Acryl pushes GPU operations platform into North America

Acryl used UKC 2026 in Orlando to introduce GPUBASE to North American AI infrastructure buyers, highlight a 1,272-GPU verification, and meet with University of Florida and NVIDIA leadership. The move signals a push to win university, research and data center customers with software that improves GPU utilization instead of adding hardware.

Why it matters: - Acryl is trying to enter North America with software built to raise GPU efficiency, a growing priority as AI data center builds accelerate. - The company is targeting universities and public research institutions first, then plans to expand to data center and cloud operators. - Acryl is pitching a software approach that aims to improve performance from hardware already purchased, which could lower infrastructure spending.

What happened: - Acryl took part in the 39th US-Korea Conference on Science and Technology, or UKC 2026, held Aug. 5-8 in Orlando, Florida. - Acryl was the only Korean company on the panel of the official Korea-US AI cooperation symposium. - Acryl delivered the anchor presentation at the official sponsor forum. - During the conference, Acryl visited the University of Florida’s HiPerGator supercomputing facility to discuss cooperation with its operations team. - Acryl also held a roundtable with SelectFlorida on HPC operations cooperation, on-site verification, siting and incentives. - On Aug. 7, Acryl’s management met NVIDIA co-founder Chris A. Malachowsky after the plenary session and introduced the company and GPUBASE.

The details: - GPUBASE pools GPUs and NPUs from different manufacturers into a single resource. - GPUBASE also partitions a single GPU so multiple jobs can share it. - The platform is designed to reduce network bottlenecks and improve realized performance from existing equipment. - Acryl says its North American entry point is heterogeneous accelerator environments rather than single-vendor systems. - Acryl is also focusing on Ethernet-based clusters rather than InfiniBand-based clusters. - The company said it will use references from universities and public research institutions to expand into larger commercial infrastructure markets. - Ko Sam-seok, a distinguished professor at Dongguk University and member of Korea’s National AI Strategy Committee, said the center of gravity in AI competition is shifting from models to operating compute resources. - Ko said Korean software companies can enter local markets through verification with U.S. universities and research institutions on real operating problems. - Acryl presented results from a K-Scale Evaluation completed in the first half of 2026. - The verification covered 1,272 GPUs across seven heterogeneous types in three major global cloud environments with different network architectures. - GPU utilization rose above 90% in the tests. - Training time under high load fell by up to 96%. - Job queue wait time fell by up to 93%. - The evaluation checked network performance, GPU sharing and resource utilization, heterogeneous accelerator integration, scalability, storage integration and fault recovery. - NCCL all-reduce communication bandwidth improved by between 141% and 2,375%, depending on load level. - Fault detection and recovery completed within five minutes. - Acryl plans a 1,000-GPU single-cluster performance test next and will seek a KOLAS-accredited test report. - Acryl’s own estimates put network TCO on a 512-GPU NVIDIA H100 cluster at 48% of the InfiniBand figure, equal to about US$2.2 million in savings. - Acryl’s own estimates also say raising average GPU utilization from 55% to around 85% could cut GPU purchase costs by about 35% for the same throughput. - In July, Acryl signed a contract to supply GPUBASE to Hancom’s modular data center construction project. - On July 29, Acryl added an advanced-tier contract worth KRW 3.3 billion, or US$2.4 million. - Cumulative supply to the Hancom project reached KRW 5.5 billion, or US$4 million. - The advanced tier runs on NVIDIA H200 NVL GPU servers and direct-liquid-cooled racks. - Delivery is scheduled to finish in the third quarter. - On July 24, Acryl signed a memorandum of understanding with Megazone Cloud to expand coverage to large AI data centers and distributed AI infrastructure. - Earlier MOUs with MangoBoost in May and FuriosaAI in July expanded verification to Korean-made DPUs and NPUs. - In June, Acryl was selected as lead organization for an IITP project to develop an Ethernet-based GPU cluster network fabric. - The project has a total budget of about KRW 6.7 billion, or US$4.8 million, including KRW 5.5 billion in government funding. - The project runs through December 2028 with Yonsei University, Sungkyunkwan University and Ajou University. - The project aims to replace InfiniBand with open Ethernet.

Between the lines: - Acryl is framing infrastructure software as the next battleground in AI, not just access to more GPUs. - The company’s emphasis on verification with U.S. universities and research groups suggests a market-entry strategy built on technical proof, not brand recognition. - The push toward Ethernet and heterogeneous systems aligns with a broader industry search for lower-cost, less vendor-dependent AI infrastructure. - The North American effort is backed by commercial deals and government-funded work in Korea, which gives Acryl a reference base before deeper U.S. expansion.

What's next: - Acryl plans to use university and research references from North America to approach data center and cloud operators. - The company will pursue a 1,000-GPU test and KOLAS-accredited report to add third-party validation. - Delivery on the Hancom advanced-tier project is expected to finish in the third quarter. - The IITP Ethernet cluster project continues through December 2028.

The bottom line: - Acryl is betting that AI infrastructure buyers will pay for better GPU utilization, faster training and lower networking cost, not just more hardware. The company is now trying to prove that case in North America.