Jeronimo De Leon, Senior Product Manager of AI at Backblaze – Interview Series

Jeronimo De Leon is a seasoned product management leader with over 10 years of experience driving AI-driven innovation across enterprise and startup environments. Currently serving as Senior Product Manager, AI at Backblaze, he leads the development of AI/ML features, focuses on how Backblaze enhances the AI data lifecycle for customers’ MLOps architectures, and implements AI tools and agents to optimize internal operations.

Backblaze is a cloud-storage and backup company that provides unlimited, automatic computer backups for individuals and businesses, along with scalable object storage solutions for enterprise, media, and application workloads. Its services focus on affordability, data security, easy recovery, and seamless compatibility with existing systems.

You bring over a decade of experience in AI-driven product management—from working with LLMs at Intelas and RAG at Welcome.AI to launching Bloomberg’s chatbot and now leading AI efforts at Backblaze. How have these experiences shaped your view on the role of cloud storage in scaling AI/ML workflows?

Since starting on AI projects at IBM Watson, I’ve seen the pace of innovation dramatically accelerate. What used to take years to move from research to production now happens in months. However, the fundamental infrastructure challenges remain the same: where is the data, where do we store it, and how do we access it efficiently?

Before, the constraints were compute and models, but now we have an abundance of pretrained models and there are a lot of compute providers. Yet when starting a project back then, we typically had to begin with a data collecting and processing project, that’s still the same today. I consistently see organizations hitting the same bottleneck of consolidating data from disparate sources. The organizations that succeed are those that solve data accessibility early, creating a foundation that scales with their AI maturity. Your storage architecture decisions determine how quickly you can get to model training and innovate. 

Where do you see cloud storage playing the most critical roles across the AI lifecycle—from data ingest and processing to training, fine-tuning, inference, and monitoring?

Cloud storage is critical across the AI lifecycle, with key stages in data aggregation, processing, training, and inference. At the start, systematically consolidating, cataloging, and securing archives speeds up new projects and makes it easy to test emerging models. Clean, well-processed data often beats just having more data, which makes storage central to quality as well as scale. One of my favorite Backblaze sayings is, “It’s not hoarding if it’s data.” You never know how valuable it will be, so organizations should collect as much as possible. During training, scalable storage ensures throughput of massive datasets, and at inference, capturing prediction outputs and user feedback enables continuous iteration. In the end, storage is the foundation that determines how quickly you can innovate with AI.

What are the biggest obstacles organizations face when scaling storage for AI, and how do those challenges differ between smaller startups and large enterprises?

The biggest obstacles in scaling storage for AI are cost, data management, and accessibility. Storing large volumes of data is only part of the challenge; it must also be organized, retrievable, and governed with the right controls. Clean, well-structured data is often more valuable than simply having more of it.

For startups, the initial challenge is acquiring enough data to train and refine their models. Once they have it, cost and architecture become the next barriers.

For large enterprises, the challenge is complexity. Their data is abundant but fragmented across silos, legacy systems, and compliance regimes, making consolidation and accessibility difficult.

The organizations that succeed treat storage as a strategic enabler that scales in cost, performance, and accessibility alongside their AI maturity. 

Among cost, latency, security, and compliance, which do you see as the most pressing barrier to scaling AI today, and how should organizations prioritize addressing it?

Among cost, latency, security, and compliance, latency is one of the most pressing barriers. It directly impacts both model training and inference, and inference in particular shapes the user experience. Organizations do as much as possible to reduce latency at this stage, since delays in serving predictions can undermine adoption.

Cost remains a constant challenge as data volumes grow, and compliance becomes more critical as organizations scale, especially in regulated industries. Startups often focus first on cost and latency, while enterprises must balance latency with governance and regulatory demands. The priority should be building storage that minimizes latency for training and inference, while still being cost-efficient and compliant as AI adoption expands. 

