Powering AI with High-Performance Storage

B2 Cloud Storage was literally the only technical thing we used in training these models that didn’t crash the first time we tried it. We’re in an industry where everything fails, but Backblaze didn’t.

Dean Leitersdorf

Co-Founder and CEO, Decart

0->16PB

in 90 days

10x

more efficient than competition

$0

egress spend

Situation

Decart set out to disrupt the AI market on inference speed, but they faced one simple, yet confounding challenge: Storage performance and cost were going to limit growth. They were burning through free credits from a traditional cloud provider, and they needed storage that could scale from 100s of thousands of hours of video data to 100s of millions.

Solution

Decart chose Backblaze B2 Cloud Storage to help them build the most high-performance multi-cloud training infrastructure possible. They put Backblaze through its paces, testing a wide variety of usage patterns to confirm that they would have the reliably high performance, GPU interoperability, and efficiency they needed.

Result

With Backblaze, Decart had no problem transferring petabyte-scale datasets, and with pricing at one-fifth the cost of traditional cloud providers, Backblaze unlocked significant budget. But the true game changer was free egress up to three times the amount stored, allowing them to train their models on multiple different GPU clusters at the same time with zero additional cost.

The Decart AI platform provides an order of magnitude improvement in both training and inferencing of the largest generative models. They debuted their innovations via the first real-time generative AI open world model: Oasis.

1M hours of training data
1.5PB per day moved
75% reduction in AI operation costs

The Details

Disrupting AI on speed (and storage)

It’s an understatement to say AI is a growing, highly competitive market. According to Pitchbook, one out of every three VC dollars invested globally in 2024 went to an AI startup. How does any company differentiate themselves in an explosive burst of technological change, one where small startups with a big idea can find themselves toe-to-toe with the goliaths of tech, fighting for money, chips, talent, and even raw electrical power. And one where even those goliaths don’t always succeed?

‍

Decart, an AI research lab that developed a real-time AI open world model, understood the path to disrupting the market was to deliver an order of magnitude improvement in both training and inference, and to do that by building the most high-performance multi-cloud model training infrastructure possible.

What we really needed was a place where we could store an insane amount of data and, at the same time, download it to a few different GPU clusters around the world, and for all that to not cost an insane amount of money. That’s why we chose Backblaze.

Dean Leitersdorf, Co-Founder and CEO, Decart

The scaling dilemma

Co-founder and CEO, Dean Leitersdorf, had done the hard work to pave the path to disruption. The multi-cloud AI stack for training was dialed in and the models were performing 10x more efficiently than competitors such as OpenAI’s Sora. They just had one simple, but big, problem holding them back: The price and the logistics of moving and storing training data in a traditional cloud storage provider were going to limit their growth.

‍

They were burning through free data storage credits and had data spread across a range of other cloud providers and GPU clusters. Their training data needed to scale tenfold.

BYOG: Bring your own GPUs

Decart needed a storage solution that could handle that scale in three key areas:

Reliably high performance. Decart needed to know that when they got time on a cluster, they could move data in as fast as possible the second that they were able to.
GPU interoperability. They needed to be sure that no matter which storage platform they chose, it would work well with a multi-cluster training approach. Being able to shop jobs between different GPU clouds and disperse training was essential for Dean’s team.
Efficiency. Every dollar an AI startup spends on anything other than training time is a competitive disadvantage, so ensuring that storage costs were low without any surprise fees for data retention or download was key.

Petabyte-scale streaming

After conducting rigorous testing of Backblaze in a wide variety of usage patterns, it was clear to the team that they had found the storage foundation they needed. When it came time to start moving data from Backblaze to GPU clusters, they had no problem with transferring petabyte-scale datasets. The only minor challenge was ensuring that the compute provider’s pipe could take the volume of data streaming in.

‍

Free egress unlocks flexibility

Here’s how Backblaze delivered on performance, flexibility, and affordability for Decart:

Performance. They were impressed by the performance they achieved with Backblaze—where other providers would crash, Backblaze worked smoothly out of the box.
Price. With pricing at about one-fifth the cost of traditional cloud providers, Backblaze unlocked a significant amount of budget.
Free egress. The true game changer. Decart, for a number of reasons, trains their models on multiple different GPU clusters at the same time. With Backblaze, they can egress their full dataset to up to three training sites every month with zero additional cost.

With all the fundamentals working, Backblaze and Decart are now working together to find even more efficiency and optimization and truly stand up the best infrastructure for training AI models.