# The New Math of AI Storage Performance **Author:** Boni Bruno **Published:** November 12, 2026 ![Black and white photo of a server rack unit with neatly bundled network cables inside a spacious data center.](https://cdn.sanity.io/images/ult5g8gw/production/4abbfc9657b4a1ca7e3763732243d7e2e96c4a9f-1376x768.jpg) ## TL;DR The first SPECstorage Solution 2020 AI_IMAGE benchmark of new WEKApod systems demonstrates what NeuralMesh can deliver in a dense, purpose-built storage system. - 9,000 concurrent AI_IMAGE jobs and 880 GB/s of throughput from just 16 rack units of storage. - 80% more concurrent jobs and 35% more jobs per rack unit than WEKA’s 2025 submission. - Consistent low latency, ranging from 1.0 to 1.3 milliseconds as concurrent jobs increased tenfold. _This blog was co-written by Boni Bruno, Principal Performance Engineer, WEKA; Radha Manga, Sr Director, Product Management, WEKA; and Rob Roy, Sr Engineer, WEKA._ **The New Math of AI Storage Performance** Every GPU you add needs storage that can feed more jobs and larger datasets without adding racks, network ports, power, cooling, or operational complexity at the same rate. Peak bandwidth by itself doesn't answer that challenge. What matters is how much concurrent AI work the platform can sustain, how consistently it responds as load increases, and how much infrastructure it takes to deliver that performance. That was the design goal behind WEKApod: get more AI work and active data out of every rack. The first independently audited results now measure that progress in two ways: absolute performance and how efficiently WEKApod delivers it. **The results at a glance** | Measure | WEKApod AI_IMAGE result | | --- | --- | | Concurrent AI_IMAGE jobs | 9,000 | | Aggregate throughput at peak load | 880.2 GB/s | | File operations at peak load | 3,915,514 operations/s | | Overall response time | 1.07 ms | | Average latency at 9,000 jobs | 1.3 ms | | Storage footprint | 16U of storage chassis | | Protected shared namespace | Single POSIX filesystem; 16+2 erasure coding | Source: WEKA November 12, 2026 SPECstorage Solution 2020 AI_IMAGE audited submission. **More Work From Every Rack** WEKApod pairs NeuralMesh's distributed architecture with purpose-built, high-density hardware. The result is a storage system engineered to deliver more performance and capacity from every rack. WEKA’s published AI_IMAGE results show a clear progression in independently measured concurrency: 2,400 jobs in the 2024 WEKA-on-AWS result, 5,000 jobs with 488.9 GB/s in the January 2025 HPE/WEKA result, and 9,000 jobs with 880.2 GB/s in the 2026 WEKApod submission. This sequence illustrates sustained gains across generations rather than a one-time peak. Compared with WEKA's January 2025 submission, WEKApod delivers **80% more concurrent jobs** and **35% more jobs per rack unit**. The earlier system supported 5,000 jobs from 12U. WEKApod supports 9,000 from 16U. That's 4,000 more concurrent jobs from just four additional rack units of storage. **More AI work from every rack. That's the new math in action.** ![SPECStorage AI Image Progression, 2024-2026](https://cdn.sanity.io/images/ult5g8gw/production/22d0fb35582f103994ef55c9f2a4d99e3971fd4d-2048x577.jpg) **Low latency as concurrency scales** Density matters. But packing more performance into less space only helps if the system can sustain it as demand grows. WEKApod ran ten SPECstorage AI_IMAGE workload levels, from 900 to 9,000 concurrent jobs. At 900 jobs, average latency was 1.0 millisecond. At 9,000 jobs, with throughput reaching 880 GB/s, average latency was 1.3 milliseconds. **Ten times the load. Just 0.3 milliseconds more latency.** ![SPECStorage 2020 AI Image Benchmark - Low Latency at Scale](https://cdn.sanity.io/images/ult5g8gw/production/23d33e11a7f860b2bae07de7b56a84b5a030a906-1600x840.jpg) These figures are per-load-point average latencies and are distinct from the benchmark’s reported overall response-time metrics. They demonstrate that WEKApod can sustain very high concurrency while maintaining consistently low storage latency under the same workload conditions. **Protected, shared performance** SPECstorage AI_IMAGE models the intensive file operations of AI image-processing pipelines, using traces collected from NVIDIA DGX-based systems running COCO, ResNet-50, and Cityscapes processing. It measures how storage responds as increasing numbers of concurrent jobs access a shared dataset. WEKApod delivered its result from a single shared NeuralMesh POSIX filesystem with 16+2 erasure coding enabled, driven by 40 load generators over a 400GbE Ethernet fabric. The tested system consisted of 31 independent storage servers across eight 2U WEKApod chassis, with 1.8 PiB of usable capacity. This wasn't a test of isolated servers or unprotected capacity. It demonstrates how much concurrent AI work a fully protected, distributed storage system can support in 16U. **The New Math of AI Storage** The new math of AI storage isn't just about how fast a system can go. It's about how much useful work it can sustain, how predictably it responds, and how much infrastructure it takes to get there. WEKApod's latest SPECstorage AI_IMAGE results extend that efficiency story with independently audited measurements: 9,000 concurrent jobs, 880.2 GB/s of throughput, and approximately 900 TB of active benchmark data from just 16U of protected shared storage. If you're building AI infrastructure for growing workloads on finite power and space, that's the equation that matters. **Learn more** [View SPECstorage Solution 2020 results](https://www.spec.org/storage2020/results/ai_image/) [Explore WEKApod](/product/wekapod)