If you rent GPUs in one public cloud, buy none of these four: your provider's managed Lustre covers the write bandwidth training needs, at a published price. For a cluster you own or colocate and use mainly for training, start with DDN EXAScaler. Pick WEKA if you have no Lustre operators, FlashBlade//EXA if audited checkpoint numbers matter most, and VAST only when the storage must also be your enterprise data platform, because VAST has never put a system through MLPerf Storage.
To compare like for like, every number below is applied to one workload: 1,024 GPUs in 128 eight-GPU servers, training a dense 405B-parameter model, checkpointing every 30 minutes, mostly text data. Prices and certifications were checked on each vendor's or cloud's own pages on November 11, 2026.
VAST vs WEKA vs DDN vs Pure at a Glance
| DDN EXAScaler (AI400X3) | WEKA NeuralMesh | VAST Data AI OS | Everpure FlashBlade//EXA | Cloud managed Lustre | |
|---|---|---|---|---|---|
| Latest audited MLPerf Storage result | v2.0 (Aug 2025) | v1.0 (Sep 2024) | None, ever | v3.0 (Sep 2026) | Azure Managed Lustre in v3.0 |
| Architecture | Lustre: data and metadata on separate server tiers | Distributed data and metadata across all nodes | Disaggregated: stateless CNodes, shared flash enclosures | Disaggregated: FlashBlade metadata core, commodity data nodes over pNFS | Lustre, run by the provider |
| NVIDIA status (as each vendor publishes it) | GB300/B300 NCP reference designs | DGX SuperPOD (H100), GB300 NCP reference design | DGX SuperPOD with B300/GB300 | FlashBlade//S R2 is SuperPOD-certified; EXA validation planned | n/a |
| Public price | None (Google sells it as Managed Lustre) | None | None | None (Evergreen//One subscription offered) | $0.145 to $0.60 per GiB-month |
| Pick it if | Training-first cluster, Lustre skills in house | Mixed small-file workloads, no Lustre team | Storage is also the shared enterprise data platform | You want the best audited checkpoint numbers | You rent GPUs in one cloud |
| Skip it if | Nobody on staff can tune Lustre | You can't give up client CPU cores | You're buying on throughput claims | You need a long field record | You own the GPUs or span clouds |
The loser on evidence is VAST. It has a $1.17 billion CoreWeave deal and no audited throughput number from any MLPerf Storage round, so the burden of proof sits entirely with you.
Who Has an Audited MLPerf Storage Result, and Who Doesn't
MLPerf Storage is MLCommons' peer-reviewed benchmark that simulates GPUs reading training data and writing checkpoints against a real storage system, then publishes results only after other submitters audit them. It is the only public, like-for-like throughput record for this category, and three of the four vendors here skipped the latest round.
The v3.0 round published September 1, 2026 drew 19 organizations and added KV cache and vector database tests alongside the existing training and checkpointing workloads. StorageReview's write-up of v3.0 reports that VAST Data, WEKA and DDN "sat v3.0 out", and Blocks & Files lists DDN, Dell, Huawei, IBM, NetApp, VAST Data and WEKA among the absent AI storage suppliers. StorageReview adds a fair caution: skipping a round "is not evidence that a system is slow."
Here is each vendor's record by round, per StorageReview's audited-results table, updated October 29, 2026:
- Everpure (the company formerly called Pure Storage) first submitted in v3.0. FlashBlade//EXA wrote 877.5 GiB/s on the 1.25T-parameter checkpoint test with 30 data nodes, the top checkpoint write in the round.
- DDN submitted to v2.0. Its August 4, 2025 release reports a single 2U AI400X3 at 30.6 GB/s read and 15.3 GB/s write, a Llama3-8B checkpoint saved in 7.7 seconds, and over 120 GB/s of multi-node read on the 3D U-Net training test.
- WEKA submitted to v1.0 in September 2024 and has not submitted since. Its own commentary on that round leans on modeled five-client projections that go beyond what it submitted, and the same article concedes that single-client and multi-client results can't be compared.
- VAST Data has never submitted.
Two cautions before you put these on a slide. v3.0 results "cannot be compared against v2.0 results" because the simulated accelerator changed from H100 to B200, per StorageReview. And the units differ: StorageReview's v3.0 figures are in GiB/s, DDN's release reports GB/s.
How Much Storage Throughput Does GPU Training Actually Need?
