Extra setup

NVIDIA GRID GPUs, on the appliance

Azure's RTX PRO 6000 Blackwell and A10 sizes are virtual GPU cards rather than passthrough, and the appliance treats them as one family: the same driver, the same enable step, the same deploy posture. They work, and this page says plainly what the extra setup is. If you want the easy path, the recommended sizes need no enable step at all.

Get one running, from coldSix steps, about fifteen minutes

RTX PRO 6000 Blackwell SE

Azure size names in this family follow the pattern Standard_<size>_RTXPRO6000BSE_v6. The card is 98,304 MiB; the smaller sizes are hardware slices of it, and what the guest sees is the slice.

Size	Card	GPU memory	System RAM	Worth knowing
NC24lds_xl	Smallest slice	24,576 MiB	72 GB	The cheapest way into this family. No ds twin at this tier. Slower than a full A10 at similar memory, so consider an A10 or A100 first.
NC36lds_xl	Smallest slice	24,576 MiB	72 GB	The same slice and the same host memory as NC24lds with 12 more vCPU, for pennies an hour more, and it sits in a lower eviction band.
NC36ds_xl	Smallest slice	24,576 MiB	132 GB	Same GPU as NC36lds with double the host memory, and the ds twin NC24lds never had.
NC72lds_xl	Half card	49,152 MiB	132 GB	The value pick at this tier: same GPU as NC72ds, 14% cheaper.
NC72ds_xl	Half card	49,152 MiB	264 GB	Same GPU as NC72lds with double the host memory.
NC144lds_xl	Full card	98,304 MiB	264 GB	The size where this family makes sense, and the cheaper half of the pair.
NC144ds_xl	Full card	98,304 MiB	516 GB	Same GPU as NC144lds with double the host memory.
NC288lds_xl	Two full cards	196,608 MiB	516 GB	Same price as NC288ds and a lower eviction band, so there is little reason to take the ds here.
NC288ds_xl	Two full cards	196,608 MiB	1,032 GB	The biggest Blackwell box today.

A10

Azure size names here follow Standard_NV<n>ads_A10_v5. A full A10 presents 24,512 MiB, so this is a small-model tier: it comfortably holds the likes of Phi-4 and gpt-oss-20b, and a 27B at 4-bit does not fit alongside its context even though the weights alone would.

Size	Card	GPU memory	System RAM	Worth knowing
NV18ads_A10_v5	Half card	~12,288 MiB	220 GB	Booted and served on the shipping image. Tight: plan on one model resident.
NV36ads_A10_v5	Full card	24,512 MiB	440 GB	The sensible A10 rung.
NV36adms_A10_v5	Full card	24,512 MiB	880 GB	The same card as NV36ads with double the host memory.
NV72ads_A10_v5	Two full cards	49,024 MiB	880 GB	Priced close to A100 and H100 sizes, which hold more and need no enable step.

We publish no tokens-per-second figures for either family. Blackwell runs exist but are internal sizing measurements, and no A10 figure has been cleared for publication. Our rule is that a published number is measured, dated and reproducible, so until one is, these say nothing.

What "extra setup" means here

Not "it might not work". It means the deployment posture differs from every passthrough size we support, and there are sharp edges we would rather list than have you discover.

Secure Boot and vTPM must be OFF, at deploy time

The only driver that binds these cards is the NVIDIA GRID guest driver, and it is unsigned. Both settings are deploy-time choices that cannot be changed from inside the VM: a box deployed with Trusted Launch must be deallocated and updated before the enable step can do anything. Every passthrough size we support keeps Secure Boot on; the GRID families are the exception, and it is the main reason they need extra setup.

One enable step, from the menu

The driver cannot be redistributed inside a commercial image, so the appliance fetches it from Microsoft when you run the GPU enable step from the menu. One driver covers both families. It is safe to re-run, reports status when run bare, and switches the appliance to the right engine build afterwards.

Partial-card sizes report partial cards

The quarter and half sizes are vGPU slices, and the appliance sees only the slice: 24,576 MiB on a quarter Blackwell, around 12 GB on a half A10. That is expected, not a fault. Bandwidth is sliced with the memory, so a quarter Blackwell is slower than a full A10 holding the same 24 GB. The appliance tells you when it detects a partial profile, and it reads free memory from the device rather than from a size table.

Licensing has a silent grace period

An unlicensed vGPU runs at full speed for around 20 minutes after boot, then the driver clamps compute several times slower, with no error and no warning. The enable step handles the licence; if a GRID box mysteriously slows down shortly after boot, re-run the enable step and check the licence state before blaming the model.

Re-run the enable step after a kernel update

Automatic updates are off on the appliance, so OS patching is yours to do, and on these families a kernel update needs one extra step. The GRID driver is installed for the kernel that was running at the time, so after a kernel update and a reboot there is no driver for the new kernel until you re-run the GPU enable step, which reinstalls it.

Getting one running, from cold

Six steps, about fifteen minutes, most of it waiting for a download. There is no one-click button for this family on purpose: the deploy needs two settings that a template cannot choose for you, and getting either wrong gives you a GPU box quietly running on its CPU. Follow these in order and it works.

1
Register one subscription feature, once
These cards need the Standard security type, and most subscriptions have never used it. Register it once and you never think about it again. Skip it and the deployment fails with a BadRequest that does not explain itself.
```
az feature register --namespace Microsoft.Compute \
    --name UseStandardSecurityType
az provider register -n Microsoft.Compute
```
Registration takes a few minutes to go live.
2
Check you have GPU spot quota
Spot quota starts near zero on a new subscription and is granted per family, so ask before the day you need it. The walkthrough with portal screenshots covers it. Running on pay-as-you-go instead needs no quota request and costs about five times more.
3
Deploy from the Marketplace, in West US 2
Open the offer on the Azure Marketplace and deploy it. Pick West US 2: these cards exist in only two Azure regions, and the other one has no capacity for the larger sizes today. Choose size NC144lds_xl for a full card: it is the same GPU as NC144ds_xl with half the host memory, which this appliance does not need, so it costs less and sits in a lower eviction band. Set the security type to Standard so Secure Boot and vTPM are off. That last one cannot be changed afterwards from inside the VM.
4
Sign in and let the menu open
Open the VM's serial console in the Azure portal, which needs no SSH client, or SSH in if you prefer. The llm menu starts by itself.
5
Run the GPU enable step
Choose GPU status / enable (experimental). It fetches NVIDIA's driver, licenses it and switches the appliance onto the GPU build. Safe to re-run. Do not skip the licensing: an unlicensed card runs at full speed for about twenty minutes and is then clamped several times slower with no warning.
6
Download a model, then load it
Two separate menu actions: download a model puts the weights on disk, load a model serves them. A full card holds 98,304 MiB, so anything up to about 90 GB fits. The menu declines a model that does not.

What it costs. A full card on NC144lds_xl runs about $1.41 an hour all-in on spot in West US 2, which is the VM at roughly $1.16 plus the appliance software fee. The ds twin is about $1.60 for the same GPU. Both land on the Azure bill you already have. Destroy the resource group when you are done and it stops.

What we do not claim. We publish no tokens-per-second figures for this family. Our runs are internal sizing measurements, and a published number here has to be measured, dated and reproducible.

The rest of the appliance works exactly as the user guide describes, and the full size list is on VM sizes and regions. Questions, or numbers from your own runs? We would genuinely like to hear them.

Ready to deploy one?

The appliance is one image on the Azure Marketplace, and it runs on every size listed above as well as the passthrough GPU, CPU and Azure Cobalt sizes. Work through the six steps first: on this family the deploy settings matter.

Read the six steps first