AI Storage Bursts Outgrow Traditional Capacity Planning
AI storage traffic can arrive in concentrated bursts that challenge networks and storage systems designed around steadier growth. Backblaze says measurements from its infrastructure show that the cost per terabyte is only part of the issue. Throughput, the amount of data transferred in a given time, and latency, the delay before a transfer begins or completes, can determine whether expensive computing resources remain productive or sit idle.
The company publishes three quarterly datasets based on its infrastructure and testing. Network Stats records where data enters and leaves its network. Performance Stats compares upload and download behaviour through a neutral third-party testing framework. Drive Stats, maintained since 2013, follows hard-drive failure rates in its data centres. Backblaze publishes the methodologies and says it reports results even when they are unfavourable to its own service.
Its network measurements show AI data arriving in large batches, being stored and then moving to whichever computing environment offers suitable cost or availability. This supports multi-cloud architecture, meaning an application uses infrastructure from more than one cloud provider. Preferred AI tools and access to computing capacity can change from month to month, making data movement a central design concern rather than a negligible storage detail.
Backblaze classifies connected networks as content delivery networks, traditional hosting providers, hyperscalers, regional internet service providers, Tier 1 internet carriers or neoclouds. A neocloud is a computing provider focused on AI workloads. The company measures each category using magnitude, defined as the number of data bits transferred per Internet Protocol address, or IP address, which identifies a destination on a network. It reports that neocloud magnitude is so much higher than the other categories that direct comparison is of limited use.
The shape of that traffic matters. Conventional workloads often involve many sources transferring predictable quantities. AI workloads can involve relatively few sources producing elephant flows, meaning exceptionally large network transfers, at highly variable times. Backblaze says these transfers can appear as concentrated bursts from small numbers of GPUs, or graphics processing units used for intensive AI calculations, located around major AI computing infrastructure.
That pattern undermines a traditional capacity-planning approach based on observing several months of growth and expecting demand to double every two years. Infrastructure teams may instead need to add bandwidth in larger increments, strengthen internal network switching for sudden loads and establish reliable peering relationships, which are direct connections used by networks to exchange traffic. Storage designed around conventional traffic may otherwise be undersized or incorrectly configured for AI storage demand. Backblaze cautions that Network Stats is still a young dataset: the patterns are directionally useful, but the available measurements do not yet support firm conclusions.
Performance testing presents a second problem because there is no single benchmark that represents every AI workload. Model serving and distribution place pressure on download throughput. Machine learning operations, commonly called MLOps, and telemetry can be single-threaded, meaning they perform transfers through one processing stream rather than many in parallel. Inference, the process of using a trained model to produce an answer, can require low read latency when retrieving model files quickly.
Other applications create different demands. Retrieval-augmented generation, or RAG, repeatedly reads relatively small stored objects after writing them once. Log and event analysis needs competitive performance for small writes and fast reads. Interactive data-lake queries, which search a large collection of stored information, rely on consistent throughput and rapid downloads. Backblaze therefore tests single-threaded and multi-threaded uploads and downloads across several file sizes rather than reducing performance to one score.
The company also disclosed that its initial testing reached its own bandwidth caps. It published the rate-limited results, documented what triggered the restriction and then repeated the tests with what it described as the smallest possible adjustment. Both sets of results were presented side by side. This is a useful testing principle for any buyer: document limits rather than quietly changing the benchmark when a result is inconvenient. Backblaze says two quarters of Performance Stats reveal some patterns, but are not enough for definitive conclusions across regions. Organisations should map the tests to their workloads and repeat them with their own applications.
Physical media remain part of the equation. Drive Stats covers hard drives rather than solid-state drives, or SSDs, which store data electronically and can provide very fast access. Backblaze argues that SSDs are particularly important when fast storage must keep GPUs supplied during AI training, while mechanical hard drives remain relevant for the much larger volumes of data surrounding those intensive stages because they cost significantly less per gigabyte.
Backblaze reports that hard-drive prices have risen 46% since late 2025, while lead times have extended to 12 months or longer as AI data centres consume production capacity. Western Digital has said nearly all its production for the year is committed, with hyperscale customer contracts extending into 2028 and 2029. Seagate is shipping drives of up to 44TB using its heat-assisted magnetic recording, or HAMR, platform and says multiple hyperscale cloud providers have qualified them. Western Digital announced architecture work at its Innovation Day 2026 showing a current twofold throughput improvement, with a planned route to fourfold and eventually eightfold improvement.
Higher-capacity drives are also changing storage density. At the end of 2025, drives of 20TB or more represented almost 23% of Backblaze's fleet by unit count and a larger proportion of capacity. Its first 26TB drive entered service in the fourth quarter of 2025. Replacing 8TB units with 24TB drives allows the same amount of information to occupy one-third of the rack space.
Greater density complicates protection. Backblaze uses Reed-Solomon erasure coding, a method that divides files into data fragments and recovery fragments called parity shards, then distributes them across the storage fleet. Larger drives require more parity shards, which must be placed carefully so too many do not occupy the same rack. Otherwise, equipment failure could affect multiple fragments needed to reconstruct the information.
Drive Stats covers 557 million drive-days, calculated by adding every day that every tracked drive operated over 13 years.