← learn

storage · business · market

VAST, WEKA and NetApp: three ways to feed GPUs, and three businesses

The three storage companies most often shortlisted for AI clusters were built on different bets: a parallel file system, a disaggregated shared-everything design over NFS, and the enterprise's existing file store. What each is good at, where it lags, what the evidence says, and what their businesses look like.

Before you start

This article assumes you know:

Ideas it leans on:

RDMA
· remote direct memory access: a network card writing straight into another machine's memory.
GPUDirect Storage
· storage or a network card copying data straight into GPU memory, skipping the CPU's memory.
Checkpoint
· a saved copy of the weights and optimizer state, so a run can resume after a failure.
Throughput vs capacity
· how fast storage can deliver data (GB/s) versus how much it holds (TB); the two are bought separately.

Why storage decides how busy the GPUs are

A GPU only earns its keep while it computes. Storage decides how often it waits. In a training run, storage does three jobs: it holds the prepared data, it feeds batches to every GPU, and it takes the checkpoints that let a run survive failures. In serving, two more have appeared: loading model weights when new replicas start, and holding KV caches that no longer fit in GPU memory (see the training and inference walk-throughs for where each sits).

Scale decides whether any of this matters. Our own GPT-2 run on one 8-GPU server read a 20 GB dataset over a few hours; a local SSD would have done. At thousands of GPUs it changes: Meta reports 419 unexpected interruptions in 54 days of training Llama 3 405B on 16,384 GPUs (Llama 3 paper), and every restart reads back a checkpoint of several terabytes while the whole cluster waits. That is the market these three companies fight over, and it is why buyers price storage two ways: per terabyte stored and per gigabyte per second delivered (throughput vs capacity).

Three bets

WEKA bet on a parallel file system written for flash: software that runs on ordinary servers and a client on every GPU server that reads from all storage servers at once (WEKA docs).

VAST Data bet that cheap, dense QLC flash could replace tiers of disk, and built a disaggregated shared-everything design: stateless front-end servers that can all reach every SSD, spoken to over NFS (Blocks & Files, 2019; VAST).

NetApp is the incumbent: scale-out NAS (network-attached storage) that already holds a large share of enterprises’ files, now being rebuilt in disaggregated form for AI (AFX, Novus).

Same words, every article

This article uses the same terms as the rest of /learn: RDMA and GPUDirect Storage for the data path (explained here), a server is one machine, and GB/s measures bytes while Gb/s measures network links.

How each is built

Draw all three with the same blocks: GPU servers, a network, the storage software, its metadata (the index of which file lives where) and the flash. The differences are where the metadata lives and what the GPU server has to run. (Protection works differently too: VAST, for example, uses very wide erasure-coding stripes, and ONTAP spreads one namespace across controllers with FlexGroup.)

Four designs, same building blocks
computeGPU serverNFS client (in-box or VAST’s)GPU serverNFS client (in-box or VAST’s)GPU serverNFS client (in-box or VAST’s)GPU serverNFS client (in-box or VAST’s)network · RDMA over InfiniBand or Ethernet (RoCE)storagefront-end serverstatelessfront-end serverstatelessfront-end serverstatelessfront-end serverstatelessfront-end serverstatelessinternal fabric · NVMe over fabrics: every front-end server reaches every SSDflash enclosurewrite buffer + QLCflash enclosurewrite buffer + QLCflash enclosurewrite buffer + QLCsame data as NFS, SMB, S3 and tables
Design
Disaggregated shared-everything
Data path
GPU servers mount it over NFS; plain mounts use the in-box client, while multipathing and GPUDirect Storage use VAST’s own build of the kernel NFS client. Any front-end server can answer any request.
Metadata
Kept on the shared flash, so the front-end servers hold no state of their own.
Scales by
Add front-end servers for speed and flash enclosures for capacity, separately.
Simplified: real systems have many more servers and redundant networks. Sources: VAST: how it works · VAST KB: NFS multipathing · VAST KB: GPUDirect Storage guide

