Acryl leadership visited the HiPerGator racks at the University of Florida East Campus Data Center on August 5.
Acryl executives engaged in discussions with co-founder Malachowsky
KOSDAQ-traded AI infrastructure firm demonstrates 1,272-GPU validation at UKC 2026, meets NVIDIA co-founder Malachowsky, and reports ₩21.2 billion in order backlog
Acryl Inc (KOSDAQ:0007C0)
Competition in AI infrastructure is moving from how many GPUs a company has secured to how efficiently and reliably it operates the compute it holds
”— Jin Park, CEO of Acryl IncSEOUL, SEOUL, SOUTH KOREA, September 5, 2026 /EINPresswire.com/ — SEOUL, South Korea — Acryl Inc. (KOSDAQ: 0007C0), a software company specializing in AI infrastructure, has initiated its North American push with GPUBASE, its GPU operations platform, fueled by first-half revenue that climbed 150% year-over-year to KRW 12 billion (US$8.6 million) and an order backlog reaching KRW 21.2 billion (US$15.3 million) as of the end of June.
Acryl participated in the 39th US-Korea Conference on Science and Technology (UKC 2026), held from August 5 to 8 in Orlando, Florida, as the sole Korean company featured on the panel of the official Korea-US AI cooperation symposium, and delivered the lead presentation at the official sponsor forum. During the event, the company visited HiPerGator, the supercomputing facility at the University of Florida, to explore collaboration with its operations team, convened a roundtable with SelectFlorida (the state’s economic development organization), and on August 7 met NVIDIA co-founder Chris A. Malachowsky to present the company and GPUBASE.
A comprehensive platform for heterogeneous GPU clusters
GPUBASE aggregates GPUs and NPUs from various manufacturers into a unified resource. It divides a single GPU to allow multiple jobs to share it and alleviates network bottlenecks, thereby boosting the effective performance of already purchased hardware. The commercial offering comprises three components: Booster, which minimizes data-transfer bottlenecks between GPUs; Control Tower, which manages GPU, server, storage, and network resources as a single pool; and Monitor Tower, which provides real-time visibility into infrastructure and job status. The networking technology is built on research by Acryl CTO Yeom Ik-joon, a professor of software at Sungkyunkwan University, including PeRF, which distributes data-transfer paths, and UL-MPRDMA, which simultaneously utilizes multiple paths.
Acryl’s entry point in the North American market targets heterogeneous accelerator environments rather than single-vendor setups, and Ethernet-based clusters rather than InfiniBand. Universities and public research institutions are the initial focus; references obtained there will be leveraged to expand into data center and cloud operators.
1,272 GPUs validated: utilization exceeds 90%, training time reduced by up to 96%
The evidence Acryl presented in North America is the K-Scale Evaluation, a performance verification launched in June across Amazon Web Services and other global cloud environments, whose first phase—the Horizontal Phase—concluded in July. In independent tests conducted on 1,272 GPUs of seven heterogeneous types spanning the three major global cloud environments, each with a distinct network architecture, GPU utilization rose above 90%, training time under heavy load dropped by up to 96% (a speed improvement of up to 24 times), and job queue wait time decreased by up to 93%. NCCL all-reduce communication bandwidth improved by 141% to 2,375% depending on load level, and fault detection and recovery completed within five minutes. The next step involves a performance test at a single-cluster scale of 1,000 GPUs, followed by a KOLAS-accredited test report.
Cost projections are presented as the company’s own estimates. Based on a compilation from published papers and commercial cases, network total cost of ownership for a 512-GPU NVIDIA H100 cluster amounts to 48% of the InfiniBand figure, saving roughly US$2.2 million. Raising average GPU utilization from 55% to about 85% would reduce the GPU purchase cost required for the same throughput by approximately 35%.
US$4 million cumulative supply to Hancom’s modular data center, KRW 6.7 billion Ethernet fabric project
Commercial supply continues in Korea. In July, Acryl signed a contract to provide GPUBASE to the modular data center (MDC) construction project led by Hancom. Following the basic tier, an advanced-tier contract valued at KRW 3.3 billion (US$2.4 million) was added on July 29, bringing cumulative supply for the project to KRW 5.5 billion (US$4 million). The advanced tier operates on NVIDIA H200 NVL GPU servers and direct-liquid-cooled racks, with delivery scheduled for completion in the third quarter. A July 24 memorandum of understanding with Megazone Cloud extends coverage to large AI data centers and distributed AI infrastructure, and earlier agreements with MangoBoost and FuriosaAI broadened the verification scope to Korean-made DPUs and NPUs.
In June, Acryl was selected as lead organization for a project commissioned by Korea’s Institute of Information & Communications Technology Planning & Evaluation (IITP) to develop an Ethernet-based GPU cluster network fabric. With a total budget of approximately KRW 6.7 billion (US$4.8 million), the project runs through December 2028 jointly with Yonsei University, Sungkyunkwan University, and Ajou University. Its objective is to replace InfiniBand, NVIDIA’s proprietary networking technology, with open Ethernet.
First-half results
Acryl reported on August 18 consolidated first-half revenue of KRW 12 billion, up 150% and equivalent to 90% of its full-year 2025 revenue. The operating loss ratio narrowed to 19% from 44% a year earlier as the company continued upfront investment in technology and computing infrastructure. The order backlog at the end of June was 49% higher than at the end of March. Medical AI contributed KRW 8.5 billion (US$6.1 million) in new orders this year, driven by the combination of NADIA-ANE, an AI-native electronic medical record platform, and ALLM.H, a healthcare foundation model deployed on hospitals’ own GPU servers. On September 1, the company reported that its 31-billion-parameter successor model, ALLM.H2 FRANK, answered all 435 physician-examination questions on the KorMedMCQA benchmark.
Jin Park, CEO of Acryl, remarked: “Competition in AI infrastructure is moving from how many GPUs a company has secured to how efficiently and reliably it operates the compute it holds. We will make joint verification and business cooperation concrete on the basis of the operating problems North American universities, research institutes and data centers are actually facing.”
About Acryl Inc.
Acryl is an AX (AI transformation) infrastructure software company founded in March 2011 and listed on the KOSDAQ in December 2025. Its main products are the GPU operations platform GPUBASE; the integrated AI platform Jonathan; the industry-specific foundation model family ALLM; and the healthcare AI solution NADIA. Headquarters: 704 Seolleung-ro, Gangnam-gu, Seoul, Korea. www.acryl.ai
Won-dollar conversions at US$1 = KRW 1,390. Cost-saving figures are Acryl’s own estimates.
Jerry (Jaehyuck) Yang
Acryl Inc
+82 10-9949-0296
jerry2@acryl.ai
