Black Box analysis. Published 2026-10-28. Based on Scott Sinclair’s TechTarget analysis of AWS’s re:Invent 2023 storage announcements, published December 1, 2023.
AI Amazon FSx for NetApp ONTAP AWS 2026-10-28What AWS announced
TechTarget grouped two re:Invent 2023 announcements around the same constraint: data-intensive applications need storage that can sustain their access pattern. Amazon S3 Express One Zone was the object option, described as offering ten times the performance of S3 Standard, hundreds of thousands of requests per second and consistent single-digit-millisecond request latency. Scale-out Amazon FSx for NetApp ONTAP was the file option, described as delivering as much as nine times the storage performance of the existing file systems while retaining ONTAP data-management capabilities.
Those figures are vendor claims relayed by TechTarget, not Black Box benchmarks. They also describe different interfaces and failure models: an object storage class in one Availability Zone is not interchangeable with a managed ONTAP file system. Protocol, namespace, durability design and application behavior still determine which service fits.
Object speed and file-system scale solve different problems
| Question | S3 Express One Zone | Scale-out FSx for ONTAP |
|---|---|---|
| Application interface | Object API | Managed ONTAP file storage |
| Announcement emphasis | Low request latency and very high request rate, including many small objects | Higher file-system performance with ONTAP data services |
| Placement implication | Single Availability Zone is part of the named storage class | Topology and availability must be selected and validated for the actual FSx deployment |
| First design test | Can the application use object semantics and the intended resilience model? | Do protocol, throughput, capacity and HA choices match the pipeline? |
The distinction prevents a common AI-infrastructure mistake: choosing from a headline performance number before drawing the data path. The site’s FSx for ONTAP reference sets out the managed-service boundary, while the protocol-choice guide helps separate file, block and object requirements.
The ONTAP-side implication: keep the pipeline fed and recoverable
TechTarget’s analysis correctly broadens the workload beyond model training. Data preparation and cleansing can dominate an AI project, and those stages can create a mixed profile: directory walks and small-file metadata operations, large sequential reads, concurrent writers, checkpoints and repeated scans of a curated corpus. A peak-throughput claim alone does not establish that the whole pipeline will perform.
For the ONTAP layer, the real requirement is a measurable service envelope: client protocol and mount behavior, aggregate and per-client throughput, latency distribution, metadata rate, working-set size, concurrency and recovery objective. Operators also need to know where snapshots, clones or replicas enter the workflow so that a training run or retrieval index can be reproduced from a consistent data state. This is an architectural implication, not a claim that the 2023 announcement added a new ONTAP consistency feature.
What to validate before calling it AI-ready
- Replay the real access pattern. Test the expected file-size distribution, directory depth, concurrency and read/write mix; do not substitute a single large-file bandwidth run.
- Measure the clients as well as storage. Network path, instance type, protocol settings and client parallelism can cap observed performance before the file system does.
- Define the authoritative copy. Record where source data is governed, how a consistent dataset is created, and which copy a pipeline may modify.
- Test interruption and restart. Checkpoint placement, retry behavior and recovery time matter because an expensive compute fleet can sit idle while data access recovers.
- Price the entire data path. Include provisioned storage and throughput, requests, snapshots, replicas, temporary copies, transfer and retained model artifacts.
The performance hub provides the ONTAP-side investigation map. No CLI command is prescribed here because FSx for ONTAP is managed and the correct diagnostic or configuration action depends on the selected deployment and current AWS documentation.
What the 2023 story does—and does not—establish
The story establishes that AWS was expanding both low-latency object storage and scale-out managed ONTAP file storage for demanding workloads, with AI, machine learning, media, life sciences and other data-intensive uses named as examples. It does not prove that either service is automatically faster or cheaper for a specific dataset, and it does not collapse object and file storage into one design.
Bottom line: the ONTAP contribution is more than capacity. It is a governed file data path whose throughput, metadata behavior, consistency point and recovery characteristics have to keep expensive compute useful. Scale-out raises the available performance ceiling; workload evidence determines whether the pipeline can reach it.
Source: TechTarget — “Amazon and NetApp’s new data storage options optimized for AI” (Scott Sinclair, December 1, 2023). Performance figures above are attributed to the announcement as reported by TechTarget.