For text-heavy LLM training, the reference workload needs roughly 30 GB/s of checkpoint write if it checkpoints synchronously, and about 3 GB/s if it checkpoints asynchronously. Both figures sit far below the TB/s numbers on vendor slides. Data loading barely registers.
A checkpoint is the saved model weights and optimizer state, written periodically so a failed job can restart from it instead of from zero. Its size scales with the model, never the cluster. Glenn Lockwood's analysis of checkpoint bandwidth assumes 14 bytes per parameter, so a 405B model writes about 5.7 TB per checkpoint. If training pauses while that write happens and you want the pause under 10% of a 30-minute interval, the write must finish in 180 seconds: about 32 GB/s. With asynchronous checkpointing, where GPUs dump state to host memory or local NVMe and the shared file system absorbs it in the background, the shared tier has the full 30 minutes, which works out to about 3 GB/s.
The same analysis covered more than 85,000 checkpoints from 40 production runs of 45B to over 1T parameters across 18 AI cloud clusters. Drains to global storage ran at 50 to 200 GB/s, almost every run kept median checkpoint overlap under 10% of training time, and Lockwood concludes trillion-parameter models train efficiently on "well under 1 TB/s." Weigh the source: Lockwood works for VAST, which published the analysis and has the most to gain from buyers discounting peak bandwidth. The arithmetic is still checkable with your own parameter count.
Data loading is smaller still for text. Lockwood's notes on storage for LLM training put tokenized text at "a few terabytes per trillion tokens." NVIDIA's Rubin DGX SuperPOD storage guidance sizes its Standard tier for LLM jobs whose datasets fit in the compute nodes' memory cache, at 1.25 GB/s of read per node. Its Enhanced tier, for multimodal training, is 3.91 GB/s per node, and only high-resolution computer vision gets 4 GB/s per GPU.
Applied to 128 nodes, NVIDIA's guidance gives 160 GB/s of read at Standard and 500 GB/s at Enhanced. Its B200 SuperPOD page says to plan at least half the read figure as write for LLM workloads, so 80 GB/s of write at Standard. That is a margin above the checkpoint math, and it is a sensible ceiling for a spec. The same page tells buyers to "characterize workloads and organizational needs before finalizing performance and capacity requirements."
Against those targets: by DDN's single-appliance write figure, two or three AI400X3 units cover the synchronous checkpoint and five or six cover NVIDIA's Standard write guideline. Everpure's smallest audited configuration, 10 data nodes, wrote 327.6 GiB/s, roughly ten times the synchronous checkpoint requirement. Any of the four can meet this workload, so the choice comes down to how each one fails, what it costs and how much evidence stands behind its claims.
DDN EXAScaler: The Default for a Training-First Cluster
DDN EXAScaler is the right default for an owned cluster that mostly trains, because it pairs a recent audited result with the most portable file system of the four. EXAScaler is DDN's distribution of Lustre, the open-source parallel file system most large HPC sites run. Google's Managed Lustre is built on EXAScaler, so a team that learns it on premises can burst or migrate to Google without rewriting its data layout.
The failure mode is operational. Lustre keeps metadata on separate metadata servers, and small-file or metadata-heavy workloads hit those first. File layout also has to be tuned by hand: AWS's own FSx for Lustre performance guide spends pages on stripe counts and warns there is "no single optimal layout configuration for all use cases." On NVIDIA certification, DDN publishes GB300 NVL72 and B300 reference designs for NVIDIA Cloud Partners, and the page's wording ranges from "certified for HGX compatibility" in the body to "certified for DGX SuperPOD compatibility" in its FAQ. Ask for the NVIDIA document, not the blog.
Don't buy DDN if nobody on your team has run Lustre in production. You'll pay for the throughput and lose it to metadata contention and default stripe settings.
WEKA NeuralMesh: The Pick When Nobody on Staff Runs Lustre
WEKA NeuralMesh fits teams that need parallel file system performance on mixed workloads, many small files included, without staffing a Lustre specialist. It spreads data and metadata across every node and erasure-codes stripes across failure domains, and WEKA says rebuilds enlist every core in the cluster, so recovery gets faster as the cluster grows. Nebius runs it as the premium storage tier of its AI cloud, and Andromeda standardized on it across roughly 50 GPU providers in July 2026, per StorageReview.
