Oxide CTO Bryan Cantrill argues agentic AI's CPU-heavy workloads and data governance needs favor owning compute over renting
Oxide Computer CTO Bryan Cantrill argues at SC26 that agentic AI workloads, which are CPU-heavy and involve sensitive enterprise data, create conditions where owning infrastructure becomes more attractive than renting. He identifies economics as the primary driver, with data governance as a close second. The manuscript examines this vendor-originated thesis and its structural limitations.
Bryan Cantrill, co-founder and CTO of Oxide Computer, made a case at SC26 that agentic AI is pulling enterprise compute back on-premises. Oxide is an on-premises cloud infrastructure company that raised $445 million in October 2026 to expand rack production. Cantrill's argument rests on economics as the primary driver, with data governance and security as close seconds.
"Cloud computing is great, but the model shouldn't be exclusively to rent it. You get to a certain scale, as it turns out, you want to own your own computer," Cantrill said. He identified three reinforcing factors: agentic workloads are driving an explosion in compute demand, and much of it is CPU-heavy work that enterprises would rather own than rent; there is a real enterprise drive toward open-weight models, with organizations wanting to run those models on hardware they control so their data does not leave; and multi-tenancy isolation for potentially adversarial tenants sharing the same physical resources requires what Cantrill describes as public cloud-grade controls built into the rack from the start.
What distinguishes this from a generic cost argument is Cantrill's framing of the model plane — the layer that decides where AI models run — as simultaneously an infrastructure and a data governance decision. If an agent's value depends on accessing proprietary internal data across multi-step tasks, then the location of inference is not just about price per FLOP but about whether that data crosses an organizational boundary.