Nutanix bets on 'agentic AI'

Nutanix came to its.NEXT conference in Chicago this week with a pitch that sounds modest until you hear what it is pushing against. Instead of telling customers that the answer to enterprise AI is another round of GPU buying, it said the harder problem is using scarce hardware well: measure it, share it, govern it, and reserve the expensive parts for work that earns their keep. On April 7, the company recast its cloud platform around that idea, bundling new AI controls, broader hardware options, bare-metal Kubernetes, and more sovereign-cloud support into what it called a platform for the “agentic AI era” (nutanix.com, siliconangle.com). The timing matters. Nutanix had only just introduced its Agentic AI stack on March 16 at NVIDIA GTC, describing a system for running large numbers of AI agents, models, and services inside one managed environment rather than spinning up one giant training job and calling it a day. In Nutanix’s telling, production AI now looks less like a single marathon and more like an airport: thousands of arrivals, departures, and gate changes, all happening at once, with costs piling up every time a model is called. The company said that earlier launch was built to lower and stabilize token costs, and this week’s updates turn that cost story into infrastructure policy (nutanix.com, nutanix.com). That is where the new metering features come in. Nutanix said its Cloud Manager will let service providers monitor AI infrastructure and bill by GPU usage, API calls, or model consumption. That sounds like plumbing, but it changes the conversation inside an enterprise. Once a team can see which agent is burning through inference calls, which model is idling on a reserved GPU, and which customer or business unit is driving the bill, AI stops looking like a mysterious capital expense and starts looking like a schedulable service. Even SiliconANGLE’s coverage, which is usually generous to infrastructure vendors, lingered on the same point: token usage is becoming a first-order operating metric, not a footnote on a cloud invoice (nutanix.com, virtualizationreview.com, siliconangle.com). Nutanix is also trying to widen the set of machines that can carry those workloads. The company said customers will be able to modernize virtual machines and containers with existing server and storage investments, including planned support for NetApp ONTAP with AHV later in 2026 and tighter collaboration with Dell PowerStore. It also pointed to support for Google Cloud C3 bare-metal instances with Hyperdisk in the second half of 2026, a combination meant to separate storage growth from compute growth and give customers another place to land when on-prem hardware is backordered (nutanix.com, nutanix.com, nutanix.com). The most concrete engineering move was NKP Metal, announced the same day. It extends Nutanix Kubernetes Platform so containers can run directly on bare metal instead of always sitting on top of a hypervisor. For AI training, edge inference, and other jobs that want every bit of GPU bandwidth and as little overhead as possible, that matters. Nutanix’s answer is not to abandon virtual machines, but to manage both styles together, which it calls a “dual-native” approach: VMs where isolation and compatibility help, bare metal where the workload is too hot or too latency-sensitive to waste cycles (markets.businessinsider.com, siliconangle.com). Then there is sovereignty, the word that keeps showing up whenever AI leaves the lab and enters government, healthcare, finance, or a factory floor. Nutanix tied its agentic AI push to multitenant controls for “neocloud” providers and to support for sovereign deployments, including AWS European Sovereign Cloud. The point is not only legal residency. It is operational control: who can access the model, where logs live, which region serves inference, and whether a regional provider can offer GPU capacity without handing the whole stack to a hyperscaler. Nutanix wants to be the layer that makes those choices portable across on-prem systems, service providers, and public cloud (nutanix.com, nutanix.com, press.aboutamazon.com). For a software engineering manager, the interesting part is not the phrase “agentic AI.” Every vendor now has one of those. It is the way Nutanix is translating AI ambition into classic platform questions: who gets the scarce resource, how usage is priced, which workloads deserve bare metal, and how one control plane spans old servers, new accelerators, and regulated regions. That is a hardware-software story in the old sense of the phrase. The glamour sits with the model, but the leverage sits with the scheduler, the meter, and the box you already own (nutanix.com, nutanix.com, siliconangle.com).