I follow Nutanix on infrastructure advisory calls. I have not worked with their analyst relations team on this deal or on earlier ones. The same constraint keeps coming up. Teams place work in a private hall, a public cloud account, and a neocloud because they cannot get enough accelerators in one place. A job often reserves a full graphics processor for the whole window. The model may only use the card for part of that window. The unused hours are still paid.

Nutanix said on September 22 it had acquired Ryax Technologies. Financial terms were not disclosed. The company said the deal is not material to results. The Ryax team will join Nutanix in France. Chief executive Rajiv Ramaswami said agentic artificial intelligence needs simpler operations and more work from the same hardware. Chief AI officer Debo Dutta told CRN that partners will sell the scheduling features once the integration is complete.

Ryax was founded in Lyon in 2017. Andry Razafinjatovo is chief executive. Yiannis Georgiou, the chief technology officer, is now listed as a Nutanix principal engineer on the company blog he wrote with Dutta. The product runs workflows on Kubernetes and can send heavy steps to a Slurm high-performance cluster. IntelliScale records CPU, memory, and GPU use on each step and sets the next request from that history, so teams do not have to over-reserve capacity in the job file.

How one job is supposed to use the card

A job usually asks for more memory than it needs because an out-of-memory failure stops the run. The scheduler reads earlier runs, assigns a smaller share, and retries if memory is still too low. More than one container can share a physical card through NVIDIA Multi-Instance GPU partitions or software packing. When the compute step ends, the card is released for the next job. Placement can choose a lower-cost pool, another cluster, or an existing high-performance machine before a new cloud rental starts.

Nutanix published two Ryax tests. On a 30-run deep-learning burst, sizing each run cut node-hours by 62 percent and finished 5.7 percent faster. On a document pipeline, holding the GPU only during compute shortened hold time from hours to minutes, and four jobs shared one H100 with a 52 percent lower cost per run. Those results are vendor tests. Measure the same jobs on your mix of inference, batch work, and notebooks before you change a budget line.

Nutanix Kubernetes Platform already includes the NVIDIA GPU Operator, Dynamic Resource Allocation from Kubernetes, time-slicing, and all-or-nothing job scheduling (Kueue). Nutanix listed those pieces on August 26 when it described CNCF Kubernetes AI Conformance. Ryax is planned as an extra layer that sizes jobs from history and can place work on Slurm and more than one cloud.

What other platform vendors added this quarter

Red Hat shipped shared GPU control inside Red Hat AI 3.5 on September 9. The release adds fair-share scheduling across tenants and priority so a live service keeps capacity while a batch job waits. It also shows allocated and borrowed GPU capacity in one view. In June, a Red Hat Developer article described OpenShift using the Red Hat build of Kueue and NVIDIA Multi-Instance GPU slices so a developer can reserve a partition. OpenShift Lightspeed is being connected to live infrastructure data through Model Context Protocol.

Broadcom added GPU pooling to VMware Cloud Foundation at VMware Explore on August 31 under the names VMware Private AI Cloud and VMware AI Factory. Cards can be shared across teams. A model gallery runs inference on virtual machines, containers, and shared GPUs. Token throughput and utilization appear on one dashboard. ClearML is the named partner for GPU access and agents. Certified nodes come from Cisco, Dell, Lenovo, and Supermicro. Broadcom is also working with AMD on Instinct cards and the ROCm software stack.

Microsoft put Azure Kubernetes Service flex nodes into public preview on September 22. An existing cluster can add worker nodes from another Azure region, an edge site, or an on-premises hall when a region has no GPU sizes left. Foundry Local on Azure Local already schedules inference across the GPUs in an Azure Local cluster and added vLLM as a serving engine. That path finds spare capacity in another place. The Ryax path is meant to raise use of capacity you already run.

Dell, HPE, Cisco, and Lenovo sell the servers and certified AI racks. Cisco published a mixed NVIDIA and AMD inference test across its fabric. Lenovo lists Nutanix Compute Cluster next to Red Hat OpenShift Virtualization and Azure Local on ThinkSystem. The hardware quote and the software entitlement stay separate decisions.

Other schedulers buyers already run

NVIDIA Run:ai is the commercial GPU scheduler most teams compare first. NVIDIA acquired Run:ai and later released the scheduling engine as KAI Scheduler. Run:ai adds memory isolation, Multi-Instance GPU profiles, multi-cluster control, and access rules on top of that engine.

Kubernetes already covers part of the same work. Dynamic Resource Allocation is generally available. HAMi divides a GPU by memory and compute share. Volcano and Kueue schedule jobs that must start together. Time-slicing lets more than one pod use a card without a hardware partition.

Cast AI and similar tools resize and pack jobs on public cloud accounts. Karpenter adds nodes when pods wait for capacity. Those products are built for rented clusters. Slurm remains the scheduler for much training and science work. Ryax can hand jobs to Slurm.

Ask Nutanix which Ryax functions ship in which Nutanix Kubernetes Platform and Nutanix Enterprise AI release, and whether the first drop can place a job on Slurm and on a public cloud from the same policy. Compare that list with Red Hat AI 3.5, VMware Private AI Cloud, AKS flex nodes, Run:ai, and the Kubernetes tools already on your clusters.

About Shashi Bellamkonda

Loading bio…

LinkedIn Profile

Get new posts by email:

Subscribe

Powered by