Caleta Private AI Appliance / VM sizes and regions
Supported VM sizes and regions
All 29 Azure VM sizes the appliance supports, and where in the world you can actually get one. Speeds per model, each with the box and date it was measured on, are in the model catalogue.
The appliance itself is not limited to this list: it checks what fits in GPU or system memory at serve time. These are the sizes we curate and stand behind.
Where can you actually run it?
Microsoft's datacentre map shows how big Azure is. This one shows how small the footprint gets for the box you want: pick a size, see every region it is priced in, where the hardware really exists, and the three regions we would pick, ranked the same way the appliance's own spot placement menu ranks them. Updated daily.
Loading spot and hardware data for NC24ads_A100_v4...
Reading the flags
- Limited availability means Microsoft capacity, not our software. The 8x MI300X size is fully supported by the appliance, but for most of August Azure could not allocate one on spot anywhere, in any region it is priced in. If you can get the box, the appliance runs on it.
- Extra setup means a size works, but does not deploy like the others: Secure Boot and vTPM have to be off at deploy time, and the GPU is enabled once from the menu afterwards. It applies to the NVIDIA GRID cards, the RTX PRO 6000 Blackwell and A10 families, which the appliance drives through the same code path. Detail is in the GRID section below.
- CPU (beta) sizes serve at single-user speeds. It genuinely works, and it is striking to watch a frontier model answer with no GPU at all, but GPU sizes are the choice when several people need throughput at once.
GPU sizes
Cards passed straight through to the VM: deploy with Trusted Launch, boot, serve. Each size links to its full hardware record, with exact vCPU, RAM, disk and network ceilings, on our free VM Capabilities explorer.
| Size | GPU | Notes |
|---|---|---|
| NC4as_T4_v3deploy now | 1x Tesla T4 16 GB | 4 vCPU. The cheap box. |
| NC8as_T4_v3 | 1x Tesla T4 16 GB | 8 vCPU, same card - more cores does not mean faster GPU serving. |
| NC16as_T4_v3 | 1x Tesla T4 16 GB | 16 vCPU, same card. |
| NC64as_T4_v3 | 4x Tesla T4 16 GB | 64 vCPU, 4 cards, 64 GB total. |
| NC24ads_A100_v4deploy now | 1x A100 80 GB | The reference box. |
| NC48ads_A100_v4 | 2x A100 80 GB | 160 GB total. |
| NC96ads_A100_v4 | 4x A100 80 GB | 320 GB total. |
| NC40ads_H100_v5 | 1x H100 NVL 95 GB | |
| NC80adis_H100_v5deploy now | 2x H100 NVL 95 GB | 188 GB total. |
| ND96is_MI300X_v5 | 8x MI300X 192 GBlimited availability | ~1.5 TB total - the frontier-giant box. Azure capacity is scarce; software support is not the constraint. |
NVIDIA GRID sizes
These are virtual GPU cards rather than passthrough, and they take a different route onto the box. The appliance treats them as one family and supports both.
What extra setup means
Deploy these sizes with Secure Boot and vTPM switched off (the Standard security type, not Trusted Launch), then run the GPU enable step once from the llm menu. That step installs NVIDIA's GRID driver and licenses it. It is a one-off: the box serves normally afterwards.
Get the licensing step wrong and the card runs at full speed for roughly twenty minutes, then NVIDIA clamps it hard. The menu handles this for you, which is why it exists.
We publish no speeds for the A10. It runs, and a half card holds one small model at a time, but no A10 figure has been cleared for publication. The NVIDIA GRID page covers both families in full, including the partial-card profiles and the licensing grace period.
| Size | GPU | Notes |
|---|---|---|
| NC24lds_xl_RTXPRO6000BSE_v6 | Quarter RTX PRO 6000 Blackwell SEextra setup | Quarter card, 24 GB visible. Slower than a full A10 at the same visible VRAM. |
| NC72ds_xl_RTXPRO6000BSE_v6 | Half RTX PRO 6000 Blackwell SEextra setup | Half card, 48 GB visible. |
| NC144ds_xl_RTXPRO6000BSE_v6 | 1x RTX PRO 6000 Blackwell SE 98 GBextra setup | |
| NC288ds_xl_RTXPRO6000BSE_v6 | 2x RTX PRO 6000 Blackwell SE 98 GBextra setup | 196 GB total. |
| NV18ads_A10_v5 | Half A10extra setup | Half card, 12 GB visible - fits only the smallest models, one at a time. |
| NV36ads_A10_v5 | 1x A10 24 GBextra setup | |
| NV36adms_A10_v5 | 1x A10 24 GBextra setup | Full card with 880 GB RAM. |
| NV72ads_A10_v5 | 2x A10 24 GBextra setup | 48 GB total. |
CPU sizes (beta)
x86 boxes with no GPU. Mixture-of-experts models with small active parameter counts are the CPU sweet spot; dense models run at a few tokens per second.
Azure Cobalt sizes (Arm plan)
The ARM64 build is a separate plan in the same Marketplace offer, on Azure Cobalt 100 CPUs: a full physical core per vCPU, D-series at 4 GiB of memory per vCPU and E-series at 8 GiB. It has its own curated model list, measured on Cobalt, in the ARM section of the guide.
| Size | Notes |
|---|---|
| D16ps_v6 | Entry general-purpose size, 64 GiB RAM. |
| D32ps_v6 | The ARM plan's primary recommendation, 128 GiB RAM. |
| D64ps_v6 | 256 GiB RAM, for the larger models. |
| E8ps_v6 | 64 GiB RAM - fits the smallest models. |
| E16ps_v6 | 128 GiB RAM - minimum for the mid-tier models. |
| E32ps_v6 | 256 GiB RAM - fits most of the ARM list; the box the ARM figures were measured on. |
| E64ps_v6 | 512 GiB RAM - the largest ARM models, including GLM-5.2. |
Related pages
- Model catalogue - every model with its deployment pairings and all-in prices
- VM Capabilities explorer - the full hardware record behind every size on this page
- Spot pricing - live prices for these sizes in every region Azure sells them
- Known issues and sharp edges