Webcast
🎙 DataStorage.com x
Backblaze:
Scaling Always‑Hot Storage for AI, Media, and SaaS Workloads
Hear how cloud storage is evolving to support high‑throughput AI, media, and SaaS infrastructure.
As AI workloads scale, many teams discover that compute is no longer the limiting factor. Storage becomes the constraint.
During the conversation, Gleb Budman, CEO of Backblaze, explained that teams often over‑invest in GPUs without addressing the data layer.
“You can throw GPUs at the problem, but if you can’t feed them fast enough, they sit idle. That’s wasted money. Storage throughput becomes the bottleneck.”
Legacy object storage systems were designed for backup and archival use cases, not for sustained, high‑throughput access. When training models or running large‑scale inference, latency at the storage layer translates directly into higher infrastructure costs.
Backblaze B2 Cloud Storage was built with a different assumption. Data should always be available, without tiering or retrieval delays.
“Our customers are using B2 to feed training pipelines, serve media, or move large volumes of data continuously,” Budman said.
“The expectation is that it’s always available.”
Instead of moving data between hot, warm, and cold tiers, the platform keeps everything accessible at all times. That simplicity removes friction for AI teams, media workflows, and SaaS platforms that depend on fast access to large datasets.
Budman described how storage throughput directly affects GPU efficiency.
“If you’re training a vision model and every image fetch adds latency, that time stacks up very quickly.”
By removing storage delays, teams reduce idle GPU time and complete training runs faster. While the exact gains vary by workload, the principle is consistent: faster data access leads to better utilization of expensive compute resources.
One of the strongest themes from the discussion was how AI is changing the value of older data.
“There’s all this data people thought wasn’t useful anymore,” Budman said.
“But now with AI, they’re realizing it’s incredibly valuable, and they need access to it.”
Media archives, historical logs, and long‑tail datasets are now actively reused for training, analytics, and multimodal models. Storage systems that assume infrequent access create delays and unnecessary cost when that data suddenly becomes critical.
As storage becomes performance‑critical, teams should evaluate platforms based on real workload needs:
“It’s not just about how much data you can store,” Budman said.
“It’s about how quickly you can use it.”
🎧 In the full webcast, John Kosturos and Gleb Budman discuss:
