# Bare Metal Compute Bare metal means a tenant receives a physical machine, with the operating system installed directly on the hardware and no hypervisor between. For accelerator work this is the common arrangement rather than an exotic one, and the reasons are specific. > [!note] The one-sentence version > The whole physical machine, no hypervisor, nothing between the job and the silicon. ## Why accelerator work prefers it **Performance fidelity.** Virtualization overhead on the data path, however small, matters when thousands of devices must stay synchronized; the fabric, the direct memory access paths, and the accelerator interconnect are all simplest and fastest when the operating system owns the hardware directly; see [[wiki/RDMA|RDMA]] and [[wiki/NVLink|NVLink]]. **Topology visibility.** Distributed training must know the machine's real internal structure, which devices share a fabric domain, which memory is local to which processor, to place work correctly; see [[wiki/Scale-Up and Scale-Out|Scale-Up and Scale-Out]]. Abstractions that hide topology destroy the information the software needs. **No noisy neighbor.** Memory bandwidth and interconnect are the contended resources, and a co-tenant consuming them produces exactly the variable slowness that paces a collective; see [[wiki/Collective Communication|Collective Communication]]. **Driver and firmware control.** Accelerator work depends on specific driver, firmware, and library versions, which a tenant frequently needs to choose; see [[wiki/OCI Container|OCI Container]]. ## How tenants are isolated without a hypervisor Isolation moves into the hardware and the network. The management controller, with its separate network, provisions and wipes the machine between tenants; see [[wiki/Baseboard Management Controller|Baseboard Management Controller]]. Network isolation is enforced on the adapter rather than by host software, which keeps the provider's enforcement outside the tenant's reach; see [[wiki/DPU and SmartNIC|DPU and SmartNIC]]. Storage is presented over the network with per-tenant access control; see [[wiki/NVMe over Fabrics|NVMe over Fabrics]]. Firmware is verified and reset between tenants, and attestation reports what is actually running; see [[wiki/Supply Chain Security|Supply Chain Security]]. Where stronger guarantees are required, hardware memory encryption provides them without a hypervisor; see [[wiki/Confidential Computing|Confidential Computing]]. The honest trade is that bare metal gives the tenant more control and places more obligation on the provider's hardware-level hygiene, since the usual software boundary is absent. Done properly it is both faster and cleaner than virtualization, because the isolation lives in silicon rather than in a software layer the tenant shares. ## Relationships - **provisioned by:** [[wiki/Baseboard Management Controller|Baseboard Management Controller]] - **isolated by:** [[wiki/DPU and SmartNIC|DPU and SmartNIC]] · [[wiki/Confidential Computing|Confidential Computing]] · [[wiki/Supply Chain Security|Supply Chain Security]] - **suits:** [[wiki/Collective Communication|Collective Communication]] · [[wiki/Scale-Up and Scale-Out|Scale-Up and Scale-Out]] - **offered by:** [[wiki/Neocloud|Neocloud]] · [[wiki/Hyperscaler|Hyperscaler]] · [[wiki/Colocation|Colocation]] - **part of:** [[wiki/Cloud Deployment|Cloud Deployment]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** virtualization overhead on data paths and its significance for synchronized distributed work; the dependence of distributed training on visible machine topology; memory bandwidth and interconnect as the contended resources producing neighbor interference; tenant need for specific driver and firmware versions. - **Established:** hardware-level isolation through management-controller provisioning and wiping, adapter-enforced network isolation, network-presented storage with per-tenant access control, firmware verification and attestation, and hardware memory encryption. - **Established (datestamped):** as of October 2026, dedicated bare-metal access to accelerator systems is standard industry practice alongside managed services. ## Sources - Bare-metal cloud provisioning and isolation practice, as current October 2026.