Two costs are easy to miss. WEKA's client takes dedicated CPU: one physical core by default, its highest-performance networking mode requires dedicated cores, and a Lenovo design guide notes those cores can't be used for computation. The certification record is also a generation behind: WEKApod was certified for DGX SuperPOD with DGX H100 in March 2024, and WEKA's GB200 and GB300 NVL72 design is described as "validated" for NVIDIA Cloud Partner storage, not certified. The newest audited throughput number is two years old. WEKA's July 2026 launch of NeuralMesh 6 and a third-generation WEKApod claims 10.2 TB/s per rack and 1.1 EB effective capacity, and none of it has been through MLPerf.
Don't buy WEKA if your GPU servers' CPUs are already saturated by data preprocessing, or if your procurement rules require an audited benchmark from the current hardware generation.
Everpure FlashBlade//EXA: The Best Audited Checkpoint Numbers
FlashBlade//EXA is the pick when audited checkpoint and KV cache numbers carry the most weight in your decision. It splits the system in two: an existing FlashBlade acts as the metadata core, and capacity sits on separately scaled QLC data nodes reached over an RDMA fabric, with clients using pNFS. Everpure chose off-the-shelf servers for the data nodes because customers said they already owned that hardware.
That design is also where it can fail. The data nodes are commodity Linux servers that you rack, patch and replace, and the published architecture coverage says little about how data-node loss is protected. Ask for the failure-domain design in writing. On NVIDIA, Everpure's GB200 and GB300 DGX SuperPOD certification is for FlashBlade//S R2, and as of March 17, 2026 Everpure only planned to extend NVIDIA-Certified Storage validation to EXA.
Don't buy EXA if your integrator or NVIDIA support contract requires a SuperPOD-certified storage product today, or if you want a platform with several years of field failures behind it.
VAST Data: The Platform Pick With No Audited Number
VAST is the pick when one system must serve training, inference, NFS home directories, S3 and a database for many tenants, and the loser when you're buying training throughput. Its architecture separates stateless CNodes, which run the protocols, from shared DBox flash enclosures whose media every CNode can reach, so losing one reduces protocol capacity while the data stays reachable. CoreWeave signed a $1.17 billion commercial agreement in November 2025 for VAST storage and software in its AI cloud, and VAST says it is certified for the DGX SuperPOD reference architecture with B300 and GB300 systems (May 13, 2026).
The failure mode is rack loss. VAST's own engineering post on rack-level resilience says the standard layout tolerates four simultaneous SSD failures but "does not protect overall cluster availability if an entire rack - and thus its DBoxes - goes offline." The newer DBox-HA layout survives one DBox plus two more SSD failures. Stock NFS clients connect at low performance; multipath and RDMA throughput need enhanced NFS clients, per Lockwood's notes on VAST, which also describe it as "neither the fastest nor the cheapest."
Don't buy VAST on a throughput claim alone. If you shortlist it, run the open MLPerf Storage benchmark on its proof-of-concept hardware yourself.
The Inference and KV Cache Claims Are Mostly Unaudited
Every vendor now pitches its storage as GPU memory extension for inference, and only Everpure has an audited number for it. A KV cache is the attention state a model computes for a prompt; offloading it to storage lets a server reuse it instead of recomputing it. In MLPerf v3.0, FlashBlade//EXA posted 1,623.4 GiB/s on the KV cache test with 51 hosts, per StorageReview.
WEKA's Augmented Memory Grid claims 10x token throughput on Oracle Cloud H100s and 4.2x more tokens per GPU at CoreWeave, both against a DRAM baseline, per Blocks & Files. Those are vendor claims with no audited counterpart. VAST and DDN both market inference data platforms and submitted nothing to the v3.0 KV cache test. If inference reuse is part of your business case, make the vendor run the MLPerf KV cache workload on your model sizes.
What GPU Cluster Storage Costs When Nobody Publishes a Price
None of the four publishes a rate card, so the only public price for any of these software stacks is Google's rate for EXAScaler sold as a managed service. Everything else is a private quote.
What is on the record, checked November 11, 2026:
- DDN EXAScaler via Google Cloud Managed Lustre: $0.145, $0.21, $0.34 and $0.60 per GiB per month for the 125, 250, 500 and 1,000 MB/s-per-TiB throughput tiers, per Google's pricing page and its launch post.
- VAST: the CoreWeave deal is $1.17 billion with capacity and term undisclosed, so it gives no per-PB figure.
- Everpure: Evergreen//One now covers FlashBlade//EXA as a pay-as-you-go subscription. No rate was disclosed.
