Feeding the GPUs: File Storage for AI and Cloud-Native Workloads on Azure

Hey People!

If you’re operating AI workloads on Azure, you’ve in all probability discovered the exhausting manner that the flawed storage selection can depart a rack of very costly GPUs sitting idle, ready for knowledge. On this session through the Microsoft Azure Infra Summit 2026, Wolfgang de Salvador and Reena Shah from the Azure Storage workforce walked by way of how Azure Managed Lustre and Azure Information map to the distinct phases of the AI pipeline, and why choosing the right file system per stage is among the highest leverage choices you’ll make.

📺 Watch the session:

You in all probability didn’t get into IT to babysit checkpoint writes or debug Hugging Face egress payments at 2 a.m. However that’s precisely the form of work that lands in your plate when storage is just not matched to the workload. Right here is why MAIS28 issues for the IT execs, platform engineers, and Azure architects within the room:

• GPU time is the most costly compute you’ll ever purchase. Gradual knowledge loading and sluggish checkpoints flip that into burned money.
• AI workloads usually are not one workload. Knowledge prep, coaching, fine-tuning, and inferencing every have a distinct storage profile.
• Cloud-native AI on AKS and Azure Container Apps lives or dies on the ReadWriteMany expertise. If mannequin loading is sluggish or shared mannequin caches don’t exist, each chilly begin re-downloads lots of of gigabytes.
• Storage selections ripple into safety and compliance. Encryption in transit, redundancy, and snapshots usually are not elective in 2026.

In brief, this session is for anybody who has to reply the query, “What persistent quantity ought to we use for this AI workload?” and needs a defensible reply.

Wolfgang opened with the storage profile of each stage of an AI workflow. Knowledge preparation wants lots of of petabytes at the most effective TCO (suppose Azure Blob Storage because the sturdy core). Coaching and fine-tuning want excessive throughput so GPUs keep fed throughout knowledge loading and so checkpoint writes full quick. Inferencing wants quick mannequin masses, low-latency KV cache, and grounded knowledge for RAG. One filesystem doesn’t match all of these without delay, and making an attempt to make it match is the place groups overspend.

Azure’s reply is a tiered, file-based portfolio that strains up with these phases:

• Azure Managed Lustre (AMLFS) is a totally managed, accelerator-tier filesystem. It scales to 25 PB of capability and as much as 512 GB/s of throughput, integrates with Azure Blob Storage because the sturdy core, and exposes a regular Lustre shopper plus a CSI driver for AKS.
• Azure Information is the pure ReadWriteMany selection for cloud-native AI on AKS and Azure Container Apps. It tops out at 256 TB of capability and 10.4 GB/s of throughput, provides LRS and ZRS redundancy with snapshots and gentle delete, and ships with a 99.99% SLA.
• Azure Blob Storage sits beneath each of those as a budget, sturdy core for knowledge prep and long-term retention.

The psychological mannequin the audio system used is “accelerator and core”. Blob is the core, sturdy and economical. AMLFS is the accelerator for coaching. Azure Information is the accelerator for inferencing and shared state. Decide the appropriate pair for the stage you’re operating.

For coaching, Wolfgang confirmed the demo most folk got here to see: a 32 x H100 ND H100 v5 AKS cluster deployed from the Azure AI Infrastructure repository, operating a 30B-parameter GPT-3 coaching job backed by AMLFS. Two issues matter right here.

First, AMLFS absorbs checkpoint write bursts. When 32 H100s flush state on the similar time, you want a filesystem that may take the punch with out stalling. AMLFS does, which retains GPU utilization drops quick and contained.

Second, the AMLFS Lustre CSI driver for AKS helps each static and dynamic provisioning with availability-zone placement, and there are 5 SKU tiers from MLFS20 (least expensive by capability) to MLFS500 (least expensive by bandwidth). Which means you’ll be able to choose a cost-performance level that matches your coaching funds as an alternative of shopping for the highest SKU and hoping for the most effective.

