AI Workloads Are Breaking Old Storage Assumptions

Backblaze is expanding its quarterly operational statistics program to three franchises—Network Stats, Performance Stats, and Drive Stats—to document how storage infrastructure behaves under sustained parallel AI workloads, with Drive Stats having tracked hard drive failure rates from Backblaze data centers since 2013. Backblaze says AI workloads differ structurally from traditional ones, producing concentrated bursts of "elephant flows" from a small number of GPU sources rather than many predictable sources, and argues that cheap storage that performs poorly under AI load leaves expensive GPU compute idle. The company reports only on its own network and invites other infrastructure providers to share comparable data.

AI’s impact on storage is often framed as a cost story if it’s included in the conversation at all, given the industry’s propensity to focus on compute. That framing isn’t wrong, but it’s incomplete. The harder questions are about end-to-end performance and architecture. How does a storage system actually behave when an AI pipeline is pushing it? Where does it break? What does real throughput look like under sustained parallel load? Most providers don’t answer those questions. Sometimes it’s because the impact of storage is incorrectly viewed as negligible; other times it’s because the answers aren’t flattering. The idea of a multi-cloud architecture isn’t new; but the current rapid evolution of preferred tools and high demand for compute means that it can no longer be ignored. Backblaze publishes three ongoing quarterly stats franchises: Network Stats https://www.backblaze.com/blog/tag/networkstats/ , Performance Stats https://www.backblaze.com/blog/tag/performancestats/ , and Drive Stats https://www.backblaze.com/cloud-storage/resources/hard-drive-test-data . Each one tracks real operational data from our own infrastructure and testing, with the methodology published openly and the results reported in full. - Network Stats https://www.backblaze.com/blog/network-stats-for-q1-2026-neocloud-traffic-trends/ tracks how data actually moves into and out of Backblaze’s network and where it’s going. - Performance Stats https://www.backblaze.com/blog/backblaze-performance-stats-q1-2026/ benchmarks upload and download performance against competitors using a neutral third-party framework. - Drive Stats https://www.backblaze.com/cloud-storage/resources/hard-drive-test-data has been tracking hard drive failure rates from Backblaze’s data centers since 2013. Cost is a significant factor and belongs in the conversation https://www.backblaze.com/blog/the-hidden-cost-of-cloud-storage-what-400-it-leaders-wish-they-knew-sooner/ . But, more importantly, a storage system that’s cheap and unreliable under AI load will slow you down, cost you more in GPU time due to the expensive compute resources sitting idle, and create exactly the kind of unpredictable behavior that makes AI pipelines expensive to operate. What you pay for storage is far from being the only factor to consider; it’s what you get when the workload gets hard https://www.backblaze.com/blog/architecting-your-ai-data-pipeline-using-b2-overdrive/ . These three stats franchises together work to answer those questions—what actually happens when you push AI workloads through a storage system, reported in full. Network Stats Network Stats makes the AI data movement visible. Data arrives in large batches, gets stored, and gets pushed out to whatever compute environment makes the most sense at the moment which can change month to month depending on cost or availability . On a network, that means neocloud traffic https://www.backblaze.com/blog/neoclouds-are-winning-on-compute-storage-shouldnt-slow-them-down/ arriving in concentrated bursts from a small number of GPUs, clustered where AI compute infrastructure is densest. This is structurally different from how other workloads use storage. Traditional workloads involve many sources moving predictable amounts of data, which makes them easy to model and plan around. AI workloads mean few sources, elephant flows, https://www.backblaze.com/blog/making-the-backblaze-network-ai-ready/ and highly variable timing. One small caveat: It’s true that we can only report on our own network, but we do so in a structured way—measured, classified, and tracked quarter over quarter. For any other infra providers out there, we’re here mailto:drivestats@backblaze.com if you want to collaborate and share similar data. What the data shows The old capacity planning model was straightforward: track trends over a few months, expect demand to double every two years, plan accordingly. But AI-scale storage demand doesn’t work that way. It arrives in bursts, at magnitude. Planning for it means adding bandwidth in larger and wider increments, https://www.backblaze.com/blog/point-to-point-at-100gbps-what-ai-grade-infrastructure-actually-requires/ hardening internal switching for burst traffic, and establishing solid peering relationships. Backblaze captures network flow data across all of our data center https://www.backblaze.com/blog/data-centers-temperature-and-power/ regions every month. To make AI traffic visible as a distinct category, Backblaze classifies every network we connect to by type: - CDN: Networks that use Backblaze as an origin store for content delivery - Hosting: Traditional hosting providers that runs workloads like physical or virtual servers for web, database, or application tasks - Hyperscaler: Large, traditional cloud providers - ISP Regional: Local or regional ISPs, think of these as the “last mile” paths as these networks are very close to customer equipment and efficient - ISP Tier1: National or international ISPs that carry our traffic long distances - Neocloud: AI -focused compute networks For each type, we measure magnitude https://www.backblaze.com/blog/network-stats-for-q3-2025-the-magnitude-of-ai-workflows/ —it’s a metric that our Network Engineering Manager Brent Nowak https://www.backblaze.com/blog/author/brentnowak/ started tracking, defined as bits of data transferred per IP address. But neocloud traffic is so far above everything else that the comparison is almost meaningless. Storage systems sized and designed for traditional traffic patterns will be undersized and misconfigured for AI workloads. A note on the data: Network Stats is a young dataset. The patterns above are directionally meaningful, but drawing firm conclusions at this stage of the data’s maturity would go further than the numbers support. Performance Stats It’s hard to find good numbers on cloud storage performance. You may be able to find some considered conversations in developer spaces https://www.reddit.com/r/backblaze/ . But, anytime anyone talks about how good or bad a company is as a storage provider, there are limitations. Let’s pose them as questions: - How do they work with my other tools? - How do I know they’re reliable and fast? - Does it work for my use case? - Can I trust these numbers? This is precisely why we publish the methodology as the primary focus of the report. We treat it like science—we want the testing to be repeatable, even if we’re not the ones running it. If you ever use a similar methodology and want to share your results with us, email our team mailto:drivestats@backblaze.com . One example of what that looks like in practice: when our own testing https://www.backblaze.com/blog/backblaze-performance-stats-q1-2026/ hit Backblaze’s bandwidth caps, we didn’t adjust the methodology to make the results look cleaner. We published the rate limited results, documented what triggered the limit and how you can reach out; then, retested with the minimal possible change and tested both the rate limited and adjusted results side-by-side. If that sounds like we did a lot—sure. But we’ve been on the other side https://www.backblaze.com/blog/introducing-the-backblaze-flamethrower-startup-program/ when we want to understand architecture and build a product. We want to understand what you actually might run into and how it could affect your product; we also want to be transparent. That’s the standard we’re holding ourselves to. If you have questions, check out the reports https://www.backblaze.com/blog/tag/performancestats/ . If you have more questions, do your own testing, and then let us know how your results were different. mailto:drivestats@backblaze.com What the tests actually measure There is no single test that represents an AI workload. What you need from storage depends entirely on what you’re building, and how that specifically impacts storage performance. For example: - Model serving and distribution puts pressure on download throughput. - MLOps pipelines and telemetry tend to be single-threaded. - Inference workloads that need to pull model artifacts quickly stress read latency. - RAG systems write once and read repeatedly at small object sizes. - Log and event analytics require competitive small-write performance and fast reads. - Interactive data lake querying depends on consistent throughput and fast download speeds. Each of these maps to a different part of the benchmark suite. That’s why Performance Stats publishes the full range—single-threaded and multi-threaded tests for both upload and download across multiple file sizes. This gives you a starting point to evaluate which provider performs best for your specific workload type. As the dataset matures across quarters and regions, the picture gets more useful. Two quarters of data is enough to see some patterns, but not yet enough to draw firm conclusions. However, it is enough to start mapping your own workload requirements against what the data shows, and to run your own tests against your own stack. Drive Stats If