- WEKA: no price. Its NeuralMesh 6 upgrade is free for existing customers, and WEKA says the QLC flash in its mixed TLC/QLC "AlloyFlash" systems costs roughly 30 to 40% less per TB, per Blocks & Files.
We found no public-sector contract record with a disclosed price per petabyte for any of the four. Negotiate on two numbers written into the order: dollars per usable TB, and dollars per GB/s of sustained write at your node count.
When to Buy None of Them: FSx for Lustre and Managed Lustre
If your GPUs are rented in one cloud region, the managed Lustre service in that cloud meets the reference workload's needs, and you avoid a multi-year hardware commitment against a GPU lease that may be shorter. Azure Managed Lustre wrote 642.2 GiB/s on the MLPerf v3.0 checkpoint test from a 4,096 TiB deployment with 128 clients, an audited result, per StorageReview. That is twenty times the reference workload's synchronous checkpoint requirement.
Costed on Google's 1,000 MB/s-per-TiB tier, NVIDIA's 160 GB/s Standard read target needs about 160 TiB, which is 163,840 GiB at $0.60, or about $98,300 a month. Google's performance tiers page caps that tier at about 9.5 PiB per instance. On AWS, Amazon FSx for Lustre pricing uses $0.145 per GB-month for persistent SSD storage at 125 MB/s per TiB in its US East example, and an EFA-enabled file system with GPUDirect Storage can drive up to 1,200 Gbps to one client, per the FSx performance guide.
Don't use the cloud service if you own the GPUs, if training spans clouds, or if your capacity need is large and stable enough that the monthly bill exceeds a three-year amortized system. Our GPU cloud comparison covers which neoclouds bundle which storage.
The Decision: Six Questions That Predict Regret
The regret in this category comes from buying peak throughput you never use and discovering the operational cost later. These questions decide it:
- Do you own the GPUs? If not, use your cloud's managed Lustre and stop here.
- What is your real checkpoint write requirement? Parameters × 14 bytes ÷ (checkpoint interval × 10%). For most clusters it is tens of GB/s.
- Who will run it? Lustre skills point to DDN. No storage specialists points to WEKA or VAST.
- Is this file system training-only? If it must also be the company's NFS, S3 and analytics store, VAST's case strengthens; if it's a training scratch tier, it weakens.
- What evidence will you accept? An audited MLPerf result from the current round, or your own MLPerf Storage run on proof-of-concept hardware. A vendor slide is neither.
- What happens when a rack loses power? Ask each vendor to show it in the proof of concept.
What changes the answer: moving to multimodal or video training roughly triples NVIDIA's read guidance, which favors the systems with audited multi-node results. A large inference footprint with KV cache reuse favors whichever vendor will run the MLPerf KV cache test on your models.
This Week:
- Compute your checkpoint requirement with the formula above and compare it with the throughput each vendor is quoting.
- Ask each shortlisted vendor for its MLPerf Storage submission by round, or a written reason it has none.
This Month:
- Run the open MLPerf Storage benchmark on each proof-of-concept system at your real client count.
- Pull power on one enclosure or rack during the test and time the recovery.
Before You Sign:
- Write dollars per usable TB and dollars per GB/s of sustained write into the order, with the benchmark result as an acceptance test.
- Price the cloud managed Lustre alternative for the same term, so the hardware quote has a public number to beat.
The Bottom Line on GPU Storage
Storage buying for GPU clusters looks like the scale-out NAS buying of a decade ago: capable products, spec sheets that quote peak numbers, and buyers who rarely run at the peak. This time an audited benchmark exists, and three of the four vendors skipped its latest round. Spec the checkpoint write you need, demand an MLPerf Storage run at that number, and price your cloud's managed Lustre before you sign a hardware quote. If the GPUs themselves are on a lease, check the term against our reserved vs spot GPU break-even so the storage contract doesn't run longer than the compute.
Continue Reading
- Reserved vs Spot GPUs: Commit Six Months, Not Twelve
- Clockwork Raises $31M as LinkedIn Adopts Its GPU Failover
- GPU Clouds Compared: Nebius on Price, CoreWeave at 3AM
- NetApp to Acquire PEAK:AIO to Scale AI Storage in ONTAP
- When the Colo's Chillers Fail, the GPU Loss Is Yours
- Dropbox's $350M Tigris Buy Targets AI Storage and Agents