Alluxio delivers sub-millisecond AI data platform & 50% growth

Alluxio has reported its results for the second quarter of its 2026 fiscal year, highlighting new customer acquisitions and performance benchmarks for its AI data platform.

The company released Alluxio Enterprise AI 3.7 during the quarter, described as providing sub-millisecond latency for artificial intelligence workloads accessing cloud storage. This update comes as part of Alluxio's drive to address latency and throughput challenges faced by organisations scaling their AI initiatives.

Product updates

Alluxio Enterprise AI 3.7 introduces several features aimed at improving AI data infrastructure. A distributed, transparent caching layer has been deployed to lower latency to sub-millisecond levels when retrieving AI data from cloud storage. The company has stated this can deliver up to 45 times lower latency than S3 Standard and five times lower than S3 Express One Zone. Additionally, the platform is reported to achieve throughput rates of up to 11.5 GiB/s (98.7 Gbps) per worker node, with throughput scaling linearly as more nodes are added.

The updated Distributed Cache Preloader in version 3.7 now supports parallel loading, which is designed to accelerate cache preloading by up to five times and improve AI training and cold start inference speeds. Furthermore, the new version includes granular role-based access control (RBAC) for S3 access, allowing integration with identity providers such as OIDC/OAuth 2.0 and Apache Ranger to manage user authentication, authorisation and permissions on cached data.

Customer growth

Alluxio reported over 50% customer growth in the first half of 2025 compared to the previous period. Sectors including technology, finance, eCommerce and media have implemented the platform to increase AI model training throughput, streamline feature store access, and reduce inference latencies. The company noted that Salesforce, Dyna Robotics and Geely were among its new customers during this period.

Usage of Alluxio's platform has also expanded across hybrid and multi-cloud environments, reflecting a growing organisational demand for high-performance, low-latency AI data infrastructure.

"This was a phenomenal quarter for Alluxio, and I couldn't be prouder of what the team has achieved," said Haoyuan Li, Founder and CEO, Alluxio. "With Alluxio Enterprise AI 3.7, we've eliminated one of the most stubborn bottlenecks in AI infrastructure, cloud storage performance. By combining sub-millisecond latency with our industry-leading, throughput-maximizing distributed caching technology, we're delivering even greater value to our customers building and serving AI models at scale. The strong customer momentum and outstanding MLPerf benchmark results further reinforce Alluxio's critical role in the AI infrastructure stack."

Benchmarking results

Alluxio has released its MLPerf Storage v2.0 benchmark results, a standardised set of performance tests for AI infrastructure. In ResNet50 training using 128 accelerators, the system achieved a throughput of 24.14 GiB/s with 99.57% GPU utilisation, scaling linearly with the number of clients and workers. The 3D-Unet model test with eight accelerators achieved 23.16 GiB/s throughput and 99.02% GPU utilisation. In the CosmoFlow test, eight accelerators reached 4.31 GiB/s and 74.97% utilisation, nearly doubling in performance when the number of clients was increased.

The company also reported results for large language model (LLM) checkpointing: the Llama3-8B model achieved read and write speeds of 4.29 GiB/s and 4.54 GiB/s, with read and write durations of 24.44 seconds and 23.14 seconds, respectively. For the larger Llama3-70B model, the platform delivered 33.29 GiB/s read and 36.67 GiB/s write performance, with read and write operations completing in 27.39 seconds and 24.86 seconds.

"The MLPerf Storage v2.0 results validate the core value of our architecture: keeping GPUs fed with data at the speed they require," Li added. "High GPU utilization translates directly into faster training times, better throughput for large models, and higher ROI on infrastructure investments. These benchmarks show that Alluxio can deliver performance at scale, across diverse workloads, without compromising flexibility in hybrid and multi-cloud environments."

Alluxio Enterprise AI version 3.7 is now available for customers and organisations aiming to address data access and performance issues in AI and machine learning applications.