Nutanix and GPUs: building an AI cluster

Which nodes take cards, and which do not

A hyperconverged node is a server that combines storage, compute and networking in one system under AOS. Drive bays and fans often leave no room for a full-size card, so some NX models take up to four accelerators and others take none.

Platforms per Nutanix’s NX FAQ (February 2026); card counts per the node spec sheets as reprinted by resellers (G10 nodes have shipped since October 2025). Nutanix’s spec-sheet list (September 2026) also shows a 4U NX-9183-G10 for AI workloads; confirm the spec sheet of the exact model before ordering.

The model code does not tell you whether a node takes GPUs; the table above does. Two checks are reliable: the GPU line in the model’s spec sheet, and the card list validated for that chassis. The NX-9151-G9 accepts the H100 NVL; the NX-3155-G9 with the same 2U1N form factor stops at two L40S; the H100 80 GB PCIe is qualified but not sold by Nutanix. The limit is set by power and airflow rather than by the size of the case. We supply the L40S and L4 cards that Nutanix lists for these nodes. Describe your node and workload and our team will help.

Power and airflow limits for GPUs in NX nodes

The power spread across the line is large: an edge node draws a few hundred watts, while an NX-9151-G9 with four 350 W cards needs a power budget of several kVA and a matching cooling budget, a separate conversation with your data centre about the phase and row cooling.

Cooling is the second limit, because server cards are passive. The L40S and RTX PRO 6000 Server Edition have no fans of their own and depend on front-to-back chassis airflow. Without directed airflow the card overheats: on NVIDIA’s forum, an L40S in a workstation case was reported at 100 °C while idle. A passive card needs a server chassis that pushes air through it.

Passthrough, time-sliced vGPU and MIG-backed vGPU

MIG has one limit that matters for multi-card node plans. According to the MIG documentation, with MIG enabled there is no peer-to-peer communication between instances on different GPUs, over PCIe or NVLink, and NCCL is not supported. NVIDIA’s position on its forum is that distributed training over MIG can in theory run through GPUDirect RDMA over the network, but more slowly than on a larger slice or with MIG disabled. Partitioning a card and fast card-to-card exchange therefore exclude each other.

MIG instances per card

The last row matters when choosing a node: the L40S is listed for both the NX-9151-G9 and the NX-3155-G9, and it does not support MIG. Teams that choose the L40S for its compute sometimes find out later that it cannot be partitioned in hardware; on these nodes it is shared by time slicing only. The H100 NVL in the table is the one Nutanix lists for the NX-9151-G9; a reseller notice set its last-time-buy date at 1 November 2025, and we do not supply it.

vGPU and MIG support on Nutanix AHV

NVIDIA’s Nutanix AHV support page (updated 29 September 2026) supports AHV “as a generic Linux with KVM hypervisor”, and for such hypervisors NVIDIA leaves the supported GPUs and features to the hypervisor vendor. Ordinary time-sliced vGPU works. NVIDIA’s validated-platforms page for Linux with KVM (updated 29 September 2026), which names Nutanix AHV among the hypervisor vendors, lists 1:1 and time-sliced MIG-backed vGPU for the RTX PRO 6000 and RTX PRO 4500 Blackwell Server Edition and leaves the supported releases to the vendor, so ask Nutanix whether your AOS and AHV release supports it before planning MIG slices for VMs. Verify compatibility for your AOS version before ordering: NVIDIA AI Enterprise 8.2 lists Nutanix AOS 6.5 to 6.10, marked “single-node only”, and NVIDIA supports AI Enterprise only on NVIDIA-Certified Systems, a list on which the NX-9151-G9 is the only NX node as of October 2026.

Two further issues appear only after deployment. The first is live migration, which for a vGPU VM requires the identical physical card on the receiving host and a driver from the same release branch that is the same or newer, and never works across different MIG profiles. Nutanix states that AHV live-migrates vGPU VMs to a host with a matching free vGPU profile; NVIDIA’s support page treats AHV as generic Linux with KVM, so confirm with Nutanix that your AOS and vGPU release support it before you plan maintenance around it. The second is behaviour without a valid licence: NVIDIA’s current client licensing guide (updated 29 September 2026) states that a graphics vGPU’s performance is degraded if the VM has no licence after 20 minutes and further degraded after one day, without naming figures; the 15 fps and 3 fps caps come from the guide for vGPU release 16, which does not support the RTX PRO Blackwell cards. A compute (C-series) vGPU drops to idle-level compute after a 20-minute grace period, according to NVIDIA AI Enterprise’s licensing page (2 September 2026). C-series compute profiles come only with NVIDIA AI Enterprise, licensed per GPU, while graphics vWS is licensed per concurrent user, so these are two separate budget lines.

We supply the NVIDIA AI Enterprise and vGPU licences an AHV cluster needs, on the same quote as the GPUs. Tell us your nodes and the vGPU profiles you plan, and we reply within one business day.

How many VMs per card

Compute profiles for the RTX PRO 6000 Blackwell Server Edition

For virtual desktops NVIDIA’s sizing guide gives two to four heavy users per RTX PRO 6000 (DC-12Q to DC-24Q profiles) and six to eight light users per RTX PRO 4500 Server Edition (DC-4Q). NVIDIA’s guide adds that the default best-effort scheduler often allows two to three times that user density. For monitoring, VM video-memory use should rarely exceed 90% and should average below 70%.

Source: Red Hat blog, “Sharing is caring: How to make the most of your GPUs part 2 – Multi-instance GPU”, 6 September 2024: an A100 40 GB split into 7 × 1g.5gb; a measurement on Ampere, not on the cards discussed above.

Each slice is slower than the whole card, but a small model does not saturate an A100, so in Red Hat’s test seven slices together delivered about 2.3 times the throughput of the undivided card.

Cluster size and GPU pool: where to start

The working starting configuration is three nodes: RF2 replication runs on three, and the cluster keeps its data available when one node fails, while high availability restarts that node’s VMs on the others. A VM with a GPU restarts only on a node with a free card of the same type, which is one more reason for the next paragraph. Cards are not needed in every node; a common pattern is three nodes for general workloads, one or two of them with accelerators for inference.

Keep the GPU pool homogeneous, because live migration needs the same card on both hosts. A second node with the same card gives headroom and a place for VMs to move during maintenance. A mixed cluster with an L40S in one node and an RTX PRO 6000 in another works too, but every planned maintenance then means stopping the services on the node with the different card.

What we supply

We supply the L40S and L4 cards that Nutanix lists for NX nodes, and the RTX PRO 6000 Server Edition, with manufacturer warranty on one EU contract and invoice. The NVIDIA AI Enterprise and vGPU licences that C-series and graphics vGPU profiles need come on the same contract. Nutanix NX nodes can be quoted on request; which cards a given node takes, and whether they must ship with the node, is for Nutanix to confirm. Deployment and virtualisation engineering comes from our engineering partner Vixen.UNO under the same contract.

FAQ