Sep 4, 2026 · 27 min
Storage bottlenecks leave AI’s GPUs waiting
EP 65: The AI Bottleneck Nobody's Watching | Troy Liljedahl, Backblaze
As AI systems move toward production, storage capacity, throughput, cost, and sustainability can matter as much as model performance.
- 1Storage constraints can stall generative AI projects between promising pilots and production deployments.
- 2Idle GPUs create opportunity costs, making data preparation and storage economics central to AI infrastructure decisions.
- 3Engineers will increasingly manage AI agents while relying on architecture, review, logic, and business judgment.
Don't miss
Troy Liljedahl’s advice to use intelligent tiering across warm object and spinning-disk storage crystallizes the episode’s cost-and-throughput argument.
The brief
Sam Cooper and Troy Liljedahl frame a less visible AI constraint: storage bottlenecks can keep expensive GPUs idle and prevent promising systems from reaching production.
Liljedahl, a senior director of solution engineering at Backblaze, connects storage performance to data preparation, infrastructure costs, and the competitive pressure surrounding AI adoption.
The discussion turns to scale, from intelligent tiering across warm object and spinning-disk storage to the expanding data demands of physical AI and high-performance workloads.
The sustainability debate broadens the infrastructure question: AI growth must be matched by efficient systems and data-center choices that limit harm to communities and resources.
For engineers, the shift is not simply job loss; it is a move toward managing agents while preserving architecture, review, logic, and business understanding.
Featuring
Backblaze
OpenClawListen to the full episode and explore every guest, topic, and moment on PodLume.