AI Cloud Storage Needs Tiered, Not Just Flash
AI cloud storage built around GPUs is running into a mismatch: the fastest storage, flash, is being used for jobs that do not need its speed, and that pushes data into hyperscaler platforms with expensive exit fees. Backblaze product marketing director David Johnson lays out the problem in a September 24, 2026 post, arguing that neoclouds, the GPU-focused cloud providers, can keep the whole customer relationship if they adopt a tiered storage architecture. Neoclouds are built around renting out GPUs, the processors that train AI models. GPUs can only run as fast as the data coming in.
Johnson explains that a neocloud's infrastructure is built around flash. Flash storage is the type of very fast solid state memory used for active training; it delivers data at extreme speed and low latency, meaning very little delay. But the storage work that surrounds active training, saving checkpoints, storing datasets, keeping finished models, does not need flash-grade speed or flash-grade prices. So for storage that does not need flash, neoclouds often point customers to hyperscaler object storage instead. It is cheaper, it works, and it keeps the neocloud focused on its core GPU product. The problem is data gravity: the more data you have stored somewhere, the harder it becomes to leave. Hyperscalers make this worse by design. They let you put data in for free but charge prohibitively expensive egress fees, the charges for downloading or moving data out. So once your data is entrenched on a hyperscaler's platform, the cost of leaving keeps climbing. And wherever your data lives, that is where you tend to buy more services. So while the neocloud keeps getting paid for GPU computing, the hyperscaler quietly takes over the broader account.
To see why flash alone is not the answer, Johnson walks through each stage of the AI lifecycle. Data ingestion first: raw datasets have to be cleansed, labeled, and prepared. This stage needs high-capacity storage that can move large volumes of data efficiently, but not flash. Active training is the stage where flash earns its place. Data moves through several storage tiers before it reaches the GPU, but at that final step the model needs it delivered continuously and at very high speed. Only flash is fast enough to keep thousands of GPUs running without interruption. Checkpointing runs throughout training. Because AI training is a synchronous process with every node in the cluster working in lockstep, a single component failure can bring the entire training run to a halt. At the scale of a large modern cluster, failures are not rare; the failure rate scales exponentially with cluster size. Checkpoints are frequent snapshots of the model's current state that allow training to resume from where it left off. At large scale, a checkpoint may need to capture 15TB of data in under five seconds, repeated every few minutes, around the clock. This stage needs storage fast enough to keep up and reliable enough that nothing gets lost. Once training finishes, the outputs, model weights, logs, and metadata, go into longer-term storage. Fine-tuning experiments and additional training runs add more over time. This stage needs capacity and durability, not speed. Finally, inference, serving a deployed model, requires repeated access to model weights on demand. The data volumes are far smaller than training, and while latency-sensitive deployments may still benefit from flash, the storage demands are far less intensive.
Flash storage is purpose-built for only one of these stages, active training, according to Johnson. At every other stage, what is needed is a cost-effective, high-capacity storage tier, which is precisely what neoclouds are currently sending to hyperscalers. NVIDIA's certified storage program validates storage solutions separately for each of these stages, recognizing that a single storage architecture cannot serve all of them well. Flash is expensive because it is built for speed. But you pay for that speed whether you need it or not. Using flash to store datasets, checkpoints, and finished model artifacts is like shipping all of your mail overnight when most of it could go by ground; you are paying a premium for a capability you do not need. There is a second problem beyond cost. Flash that is tied up storing checkpoints and datasets is not available for GPUs, and GPUs are what neoclouds sell. Every terabyte of flash used in the wrong place is capacity taken away from the core product. The economics have always favored a tiered approach: fast storage where speed is needed, cheaper high-capacity storage everywhere else. Google, Meta, and Microsoft have run tiered storage architectures for years, and Meta's engineering team describes their approach in four explicit tiers, each selected for the job it does best. But building out a full storage tier below flash is a significant undertaking, and most neoclouds have stayed focused on their core GPU product. With SSD prices up 257% in under a year, that gap has become too expensive to ignore.
The solution is to match storage to workload. For active training, flash remains the answer. VAST, WEKA, and DDN handle this tier with fast, GPU-adjacent storage that keeps training running at full speed. For everything else, Backblaze's B2 Neo can fill the gap. Datasets live there before training starts. Checkpoints write there continuously during training. Finished model weights and records accumulate there over time. The storage is fast and durable, but priced for capacity workloads rather than flash workloads. B2 Neo connects directly to the hot tier, so data moves between them without a detour through a hyperscaler. The neocloud does not have to build any of this. B2 Neo is available as a white-label product, deployed under the neocloud's own brand. The customer only sees one seamless platform; their data stays on the neocloud's infrastructure from raw dataset to deployed model and never touches a hyperscaler.
Flash is the right storage for one stage of the AI lifecycle. At every other stage, it is an expensive tool for a job that does not need it. Neoclouds that only serve the active training tier leave a gap, and customers often fill that gap by sending their data to a hyperscaler. Once it is there, egress fees keep it there. The neocloud ends up handing the broader customer relationship to whoever holds the data. Closing the gap does not require building anything from scratch. The tiered architecture that Google and Meta run at scale is available to neoclouds today as white-label infrastructure. In this market, the GPU contract is just the beginning. Keeping the data is what keeps the customer. For organizations facing similar data gravity and storage tiering decisions, AEU Data offers managed object storage and encrypted automated backups that fit the cheaper capacity tier without exit-fee lock-in, helping teams keep control of their own data.