Network Stats shows how AI traffic behaves and Performance Stats shows how storage systems hold up under AI load, Drive Stats https://www.backblaze.com/cloud-storage/resources/hard-drive-test-data shows what’s happening to the physical infrastructure underneath. We’ve been tracking hard drive failure rates in our data centers for over 13 years, and while drives are still fundamentally mechanical, we’ve also seen them improve significantly over time https://www.backblaze.com/blog/are-hard-drives-getting-better-lets-revisit-the-bathtub-curve/ . It’s worth noting that we’re not reporting on SSDs. While SSDs are arguably essential at certain points in the AI tech stack—notably, when you need super hot storage to feed GPUs at the training stage—hard drives are still relevant. Hard drives cost significantly less per gigabyte than SSDs, which is important for the huge amount of data that you need for AI workflows. And as hard drives get faster and more capable, that case gets stronger. This is the same logic that drove processor specialization: just as AI workloads pushed the industry toward GPUs for heavy computation and CPUs for everything else, different storage types are becoming suited to different parts of the workload, each optimized for its specific job in the AI stack. The hard drive industry is responding to that opportunity under enormous pressure. AI data centers have consumed production capacity so aggressively that hard drive prices https://www.gadgetreview.com/hard-drives-are-sold-out-for-2026-and-ai-data-centers-are-to-blame have surged 46% since late 2025, with lead times stretching to 12 months or more. Western Digital has said https://www.zmescience.com/future/hard-drive-prices-have-surge/ nearly all of its production for the year is already spoken for, with contracts with hyperscale customers extending to 2028 and 2029 https://www.tomshardware.com/pc-components/ssds/crushing-shortages-have-pushed-long-term-supply-agreements-for-ssds-and-hdds-to-record-five-years-large-customers-are-signing-large-contracts . The major drive manufacturers are racing to meet demand with next-generation technology: Seagate’s HAMR platform is now shipping drives up to 44TB https://www.techedt.com/seagate-pushes-hamr-hard-drives-toward-the-next-phase-of-ai-scale-storage , qualified by multiple hyperscale cloud providers, while Western Digital announced https://www.businesswire.com/news/home/20260203315373/en/Western-Digital-Accelerates-Storage-Innovation-for-AI-Era drive architecture innovations at its Innovation Day 2026 that demonstrate a current 2x throughput improvement with a roadmap path to 4x and eventually 8x. Backblaze’s fleet is responding to the same pressures. At the end of 2025, drives 20TB and larger made up nearly 23% https://www.backblaze.com/blog/backblaze-drive-stats-for-2025/ of the fleet by count. And because of their size relative to the legacy drives they’re replacing, they represent a significantly greater share of total storage capacity than that number suggests. Backblaze deployed its first 26TB drive in Q4 2025. Migrating from 8TB drives to 24TB drives means storing the same data in a third of the rack space. Larger drives also change how data protection works. Backblaze uses Reed-Solomon erasure coding, breaking files into data and parity shards distributed across the fleet. Bigger drives mean more parity shards are needed, and those shards have to be spread carefully to avoid ending up on the same server rack. More density means more complexity in how data is protected and reconstructed. Drive Stats has been tracking this fleet for 13 years across 557 million drive-days the sum of every drive running for every day across the entire fleet . That accumulated data is what lets you tell the difference between a bad quarter and a real trend, or between a drive model that’s aging out and one with a genuine reliability problem. The 2025 annual failure rate came in at 1.36%, the lowest since 2022, across a fleet actively transitioning to meet the storage demands of a new era. Use data to select your infrastructure providers Price per terabyte can’t tell you how a system holds up under the pressure of AI workloads. AI traffic arrives in bursts, runs in parallel, and punishes any weak link in the pipeline–often while GPUs sit idle. Backblaze’s Network Stats https://www.backblaze.com/blog/network-stats-for-q1-2026-neocloud-traffic-trends/ , Performance Stats https://www.backblaze.com/blog/backblaze-performance-stats-q1-2026/ , and Drive Stats https://www.backblaze.com/cloud-storage/resources/hard-drive-test-data can–and we publish all three in full, including the results that are unflattering for us. Use these as a starting point and test against your own stack. Ready to match your workload in real numbers with a proof of concept? Contact our cloud storage experts. https://www.backblaze.com/contact-sales