Eight AI Companies. One AI Storage Cloud.

A multicloud platform serving multimodal training datasets to frontier AI labs

Who they are. A large-scale web data acquisition company that operates a distributed network to collect public web content at petabyte scale. The company scrapes, cleans, and curates multimodal datasets, including text, video, image, and audio content, and sells enriched datasets to LLM developers and multimodal AI labs.

The storage problem. Storing and serving scraped multimodal data during peak training delivery cycles requires sustained throughput that most storage platforms can't handle. The data lifecycle requires hot storage for active training alongside an archive tier for data not currently in rotation. Their infrastructure was split across multiple cloud environments, including Azure, which couldn't economically sustain the scale or the throughput variability of their largest workloads.

The Wasabi solution. Compute clusters at third-party GPU provider Nebius pull directly from Wasabi buckets at high throughput during active training delivery windows, moving petabytes of data per day. Because Wasabi doesn't charge egress fees, moving that much data to compute every day costs nothing extra beyond the storage itself.

S3 compatibility meant no pipeline changes were required to make the switch. The company now stores roughly 150 petabytes in Wasabi, with another 20 petabytes in discussion to migrate off Azure. The platform has grown without storage becoming a limiting factor on delivery capacity.

A frontier AI lab building production-grade text-to-image editing models

Who they are. An AI lab building and licensing image generation models. The company develops both open-weight and proprietary models distributed via API and direct licensing. Enterprise contracts span major creative software platforms and consumer applications. The company competes at the frontier of image generation and has recently expanded into video generation.

The storage problem. The lab had accumulated large amounts of hot training data in Microsoft Azure, and the high costs and constraints on data movement had become unsustainable. As the team evaluated GPU providers outside the hyperscaler’s ecosystem, the egress fees they faced made the move impractical.

The Wasabi solution. The lab needed pricing that would scale gracefully beyond its existing hot training data. A rigorous proof of concept, including ingress and egress throughput testing from a third-party GPU compute cluster, confirmed no bandwidth limitations per account or bucket, and no throttling as workloads scale. The lab also cited Wasabi's responsiveness during the POC and contracting process as a differentiating factor.

Training data is now migrating from Azure to Wasabi Hot Cloud Storage with compute on third-party GPU infrastructure reading directly from Wasabi. The lab moved 50 petabytes of hot training data within the first three months of deployment and are now at 175 PBs with room to grow. The move eliminated hyperscaler dependency for primary training data and meaningfully improved storage economics.

A petabyte-scale shared data lake for robot foundation models

Who they are. A Japan-based non-profit building a Robot Data Ecosystem that acquires and integrates real-world operational data from retail, manufacturing, and logistics environments. The organization uses this data to enable cross-domain learning for foundation model development and real-world robot deployment. At the core of the ecosystem sits a petabyte-scale data hub that distributes raw data across GPU clusters and cloud environments.

The storage problem. The data hub continuously uploads dozens of terabytes and depends on extensive API activity — uploads, downloads, differential backup checks, and object listings — as its core usage pattern. Variable hyperscaler pricing due to burdensome API request fees didn't fit a non-profit budget. Beyond price predictability, the team required a platform that could guarantee integrity throughout the entire data lifecycle, as data loss anywhere would render it unusable for training.

The Wasabi solution. No transfer fees and no API fees made the difference, since high-frequency API operations are exactly how this organization uses storage day to day. A proof-of-concept validated Wasabi upload speed and confirmed the data remained unaltered through full upload, storage, and retrieval cycles. Wasabi has continued to perform reliably as usage scales dynamically. The organization projects savings of more than JPY 100 million compared to alternative providers and has stored more than 2.2 petabytes with Wasabi, with that figure still growing.