Backblaze and WEKA Are Building the Storage Architecture AI Needs to Deliver at Scale

Key Highlights

• Backblaze and WEKA are validating an architecture that pairs WEKA NeuralMesh with Backblaze B2 Cloud Storage for different stages of the AI data lifecycle.
• NVIDIA's Cloud Partner architecture assumes a high-speed file storage cluster and an optional object storage cluster, giving this two-tier model greater visibility as AI infrastructure moves toward production.
• WEKA provides the performance-sensitive tier and global namespace, while Backblaze provides scalable object capacity for large datasets, checkpoints, outputs, and retained AI assets.
• The partnership also gives Backblaze a validated sell-with model that can extend B2 into broader AI architectures while adjacent infrastructure capabilities remain with partners.

The News

Backblaze and WEKA announced a collaboration to simplify data management throughout the AI lifecycle by combining Backblaze B2 Cloud Storage with WEKA NeuralMesh. The companies are validating the platforms together so performance-sensitive workloads can use NeuralMesh while large datasets, checkpoints, outputs, and retained AI assets reside in B2. WEKA Snap-to-Object is also being tested with B2, with certification of B2 Cloud Storage for NeuralMesh underway. For more information, read the official Backblaze-WEKA partnership release.

Analyst Take

The announcement came as Backblaze held its first Investor Day since going public in 2021. In a recorded conversation shared during the event, WEKA Chief Strategy Officer Nilesh Patel and Backblaze CTO Dan Spraggins framed the AI storage challenge around two requirements: very high performance close to accelerated compute and rapidly expanding capacity behind it.

NVIDIA’s Cloud Partner reference architecture assumes a file storage cluster and an optional object storage cluster, while anticipating access to both high-speed file systems and object storage for different workload requirements. In our view, the Backblaze-WEKA architecture maps to that model, with WEKA serving the performance-sensitive layer and B2 providing an external object capacity tier. WEKA provides premium performance where the workload requires it. B2 competes for the larger capacity tier on economics, sustained throughput, rapid capacity availability, and egress terms. WEKA also pointed to infrastructure availability as part of the value proposition: customers can add capacity now instead of waiting for storage supply-chain constraints to ease. NeuralMesh can tier to any S3-compatible target, so Backblaze has to earn its position through those factors and through the integration work behind the validated design.

WEKA's namespace also matters as enterprise data becomes more distributed. Active datasets can move closer to GPU infrastructure when needed, while larger data estates remain on scalable capacity. That gives customers a way to reserve premium storage for active workloads while keeping more data available for training, inference, and reuse.

HyperFRAME Research Lens: State of the Enterprise Infrastructure and Operations (1H 2026)data shows why this architecture is relevant beyond specialized AI infrastructure providers. Only 14% of enterprises say their current data architecture is AI-ready, while 53% identify performance as a major infrastructure priority. At the same time, 55% are targeting hybrid architectures, creating a larger requirement to make data available as compute moves between environments.

The partnership also speaks to Backblaze's corporate strategy. The company says it is concentrating on making its object storage platform faster, larger, and easier to consume, while complementary infrastructure capabilities remain with partners. That focus is backed by more than a decade of experience operating high-performance HDD infrastructure at scale, along with the software engineering and published drive-performance and reliability data that support it. Backblaze and WEKA handle the integration, sizing, tuning, and testing in advance, creating a tested configuration customers can deploy with less engineering work. That gives Backblaze a new sell-with motion while keeping its focus on the object tier.

Backblaze also has more financial flexibility to support that focus. In August, the company priced $175 million of 0% convertible senior notes due 2031, with a further $26.25 million purchaser option and proceeds intended for capped call transactions, general corporate purposes, and capital expenditures. In our view, Backblaze has defined the part of the AI infrastructure stack it intends to own and is directing investment and partnerships around that role.

Certification remains underway, and customers will need to see sustained throughput when large datasets move between B2 and NeuralMesh environments close to GPU compute. The value of the design will depend on whether customers can gain capacity-tier economics while maintaining the performance required for production training and inference.

What Was Announced

Backblaze B2 Cloud Storage and WEKA NeuralMesh are being validated as complementary storage layers for AI environments. Large datasets, training data, media libraries, and other retained assets can remain in B2 until a workload requires higher performance. Data can then move into the NeuralMesh environment for training, inference, or other GPU-intensive processing, while B2 continues to provide the larger capacity tier. The architecture also supports the return path. Checkpoints, model outputs, saved inference data, and other assets can move back to B2 after their high-performance phase for retention, reuse, or recovery. WEKA Snap-to-Object has been tested with Backblaze to support checkpoint and inference-data recovery from the object tier. Certification of B2 Cloud Storage for NeuralMesh remains underway.

The companies are also doing the integration work in advance. Sizing, tuning, testing, and validation are intended to give customers a tested configuration ready for deployment. The partnership is selling a deployable architecture, and that scope is the substantive part of the announcement.

Looking Ahead

Training established the need to keep GPUs supplied with data at very high throughput. Inference and agentic AI add pressure from larger context windows, cache, memory, persistent state, and growing volumes of generated data. Capacity requirements will grow in parallel as enterprises bring more existing information into AI workflows.

That increases the value of separating the economics of capacity from the performance requirements of GPU infrastructure. NVIDIA's reference architecture is making this model more explicit, with an assumed high-speed file tier and an optional object storage capacity tier. In our view, that should increase its visibility with neoclouds, AI factories, and enterprises as more inference workloads move into production.

Backblaze and WEKA provide one implementation of that model. WEKA keeps the active data path close to compute and provides a global namespace for distributed data, while Backblaze provides scalable object capacity behind it. Validated designs with complementary partners can extend B2 into more AI environments while allowing Backblaze to remain focused on the performance, capacity, and economics of the object tier.