How a GPU server reads from each

All three support RDMA and GPUDirect Storage, so all three can put file data straight into GPU memory without the host-RAM detour (how that works). The real difference is the client. NetApp works with the NFS and pNFS client already in Linux, spreading a mount over several connections with session trunking (NetApp docs). VAST works with the in-box client for plain NFS, but its GPUDirect Storage guide installs VAST’s own NFS client, a build of the kernel client that adds multipathing (VAST). WEKA’s top speed needs its own client, which runs in user space and, in its performance mode, takes dedicated CPU cores on each GPU server (WEKA networking docs). That buys parallelism without waiting on the operating system, at the price of software to install, update and support on every GPU server.

What runs on the GPU server

VAST Data

  1. front-end servers→
  2. NFS over RDMA, multipath→
  3. set up by VAST’s NFS client (a build of the kernel client)→
  4. NIC→
  5. GPU memory (GDS)

Plain NFS mounts work with the in-box Linux client; VAST’s GPUDirect Storage guide installs VAST’s own NFS client, a kernel-NFS build rather than a user-space client like WEKA’s. source

WEKA

  1. storage servers→
  2. RDMA or DPDK→
  3. set up by the WEKA client (user space)→
  4. NIC→
  5. GPU memory (GDS)

The client is what makes it a parallel file system; in its performance mode it dedicates CPU cores on every GPU server. source

NetApp

  1. controllers→
  2. NFS / pNFS over RDMA, session trunking→
  3. set up by the in-box NFS / pNFS client→
  4. NIC→
  5. GPU memory (GDS)

Standard clients throughout; Novus adds a metadata service the pNFS client asks first. source

With GPUDirect Storage the client only sets each transfer up; the bytes go from the network card straight into GPU memory. The difference is the software each asks you to run on the GPU servers, and so how much of their CPU it takes and how tied you are to the vendor.

Side by side

The same questions asked of each, with the source behind every answer.

Capabilities, as of 20 Oct 2026
CapabilityVAST DataWEKANetApp
ArchitectureHow the storage is builtDisaggregated shared-everything: stateless servers reach every flash shelf over a fabricVAST: how it worksParallel file system: data and metadata spread across every storage serverWEKA docs: system overviewScale-out NAS (ONTAP on controller pairs), now adding disaggregated AFX and NovusBlocks & Files: NetApp AFX and AI Data Engine (Oct 2025)
Standard Linux clientCan a GPU server read it without installing vendor software?partly: Plain NFS mounts, yes; multipathing and GPUDirect Storage use VAST’s own build of the kernel NFS clientVAST KB: GPUDirect Storage guidepartly: NFS, SMB and S3 work, but full speed needs WEKA's own client, which takes CPU cores on the GPU serverWEKA docs: networking (DPDK, RDMA, GDS)yes: Yes, the in-box NFS and pNFS clientNetApp docs: AIPod software (GDS, pNFS, RDMA)
RDMA and GPUDirect StorageBytes straight into GPU memory, no host-RAM detouryes: Over NFS with RDMA, through VAST’s NFS clientVAST KB: GPUDirect Storage guideyes: Through its client; GDS skips the page cacheWEKA docs: networking (DPDK, RDMA, GDS)yes: Over NFS with RDMA, plus pNFS and session trunkingNetApp docs: AIPod software (GDS, pNFS, RDMA)
Audited AI benchmarkMLPerf Storage or SPEC's AI workloadno: Not in any MLPerf Storage round, v0.5 to v3.0Blocks & Files: MLPerf Storage v1.0 (Sep 2024)partly: MLPerf Storage v0.5 (2023) and v1.0 (2024), SPEC AI_IMAGE (2025); skipped MLPerf v2.0 and v3.0SPECstorage Solution 2020 AI_IMAGE: HPE + WEKA (Jan 2025)no: Not in any MLPerf Storage round; its SPEC result is for chip-design (EDA) work, not AISPECstorage EDA_BLENDED: 8-node AFF A90 (2024)
Inference featuresHelp beyond feeding trainingyes: KV cache offload through NVIDIA Dynamo (tested with CoreWeave; speed-ups are VAST’s claim), plus InsightEngine retrievalVAST: KV cache offload with NVIDIA Dynamo and CoreWeave (Dec 2025)yes: Augmented Memory Grid: a KV cache tier outside GPU memory, available since Nov 2025 (speed-ups are WEKA’s claim)WEKA: Augmented Memory Grid available in NeuralMesh and on OCI (Nov 2025)partly: AI Data Engine for curating and searching data: previewed Oct 2025, first customers from Mar 2026StorageNewsletter: NetApp at GTC 2026 (AI Data Engine rollout)
Certified for NVIDIA DGX SuperPODOn NVIDIA's list of storage for its reference AI supercomputeryes: YesNVIDIA: DGX SuperPOD FAQ (certified storage)yes: YesNVIDIA: DGX SuperPOD FAQ (certified storage)yes: Yes: AFF A90 (ONTAP), and E-Series with BeeGFSNVIDIA: DGX SuperPOD FAQ (certified storage)
Where it runsAppliance, software, or a cloud servicepartly: Software sold on qualified hardware; a managed service on Google Cloud, and Azure announcedBlocks & Files: VAST's AI OS comes to Azureyes: Software on commodity servers, WEKApod appliances, and documented installs on AWS, Azure, Google Cloud and OCIWEKA docs: system overviewyes: Appliances, plus first-party services on AWS, Azure and Google CloudAWS: FSx for NetApp ONTAP performance
Latest Gartner positionTwo different reports, see the evidence sectionpartly: Leader, File and Object 2024; not in the 2026 Enterprise Storage reportBlocks & Files: 2026 Enterprise Storage Platforms MQpartly: Visionary, File and Object 2024WEKA: Visionary, 2024 Gartner MQ for File and Object (via AAP)yes: Leader, Enterprise Storage Platforms 2026Blocks & Files: 2026 Enterprise Storage Platforms MQ

