Artificial intelligence is reshaping how businesses manage and access their data. Unlike traditional storage built for simple tasks, modern AI requires millions of agents to access and process massive data simultaneously. Older systems have complex layers that delay AI by forcing data through multiple stages before reaching the GPUs.
To meet AI’s demands, Cloudian—founded by Michael Tso ’93, SM ’93, and Hiroshi Ohta—developed a scalable storage solution. Their system unifies storage and processing with parallel computing, streamlining data flow to AI models. This design enables fast, direct data transfers between storage, GPUs, and CPUs, reducing delays and complexity.
By merging storage and computation, Cloudian enables scalable AI development and keeps pace with AI’s rapid growth.
“People often overlook that AI is fundamentally about data,” Tso explains. “A small increase in data won’t yield big performance gains—you need data on the scale of a thousandfold. The key is managing data efficiently and embedding computation in storage to process it during ingestion without moving it. That’s the direction the industry is heading.”
Transitioning from MIT to the Marketplace
As an MIT undergrad in the 1990s, Tso learned parallel computing from Professor William Dally. He also collaborated on related research with Associate Professor Greg Papadopoulos.
“It was an exciting period,” Tso recalls. “Most universities had one supercomputing project; MIT had four.”
As a grad student, Tso worked with David Clark, an internet pioneer who helped develop TCP for reliable data transfer.
“As a grad student, I focused on networking for large-scale distributed systems that were often disconnected or unstable,” Tso explains. “It’s funny — three decades later, I’m still working on the same problems.”
From Mobile Messaging to Cloud Innovation: Tso’s Journey Through the Early Internet Era
After graduating, Tso joined Intel’s Architecture Lab and developed data sync algorithms later used by Blackberry. He also created technical standards for Nokia that helped launch the ringtone download industry. He later joined Inktomi, a startup co-founded by Eric Brewer, leading in search and web content distribution.
In 2001, Tso co-founded Gemini Mobile Technologies with Joseph Norton ’93, SM ’93, and others. The company built some of the world’s largest mobile messaging systems to support the explosive growth of camera phone data. As cloud computing emerged in the late 2000s, offering scalable virtual servers, Tso observed that data volume was increasing much faster than network speeds — prompting him to steer the company in a new direction.
“Data is now being generated in many different locations, and that data tends to stay where it is — moving it costs both time and money,” Tso explains. “That’s why the future lies in a distributed cloud model that extends out to edge devices and servers. Instead of transferring data to the cloud, we need to bring cloud capabilities to where the data resides.”
In 2012, Tso officially launched Cloudian from the foundation of Gemini Mobile Technologies, shifting focus to provide scalable, distributed storage solutions that are compatible with cloud infrastructure.
“When we first started the company, we didn’t realize AI would become the primary driver for edge data storage,” he says.
Although Tso’s research at MIT dates back over 20 years, he sees a direct link between that early work and current developments in the tech industry.
MIT Lessons Resurface in Today’s AI and Edge Computing Challenges
“It feels like my whole career is coming full circle,” he says. “David Clark and I were focused on managing networks with intermittent connections — which is exactly what edge computing involves today. Professor Dally was developing high-speed, scalable interconnects, and now he’s NVIDIA’s chief scientist, where his influence is evident in their chip designs and interchip communication. With Professor Papadopoulos, I worked on speeding up application software using parallel computing hardware without needing to rewrite the code — the same challenge we’re tackling with NVIDIA now. It’s remarkable how everything I studied at MIT is resurfacing in today’s tech landscape.”
Cloudian’s platform now operates using an object storage architecture, where various types of data — such as documents, videos, and sensor outputs — are stored as individual objects along with associated metadata. This flat file structure allows for efficient management of vast amounts of unstructured data, making it well-suited for AI applications. However, traditional object storage has a limitation: it typically requires data to be copied into a computer’s memory before it can be processed by AI models, leading to delays and higher energy consumption.
Cloudian Integrates Vector Database and Partners with NVIDIA
In July, Cloudian introduced a major upgrade by integrating a vector database into its object storage platform. This enhancement allows data to be stored in a vector format that AI models can use immediately. As data is ingested, Cloudian’s system automatically computes its vector representation in real time, enabling advanced AI functions like recommendation engines, intelligent search, and virtual assistants. Cloudian also announced a collaboration with NVIDIA, allowing its storage system to interface directly with NVIDIA’s GPUs. This integration speeds up AI processing and helps reduce computing expenses.
“A year and a half ago, NVIDIA approached us because GPUs need a constant data stream,” Tso says. “People are realizing it’s more efficient to bring AI to the data than move large datasets.” Our storage now has embedded AI to handle pre- and postprocessing at the data source.”
Storage Designed for AI
Cloudian supports roughly 1,000 organizations worldwide — including major manufacturers, financial institutions, healthcare providers, and government bodies — in extracting greater value from their data.
For example, one major automotive company uses Cloudian’s storage platform alongside AI to predict maintenance needs for its manufacturing robots. Cloudian also partners with the National Library of Medicine to store research articles and patents, and with the National Cancer Database to manage tumor DNA sequences — large, complex datasets that AI can analyze to uncover new insights or help develop treatments.
“GPUs have been a game changer,” says Tso. “While Moore’s Law predicts a doubling of computing power every two years, GPUs accelerate progress by enabling parallel processing across multiple chips. That scalability is pushing AI capabilities far beyond expectations. But to fully leverage GPUs, you need to supply data as fast as they can process it — and that means eliminating the traditional layers between storage and compute.”
Read the original article on: MIT
Read more: AI Aids Chemists in Creating More Durable Plastics