Cloudian's AI-ready storage tackles data demands

Artificial intelligence is rapidly transforming how businesses manage and access data, exposing significant limitations in traditional storage systems. Designed for simpler, sequential commands from a limited number of users, these older architectures struggle to keep pace with modern AI, which demands continuous, parallel access to vast datasets by millions of agents. The inherent complexity and multi-tiered structure of legacy systems create bottlenecks, slowing the flow of critical data to the powerful graphical processing units (GPUs) that power AI computations.

Addressing this challenge, Cloudian, co-founded by MIT alumni Michael Tso and Hiroshi Ohta, has developed a scalable storage solution specifically engineered for the AI era. Their system streamlines data flow between storage and AI models by applying parallel computing directly to data storage. This innovative approach consolidates AI functions and data onto a single platform, enabling direct, high-speed transfers between storage and both GPUs and CPUs, thereby reducing the complexity and latency that hinder AI performance.

Cloudian's integrated storage-computing platform simplifies the development of commercial-scale AI tools, providing businesses with a robust data foundation capable of supporting the exponential growth of AI. Michael Tso emphasizes the fundamental role of data in AI advancement: "One of the things people miss about AI is that it’s all about the data. You can’t get a 10 percent improvement in AI performance with 10 percent more data or even 10 times more data — you need 1,000 times more data." He highlights the industry's shift towards storing data in easily manageable ways that allow for computations to be embedded and executed as data arrives, eliminating the need to move large datasets.

Tso's journey to co-founding Cloudian is deeply rooted in his foundational work at MIT. As an undergraduate in the 1990s, he delved into parallel computing under Professor William Dally and Associate Professor Greg Papadopoulos. His graduate studies with computing pioneer David Clark focused on disconnected and intermittent networking operations for large-scale distributed systems, a concept Tso notes remains central to his work today.

After MIT, Tso contributed to data synchronization algorithms for BlackBerry at Intel's Architecture Lab and developed specifications for Nokia that catalyzed the ringtone download industry. He then joined Inktomi, a startup co-founded by MIT alumnus Eric Brewer, which pioneered search and web content distribution. In 2001, Tso co-founded Gemini Mobile Technologies, which built some of the world's largest mobile messaging systems to manage the explosive data growth from camera phones.

Observing that data generation was outpacing network speeds in the late 2000s, Tso recognized a fundamental shift was needed. He concluded that "data has its own gravity," making it impractical and costly to constantly move it to centralized cloud systems. This insight led to the pivot towards a distributed cloud model, where computing power is brought closer to the data, rather than the other way around. Cloudian officially launched from Gemini Mobile Technologies in 2012, initially focusing on scalable, distributed, cloud-compatible data storage, though Tso admits they didn't initially foresee AI as the ultimate use case for edge data.

Tso sees striking parallels between his early research at MIT and Cloudian’s current endeavors. He points out that the challenges of disconnected networks, which he explored with David Clark, are now integral to every edge computing scenario. Similarly, Professor Dally's work on fast, scalable interconnects is evident in modern NVIDIA chip architecture, while his collaborations with Professor Papadopoulos on accelerating application software with parallel computing hardware without extensive rewriting directly inform Cloudian's efforts to optimize data flow for NVIDIA GPUs.

Cloudian's platform leverages an object storage architecture, where all types of data – from documents to sensor readings – are stored as unique objects with metadata. This flat-file structure is highly effective for managing the massive, unstructured datasets prevalent in AI applications. Historically, however, object storage faced limitations in directly feeding data to AI models, often requiring data to be copied into computer memory, leading to latency and energy inefficiencies.

In a significant advancement this July, Cloudian announced an extension to its object storage system: a vector database. This innovation allows data to be stored in a format immediately usable by AI models. As data is ingested, Cloudian computes its vector form in real-time, powering AI tools such as recommender engines, search functions, and AI assistants. The company also unveiled a strategic partnership with NVIDIA, enabling its storage system to work directly with NVIDIA's GPUs, promising faster AI operations and reduced computing costs. Tso notes that NVIDIA initiated the collaboration, recognizing that GPUs require a constant, high-speed data supply to operate efficiently. This partnership underscores the growing understanding that it's more efficient to bring AI processing to the data rather than moving colossal datasets. Cloudian's systems embed many AI functions, allowing for data pre- and post-processing near where the data is collected and stored.

Cloudian currently assists approximately 1,000 companies worldwide in extracting greater value from their data. Its diverse client base includes large manufacturers, financial service providers, healthcare organizations, and government agencies. For instance, a major automaker utilizes Cloudian's platform with AI to predict maintenance needs for its manufacturing robots. Cloudian also supports critical initiatives like storing research articles and patents for the National Library of Medicine, and DNA sequences of tumors for the National Cancer Database – rich datasets that AI models can process to accelerate medical research and discovery.

Tso emphasizes the transformative impact of GPUs, which have shattered traditional computing growth rates by parallelizing operations and allowing for networked configurations. This unprecedented scale is pushing AI to new levels of intelligence. However, to fully harness this power, GPUs require data to be fed at the same speed they compute. Tso concludes that the only way to achieve this is by "getting rid of all the layers between them and your data," a principle at the heart of Cloudian's innovation.