● yes · ◐ partly or with conditions · ○ no

What the independent evidence says

Vendor numbers are everywhere; audited ones are scarce. Four kinds of outside evidence exist, and each says something different.

What this means for a buyer

There is no recent, audited, like-for-like comparison of these three. Treat every throughput figure in their marketing, including the ones quoted below, as a claim to test on your own hardware and data.

Where each came from, and where it’s going

VAST Data

Started
2016, New York and Israel. CEO Renen Hallak, an early XtremIO engineer, saw that companies kept tiers of storage only because flash was too costly and disk too slow, and bet that QLC flash plus data reduction could make one fast tier affordable (Blocks & Files).
Focus shift
From “one tier of flash” to an “AI operating system”: a database, retrieval and vector search (InsightEngine), agent services, and KV cache offload for inference through NVIDIA Dynamo (VAST) on top of the storage. Its biggest customers are AI clouds: a $1.17B CoreWeave deal and xAI’s Colossus cluster (The Next Platform).
Latest
February 2026: a multi-cloud control plane (Polaris), agent-governance services and GPU-equipped servers, announced rather than all shipping (SiliconANGLE). April 2026: about $1B raised at a $30B valuation, partly as shares sold by employees and early investors; the company says it exited its last fiscal year above $500M in committed recurring revenue and is free-cash-flow positive, and Bloomberg reports it is weighing an IPO (VAST, Bloomberg).
What it implies
VAST is selling a platform, not a file system, which raises what it can charge for and what a buyer depends on it for. Its largest deals are with a few AI clouds; The Next Platform calls CoreWeave its “Customer Number One” (The Next Platform). As a private company, its figures are unaudited.

