Backblaze, the cloud storage company, published a blog post on September 24 arguing that GPU rental clouds built for AI training, often called neoclouds, are giving away their most valuable customer relationships without realizing it. The piece, written by David Johnson, Backblaze’s director of product marketing, doubles as a pitch for the company’s own B2 Neo storage product.
The mechanism Johnson describes is specific. Training needs a continuous, high-speed feed of data, and flash is the only storage tier quick enough that thousands of GPUs never sit idle waiting on it. That is why neoclouds build their infrastructure around flash. But most of an AI project’s storage needs, raw datasets before training starts, checkpoints saved during training, finished model weights afterward, don’t need flash speed. Neoclouds without a cheaper storage tier of their own point that work to hyperscaler object storage instead.
That referral is where the trouble starts, according to the post. Once a customer’s data sits inside a hyperscaler’s storage system, egress fees make it expensive to move. The customer then tends to buy more services wherever the data already lives. The neocloud keeps collecting for GPU time while the hyperscaler absorbs the rest of the account: storage, data pipelines, and whatever gets built on top.
Backblaze frames this as a lifecycle problem rather than a one-time mistake. Data ingestion needs high-capacity storage, not flash. Checkpointing, saving a snapshot of a model’s state so training can resume after a failure, needs speed and reliability but not necessarily flash-grade pricing; the post gives a concrete figure, saying a checkpoint at large scale may need to write 15 terabytes in under five seconds, then repeat that write every few minutes for the length of the training run. Finished model weights need durability more than speed. Only the active training step genuinely requires flash.
The post cites a 257 percent rise in SSD prices over the past year, a figure it attributes to the trade publication CIO.com, as the reason this gap has become too expensive for neoclouds to keep ignoring. It also notes that Google, Meta, and Microsoft already run their own tiered storage systems internally, which is less a neutral fact than a setup line for the pitch that follows: B2 Neo, sold as white-label infrastructure a neocloud can put its own brand on, positioned as the missing tier between flash and the hyperscalers.
None of this comes with independent verification. The post does not name a single neocloud customer who lost business to a hyperscaler this way, nor does it cite a third-party account of a neocloud’s margins improving after adding a cheaper storage tier. The argument is plausible on its economics, egress fees are a well-documented industry complaint, but the specific claim that this is actively costing neoclouds their accounts is Backblaze’s own framing, made in service of a product it sells.
For a neocloud operator, the practical question is not whether a cheap storage tier below flash makes sense; storage architects have known that for years. It’s whether adopting a vendor’s white-label product to build it is worth ceding another layer of infrastructure independence, given that the entire pitch is built around ceding infrastructure independence to nobody but Backblaze.
Reported by Backblaze on 24 September 2026.
