WEKA has signed Backblaze's B2 cloud object storage as the capacity tier beneath its NeuralMesh high-performance file system, letting AI teams keep training datasets and model checkpoints on cheap object storage while feeding GPUs from flash.

The two companies frame the arrangement as a hot-and-cold split. WEKA NeuralMesh serves the microsecond-latency read path that GPUs demand, while B2 holds the cold bytes that no longer need flash but cannot be deleted. The combined topology covers raw data ingestion through training, checkpointing, inference, and downstream reuse.

The pair combine WEKA's NeuralMesh, a parallel file system tuned for GPU-driven AI workloads, with Backblaze's B2 cloud, a commodity S3-compatible object store. Raw training data can sit in B2, move into NeuralMesh once a job demands microsecond latency, and return to B2 for retention or rollback.

Why split hot and cold storage for AI?

GPUs idle when the storage pipeline stalls, and the storage pipeline now spans exabytes for frontier model runs. Keeping every byte on flash burns capital that object storage offers for a fraction of the cost.

Backblaze CEO Gleb Budman framed the customer problem plainly: "AI teams need their GPUs fed and an infrastructure with the performance and capacity to support the full AI data workflow." WEKA chief strategy officer Nilesh Patel put the load profile more bluntly: "GPUs need microsecond access to data to stay fed, while datasets and checkpoints are growing to exabyte scale."

The two companies have validated WEKA's Snap-to-Object capability against B2, so a checkpoint can be written to object storage and recovered later. No retraining, no re-ingest, no extra GPU bill.

What does the deal price in?

The agreement carries no headline number, unlike Backblaze's recent $335 million strategic pact with CoreWeave. That deal marked CoreWeave as the fourth major neocloud to contract Backblaze.

The pattern tells a story. B2, a storage service the consumer backup market once carried on its shoulders, now underwrites multi-hundred-million-dollar AI infrastructure commitments. Less than three years separates the backup drive from the neocloud cornerstone.

The price gap forces the choice. Object storage runs at a fraction of flash cost per gigabyte, and AI training produces more retained data than any single flash array can hold. Customers no longer ask whether to split tiers. They ask which storage partner covers both jobs.

Who else is pairing performance and capacity storage?

• Scality and WEKA struck a similar NeuralMesh-plus-RING deal in February, with Scality's object tier standing in for B2.
• Backblaze rival Wasabi launched a dedicated AI business in the same window and hired semiconductor storage veteran Pinaki Mukherjee as SVP and GM to run it.
• Wasabi claims to hold hundreds of petabytes of AI data for frontier model labs, generative AI startups, and neocloud compute providers, after raising $70 million in January and arranging a $250 million credit facility.

Certification of B2 Cloud Storage for NeuralMesh remains in progress, and both vendors are taking customer enquiries now. The next round of contracts will decide whether one vendor stacks both tiers or two specialists keep trading customers.