Enterprises often emphasize the need for flexibility and easy access to data in order to drive AI innovation. From your perspective, what does true flexibility in data access look like, and why is it so essential?

In a recent talk I gave, I emphasized the idea of smart archiving. True flexibility in data access starts with centralizing information into a structured, searchable archive. That means unifying diverse formats, normalizing and tagging for consistency, and enabling indexing for future querying. This approach ensures that data is not just stored but made usable.

It is essential because it lays the foundation for analytics and modeling. When data is structured and searchable, teams can move faster, experiment more freely, and reduce latency in both training and inference. Without this kind of flexibility, storage quickly becomes a bottleneck instead of an enabler for AI innovation. 

Can you share real-world use cases—such as with customers like Decart AI or Wynd Labs—that demonstrate how the right cloud storage approach can directly enable AI innovation?

These are two great examples of how the right cloud storage approach directly enables AI innovation. Decart focused on model training, where moving data to compute efficiently was critical. With Backblaze B2 they scaled to 16PB in 90 days, trained across multiple GPU clusters with zero egress cost, and achieved ten times the efficiency of competitors. That reliability and efficiency freed them to innovate faster.

Wynd Labs focused on customer access to data. They ingest petabytes daily and serve tens of petabytes monthly. With Backblaze’s high performance and free egress, they could scale to enterprise demand and reinvest resources into product development. That ability to deliver data access at scale unlocked new opportunities for their platform.

In both cases, the right storage strategy transformed infrastructure from a constraint into an enabler, allowing the companies to focus on innovating in AI rather than managing cost and complexity

As AI models and datasets grow more complex, what guidance would you give organizations trying to balance storage performance with cost efficiency?

Organizations need to think about their long-term data usage with their product in mind. Collecting, processing, moving, and running inference on data will all be core to how their product evolves. If they do not account for that now, the costs and storage challenges only compound over time. Since AI will be a central part of their product and their organization, storage must be designed early to balance performance with cost efficiency so it can scale smoothly as they grow.

Security and compliance are especially pressing in regulated industries. How do you see cloud storage evolving to support governance needs while still allowing teams to innovate quickly?

Governance is a key part of storage. Streamlining access with a solid foundation for how data is managed, secured, and audited is critical. I see cloud storage evolving with stronger built-in controls such as encryption by default, fine-grained permissions, audit trails, and data residency options. Just as important is data lineage. In AI, knowing where data came from, how it has been processed, and how it feeds into models is essential for both compliance and trust.

At the same time, storage platforms are improving usability so teams can still move quickly. When governance, lineage, and accessibility work together, organizations can meet regulatory requirements while continuing to innovate with AI at speed.

For organizations evaluating or migrating to B2, what advice or guidance do you provide in terms of implementation—particularly regarding data migration, integration with existing MLOps or compute stacks, or optimizing for throughput and egress?

Because B2 is S3 compatible, it integrates directly into existing MLOps and compute stacks without re-architecture. We often work with clients on a proof of concept to validate migration, performance, and integration before scaling. From there, the focus is on optimizing throughput, data movement, and data orchestration so teams can train across clusters, run inference, and iterate quickly without being slowed down by infrastructure bottlenecks. 

As AI workloads continue to scale—especially with trends around LLMs, exabyte-scale datasets, and hybrid or multi-cloud strategies—how is Backblaze evolving its storage offerings to meet these emerging needs?

At Backblaze we are focused not just on how data is used today but how it will be orchestrated in the future. Storage is no longer only an archive, it is becoming a tool that enables fast access, efficient movement, and reliable orchestration of data across environments. With LLMs and exabyte-scale datasets, this foundation of easy access and high throughput will be critical not only for training and inference but also for the emerging class of AI agents that rely on data to make processes more autonomous. The result is a storage foundation that enables innovation now and prepares organizations for what comes next.

Thank you for the great interview, readers who wish to learn more should visit Backblaze.