The most contentious roadblock to running AI workload is not just the efficiency and high demand for GPUs but also largely depends on optimising data storage. In an exclusive interview with AIM at its India R&D centre in Bangalore, John Colgrove, the Founder of Pure Storage, shared his perspectives on how it is helping in solving these concerns, alongside ensuring fast, accessible, and reliable data storage, leading to better resource utilisation and reducing the strain on GPUs. 

“We ensure compute efficiency through scalable storage solutions that ensure data is accessible, manageable, and processed seamlessly by GPUs—through our hardware portfolio, which is GPU Direct Storage (GDS) ready and has AI-Ready Infrastructure (AIRI),” said Colgrove.

Pure Storage began its journey in California’s Bay Area in 2009, rapidly expanding to global R&D centres in Prague and Bengaluru. It expanded to India right after Covid, in June of 2022, given the opportunities and talent.

 “We wanted to open a site where we could hire top-notch engineers and give them challenging and difficult work to do,” said Colgrove about the India operations.

Competition

While it competes with major providers like DigitalOcean in the broader storage-infrastructure market. The competition is more intense in the all-flash storage market, where Pure Storage faces giants like IBM, Hewlett Packard Enterprise (HPE), NetApp, and Dell EMC.

However, Colgrove is confident in their product, describing its 75 terabytes direct flash module as a high-performance per terabyte solution. “We have by miles the best product; the direct flash gives us a huge advantage over anyone selling hard drives,” he stated.

He also focused on the importance of power efficiency in data centres—explaining its objective to create storage solutions that are both dense and power-efficient. “We want the storage footprint to be as dense and as power-efficient as possible,” he said, acknowledging the crucial role of power management in modern data storage.

Moreover, it plans to ship a 150-terabyte version of the direct flash module later in the year. This capacity significantly surpasses traditional hard drives, with Colgrove saying, “Seagate announced like we’re going to ship 30 this year. They’re going to ship 30 This year, we’re going to ship 150 This year.”

He emphasised their direct flash drives’ ecological and economic benefits, noting their power usage and electronic waste efficiency. “There’s less e-waste. It uses less power,” Colgrove explained, highlighting the environmental advantages of its technology. He also pointed out that these drives are more cost-effective in terms of total ownership cost than traditional disks.

Colgrove also talked about the ‘Evergreen’ model employed by Pure Storage, a key differentiator in its product strategy. This approach allows the arrays to be upgraded while they’re online without data migration, making them seem new even after ten years of operation. 

“The magic isn’t that they’re ten years old and still functioning; the magic is they’re ten years old, but they look like a brand new array ready for another ten years,” he said. This model eliminates the need for costly and risky data migrations common with traditional data storage solutions.

With a market share of about 6%, the company reported a revenue of $762.8 million for the third quarter of 2024. The previous fiscal year saw its revenue at $2.8 billion, a 26% increase year-over-year, with a free cash flow of $609.1 million for fiscal year 2023.

Collaborations & Partnerships 

Central to Pure Storage’s strategy is its collaboration with NVIDIA in developing AI-Ready Infrastructure (AIRI) solutions. “NVIDIA is an interesting one, and we made the first AI ready infrastructure back in 2017,” Colgrove said.

The FlashBlade hardware portfolio is set to fully support GPU Direct Storage (GDS) with upcoming software updates, enhancing this collaboration. 

“The AIRI//S solution, co-developed with NVIDIA and built on the NVIDIA DGX system, is designed to provide significant performance advancements that are non-disruptive, allowing for continuous innovation and adaptability to the evolving needs of AI,” added Colgrove.

The AIRI//S architecture is comprehensive and scalable, integrating FlashBlade//S, NVIDIA DGX systems, and NVIDIA networking. This integration creates an optimised setup for multi-dimensional GPU performance, ensuring GPUs are continuously engaged in AI workloads of any scale. This setup simplifies the development of modern AI environments for enterprises.

Additionally, the company plays a crucial role at various stages of AI development, starting from the initial data collection and storage phase—to the crucial phase of Data Access and Management, where its AI Storage Solutions automate the access to diverse data sources and storage resources, reducing the model training time from months to days, streamlining the entire process of data curation, training, and inference. 

Pure Storage has established partnerships with most major hyperscalers. “We are working with all of the hyperscalers very closely. I think we work with almost everybody other than Google Cloud at this point,” Ajeya Motaganahalli, VP of Engineering & Managing Director, Pure Storage India, told AIM, citing Portworx, a Kubernetes storage overlay.

He said this product functions as a container in the cloud and is compatible with various public cloud platforms, including GCP, Microsoft Azure, and AWS. It connects to underlying storage, whether on-premises or in the cloud, creating a unified namespace.

What’s next? 

Going ahead, the Founder said the company’s ambition is to be a leader in the inference market while continuing to support the AI training market. “We want to be leading that, obviously, and we want to continue to support the training market as well,” he stated, saying that it aims to address both the training and application phases of AI development.

Discussing the limitations of traditional hard drives in AI, Colgrove pointed out their inadequacy for AI applications due to their slow access capacity. “Hard drives don’t have the access capacity. They’re so slow on I/O ops there… AI can’t run on top of hard drives,” he explained, highlighting the need for faster, more efficient storage solutions.

Colgrove further elaborated on the role of data in generative AI, using retail and automotive industries as examples. In retail, AI can analyse small amounts of data from video cameras in stores to understand shopper behaviour and then apply these learnings to a larger dataset. 

Similarly, AI can learn from a subset of telemetry data from test vehicles in the automotive sector and then apply these insights to millions of cars. “We’re going to learn stuff by studying a subset of my data very carefully, and that’s what the training but I’m then going to apply it to millions of things,” he said, emphasising the scale at which AI operates and the consequent data requirements.