WEKA

Started
2013, by Liran Zvibel, Omri Palmon and Maor Ben-Dayan, with R&D in Tel Aviv; it came out of stealth in July 2017 pitching “the world’s fastest distributed file system” (TOP500), rooted in high-performance computing.
Focus shift
From fast files to inference memory: in 2025 it relaunched its platform as NeuralMesh (WEKA) and brought out Augmented Memory Grid, which keeps KV caches outside GPU memory: unveiled in March, generally available in NeuralMesh and on Oracle Cloud since November (WEKA, WEKA).
Latest
July 2026: NeuralMesh 6 and an exabyte-scale WEKApod appliance, promised for the second half of 2026, adding multi-tenancy, one stack for files and S3 objects, and always-on data reduction with a contractual guarantee, according to WEKA (Blocks & Files).
What it implies
WEKA is filling enterprise gaps (sharing one system safely between tenants, shrinking data) while betting that inference memory becomes a storage job. On the last figures each disclosed it is the smallest of the three: valued at $1.6B in 2024 (WEKA), above $100M in recurring revenue then, aiming for $1B, after replacing its president, CFO and revenue chief in late 2024 and reshaping sales in early 2025 (Blocks & Files, Blocks & Files).

NetApp

Started
1992 as Network Appliance, founded by David Hitz, James Lau and Michael Malcolm to sell network-attached storage appliances (Wikipedia); public on Nasdaq as NTAP.
Focus shift
From appliances to enterprise data management across data centres and clouds: ONTAP runs as first-party services on AWS, Azure and Google Cloud (AWS, Azure, Google). For AI it added GPUDirect Storage over NFS with RDMA (NetApp docs).
Latest
October 2025: AFX, which separates ONTAP controllers from flash shelves (up to 128 controllers), and an AI Data Engine previewed for curating data (Blocks & Files), which reached a first wave of customers in March 2026 (StorageNewsletter). Mid-2026: it bought DataPelago (AI data infrastructure) and JetStream (cloud disaster recovery) (earnings call). October 2026: Novus, which moves metadata onto its own servers and serves data over standard pNFS. NetApp states it scales beyond 100 TB/s; Omdia, which NetApp had audit its testing, observed near-linear scaling and projected, rather than measured, that figure. Novus is orderable now, days after NetApp said it intends to buy PEAK:AIO, a pNFS metadata specialist (NAND Research).
What it implies
NetApp is rebuilding its core product for AI-scale throughput while selling to customers whose data is already on it. Revenue is growing fast again (next section), but the new designs are young: NAND Research notes that Everpure shipped a comparable disaggregated pNFS design first and says NetApp’s “weakest points are time and proof” (NAND Research).

The businesses, in numbers

These numbers are not like for like. NetApp reports audited revenue. VAST and WEKA are private and publish recurring revenue, bookings and round valuations on their own terms, so each kind gets its own chart.

Scale, on each company's own terms

Metric

NetAppFY26, all products
$6.93B

FY ended Apr 2026 ·NetApp FY26 results (via StorageNewsletter)

NetAppFY26, all-flash arrays
$4.2B

FY ended Apr 2026 ·NetApp FY26 results (via StorageNewsletter)

Only NetApp is public. VAST and WEKA are private and don't publish revenue; their own figures are in the next tab, on their own terms.

NetApp’s most recent quarter shows the AI effect: revenue of $2.03B, up 30% in a quarter that had an extra week, all-flash arrays up 47%, and guidance raised to about $8B for the year (NetApp). On the earnings call it counted about 350 AI and data-lake modernization wins, while saying it is hard to quantify how much revenue AI drives (transcript). The wider market is growing too: IDC puts external enterprise storage at $10.34B in the second quarter of 2026, up 33.6% (Blocks & Files), with NetApp third at a 9.6% share (IDC).

Rounds, launches and certifications
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
VAST Data
WEKA
NetApp

Point at a dot to see the event and its source.

Good at, and lags at

VAST Data

Good at

  • NFS on the GPU servers, with RDMA and GPUDirect Storage; GDS uses VAST’s build of the kernel NFS client, not a user-space client (VAST).
  • Speed and capacity scale separately, and very wide data-protection stripes (up to 146 data to 4 parity) waste little flash (Lockwood, a VAST employee).
  • Runs some of the largest AI clouds: CoreWeave and xAI’s Colossus (The Next Platform).
  • One system for files, objects and tables, with retrieval built in (VAST).