For inferencing, Reena’s half of the session was simply as sensible. 5 causes Azure Information suits AKS ReadWriteMany workloads:

• Customary Kubernetes RWM quantity over NFS or SMB.
• 256 TB capability ceiling and as much as 10.4 GB/s of throughput per share.
• LRS or ZRS redundancy with snapshots and gentle delete for cover.
• 99.99% SLA so it reveals up in your availability math.
• Native help throughout AKS and Azure Container Apps, together with serverless GPU.

The headline characteristic is Azure Information Provisioned v2. Within the outdated mannequin, IOPS and throughput have been a operate of how a lot capability you provisioned, which is flawed for AI shapes that want small capability however very excessive IOPS and bandwidth. Provisioned v2 splits capability, IOPS, and throughput into three impartial knobs you’ll be able to dial with out downtime or remount. That alone adjustments the economics for lots of inferencing patterns.

The opposite massive inferencing characteristic is NFS v4.1 encryption in transit, delivered with the az-nfs utility and stunnel. You get AES-GCM TLS safety on the wire, with no Kerberos and no Energetic Listing wanted, and the appliance has no concept it’s occurring. Reena’s reside demo confirmed an AKS pod with an encrypted NFS mount, clear to the workload.

After which the sample that ties it collectively: the shared mannequin cache. Obtain the mannequin as soon as into Azure Information, mount it throughout each duplicate by way of ReadWriteMany. No per-pod chilly begin, no re-download from Hugging Face, no egress invoice. The demo used GPT-OSS 120B with VLLM on 32 x H100, and the sample scales right down to small fine-tuned fashions operating on serverless GPU in Azure Container Apps.

The session closed with the Viton case research. Viton is a Paris-based vogue AI startup. Their image-generation platform runs on Azure Container Apps serverless GPU, with Azure Service Bus for job routing and Azure Information NFS because the shared mannequin retailer. Employees pull jobs, mount the shared mannequin cache, generate the picture, and scale to zero when the queue drains. The economics solely work as a result of they don’t seem to be paying to re-download the mannequin on each chilly begin, and since they solely pay for GPU when there may be work to do.

The identical sample reveals up throughout buyer situations:

• Coaching a basis mannequin on AKS with AMLFS because the scratch tier and Blob because the sturdy archive.
• High quality-tuning smaller fashions the place AMLFS checkpoints take up the write bursts and the ultimate artifact lands again in Blob.
• Inferencing with VLLM on AKS the place Azure Information holds the mannequin weights as soon as and each duplicate reads from the identical RWM mount.
• Serverless inferencing on Azure Container Apps with the identical shared mannequin cache sample, however with scale-to-zero economics.

Sincere tradeoff: Lustre is just not the appropriate filesystem for a 10-pod internet app, and Azure Information is just not the appropriate filesystem for a 32-GPU coaching run. The entire level of the tiering is that you simply choose the appropriate one per stage. Don’t attempt to make one filesystem do all 4 jobs.

If you wish to put this into follow this week:

1. Go to the Azure AI Infrastructure repository on GitHub. Wolfgang’s demo cluster got here straight out of it, and you may spin up an AI-ready AKS cluster with GPU and InfiniBand operators in your individual dev/check subscription.
2. Set up the Azure Managed Lustre CSI driver on AKS if you’re operating coaching or fine-tuning. Begin with a smaller MLFS SKU and measurement up.
3. Activate Azure Information Provisioned v2 on a brand new share, then dial capability, IOPS, and throughput independently to match your inferencing form.
4. Allow NFS v4.1 encryption in transit with az-nfs and stunnel earlier than you set any delicate workload on the wire.
5. Strive the shared mannequin cache sample. Decide one VLLM deployment, level it at an Azure Information RWM mount, and measure chilly begin time earlier than and after.

Catch the total Microsoft Azure Infra Summit 2026 session playlist right here

Cheers!

Pierre Roman