GPU scheduling?<\/p>\n

Backend.AI is an AI infrastructure platform developed by Lablup Inc. that manages the full lifecycle of AI workloads \u2014 training, deployment, and inference \u2014 across GPU and accelerator clusters. Its core differentiation from Kubernetes-native GPU scheduling is the fGPU (fractional GPU) virtualization layer, which intercepts CUDA API calls in software<\/a> to partition a single physical GPU into precise fractional slices for concurrent workloads. Standard Kubernetes GPU scheduling allocates whole GPUs per container; GPU time-slicing in Kubernetes allows sharing but provides no memory isolation. NVIDIA MIG provides hardware-level isolation but is limited to specific datacenter GPU SKUs (A100, H100, B200) and cannot work with AMD, Intel, or domestic Korean NPU hardware. Backend.AI’s software-layer approach works across all of these hardware types from a single unified control plane.<\/p>\n

Does the KOSDAQ preliminary listing approval mean Lablup is now a public company?<\/p>\n

No. Preliminary listing approval from the Korea Exchange confirms that the company cleared the exchange’s initial technical and financial review \u2014 including A-grade evaluations from both NICE D&B and Ecredible. It does not complete the IPO. Lablup must still complete securities registration, institutional bookbuilding (where institutional investors set the final share price through demand-discovery), and a public subscription round before shares trade on the KOSDAQ. The timeline for those remaining steps has not been publicly disclosed<\/a>.<\/p>\n

What is the difference between fGPU and NVIDIA MIG, and when does each matter?<\/p>\n

NVIDIA MIG (Multi-Instance GPU) partitions a GPU at the hardware level \u2014 physically separating the streaming multiprocessors, L2 cache, memory controllers, and DRAM bus between instances. This gives each instance guaranteed performance and strong fault isolation. The constraint is that MIG only exists on specific NVIDIA datacenter GPUs<\/a> (A100, H100, B200) and creates fixed partition geometries (up to seven slices on an A100). Backend.AI’s fGPU intercepts CUDA API calls in software to enforce per-container memory and compute quotas, providing logical isolation without hardware requirements. fGPU works on all NVIDIA GPU generations including consumer RTX cards, and extends the same unified management interface to AMD, Intel, and Korean domestic NPUs that MIG cannot reach. For organizations running homogeneous NVIDIA datacenter clusters, MIG may offer stronger isolation guarantees; for organizations running mixed-vendor accelerator environments \u2014 which is increasingly the case in Korea’s national AI programs \u2014 software-layer virtualization is the only technically viable path.<\/p>\n

How does the 2026 KOSDAQ market environment affect Lablup’s IPO outlook?<\/p>\n

The 2026 KOSDAQ IPO market has been notably weak: only 30 total listings on KOSPI and KOSDAQ through July 31, the fewest in six years. Of 17 companies that listed in the first half of 2026, thirteen were trading below their offer prices by mid-year. Capital has rotated toward semiconductor giants (SK Hynix, Samsung Electronics) rather than smaller technology companies. Lablup’s preliminary approval speed \u2014 two months from May 21 filing to July 30 approval \u2014 is a genuine record for 2026 general companies, but the shares will ultimately be priced during bookbuilding against a KOSDAQ that fell 2.7% on the day the approval was announced<\/a>. The final IPO valuation will depend on investor appetite for AI infrastructure software compared to the AI hardware companies that have dominated Korean equity flows this year.<\/p>\n","protected":false},"excerpt":{"rendered":"On the same morning the KOSDAQ lost 2.7%, South Korean AI infrastructure software company Lablup Inc. received word…\n","protected":false},"author":2,"featured_media":105880,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[18],"tags":[1366,54601,54603,336,54602,872,2295,127,276,5766],"class_list":["post-105879","post","type-post","status-publish","format-standard","has-post-thumbnail","category-samsung-electronics","tag-ai-infrastructure","tag-backend-ai","tag-gpu-virtualization","tag-kosdaq","tag-kosdaq-ipo-2026","tag-lablup","tag-nvidia","tag-samsung","tag-samsung-electronics","tag-south-korea-ai"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/posts\/105879","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/comments?post=105879"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/posts\/105879\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/media\/105880"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/media?parent=105879"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/categories?post=105879"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/tags?post=105879"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}