Lags at

  • No audited benchmark: it has not entered any MLPerf Storage round (v1.0, v3.0).
  • Sold as software but only on qualified hardware (Lockwood, a VAST employee).
  • Even a VAST strategist, in personal notes, calls it “neither the fastest nor the cheapest, but it does both reasonably well” (Lockwood).
  • Large deals with a few AI clouds (The Next Platform); financials unaudited.

WEKA

Good at

  • The most independent benchmark evidence of the three: MLPerf Storage (2023, 2024) and SPEC AI_IMAGE (2025) (SPEC).
  • A true parallel file system: each GPU server reads from every storage server at once (WEKA docs).
  • A KV cache tier for inference, available in NeuralMesh and on Oracle Cloud since November 2025 (WEKA).
  • Runs as software on commodity servers, as appliances, or in four clouds (WEKA docs).

Lags at

  • Its own client on every GPU server for full speed, taking dedicated CPU cores in its performance mode (WEKA docs).
  • Skipped the last two MLPerf Storage rounds, so its 2026 claims are unaudited (StorageReview).
  • Multi-tenancy and always-on data reduction arrive only with NeuralMesh 6, still to ship (Blocks & Files).
  • Smallest of the three on the last figures each disclosed, with leadership and sales changes in 2024 and 2025 (Blocks & Files).

NetApp

Good at

  • The enterprise’s data is often already on it, with mature data protection and management.
  • Standard NFS and pNFS clients, with RDMA and GPUDirect Storage (NetApp docs).
  • Native services in all three big clouds (AWS, Azure, Google).
  • Public, audited and growing: $6.93B revenue in FY26 (NetApp), a Gartner Leader (Blocks & Files).

Lags at

  • The classic design serves metadata and data from the same controllers, which AFX and Novus exist to fix (NAND Research).
  • No MLPerf Storage entry and no SPEC AI result (v1.0, v3.0).
  • Late to AI-native designs: Novus is only now orderable, and Everpure shipped a comparable design first (NAND Research).
  • Component costs are squeezing product margins: 54.6%, down 1.5 points in a quarter (transcript).

Which fits whom

Scale and starting point matter more than the logo. A rough guide, not a ranking:

SituationWhere to start looking
One server, fine-tuning or a lab like oursNone of these. Local NVMe plus an object store (S3, GCS) is enough.
An enterprise whose files already live on NetApp, tens to hundreds of GPUs (NVIDIA’s Foundation and Enterprise tiers)NetApp first: the data doesn’t move, and the GPU servers need nothing installed.
An AI cloud or a cluster of thousands of GPUs shared by many teams (NVIDIA’s Cloud Partner tier)StorageReview calls VAST, WEKA and DDN “the vendors winning the largest neocloud and AI factory deployments” (StorageReview). Novus is NetApp’s new entry.
Inference with long contexts, where KV caches spill out of GPU memoryWEKA (Augmented Memory Grid) and VAST (KV cache offload through NVIDIA Dynamo) both offer a KV cache tier; the speed-ups each quotes are its own (WEKA, VAST). Open-source pools such as Mooncake do the same job in the GPU servers’ own memory and SSDs.
A rule of no vendor software on GPU serversNetApp, on the in-box NFS and pNFS client throughout. VAST uses the in-box client for plain NFS, but its GPUDirect Storage setup installs VAST’s own kernel-NFS build.

They are not the only options: DDN, Everpure (Pure Storage renamed itself in February 2026; Channel Insider), IBM, Dell and Hammerspace compete for the same deals. Everpure entered MLPerf Storage v3.0, and DDN, IBM and Hammerspace entered v2.0 (v2.0, v3.0). Whatever the shortlist, ask each vendor to run your data loading and checkpoint pattern on hardware you can inspect.

Ask about this article

Something unclear, or something we got wrong? Ask with the button at the bottom right, or select any passage to flag it.

Sources

Checked 20 Oct 2026. Company-reported figures are labelled; Gartner positions come from vendor releases and press because the reports are paywalled.