Two announcements at Equinix Horizon point in a less obvious direction than the headlines suggest: turning the data centre into an operational node for distributed AI, with hardware on one side and connectivity on the other. The first service, Inference Exchange, was developed with NVIDIA and Together AI; the second, Fabric One, uses open specifications created by AWS and Google Cloud. This is not a simple catalogue update. It is an attempt to make a workload that many enterprises still treat as a single-cloud problem into something multi-provider and multi-region.

Distributed inference between open models and reserved capacity

Inference Exchange lets organisations place AI inference closer to data, applications and users through Equinix data centres connected by Equinix Fabric. The offer combines NVIDIA Enterprise Reference Architectures with the Together AI inference platform, which Equinix says supports more than 200 open source models. Together AI provides both multitenant deployments and dedicated single-tenant environments with reserved GPU capacity. Equinix supplies power, cooling and operational support; NVIDIA contributes the reference architectures; Together AI runs the inference platform.

This move does not start from zero. In January 2024, Equinix introduced a managed private cloud service that allowed enterprises to place privately owned NVIDIA DGX infrastructure inside its data centres. Inference Exchange extends that logic to distributed inference, with three explicit use cases: metro edge, open-model migration and sovereign AI deployments. The strategic point is that geographic placement becomes a lever for meeting data residency and sovereignty requirements. Enterprises do not have to own every server; they can choose a dedicated environment in a specific location, with contracts and reserved resources.

Connectivity becomes an API, not a manual project

The second component, Fabric One, is the more radical one for people who watch network architectures. The service automates parts of network provisioning and management that would normally require individual configuration: organisations declare their requirements through a portal, APIs, automation workflows, agent-based requests or natural-language prompts. Equinix manages routing, cloud connectivity, encryption, resiliency and failover. The foundation is an OpenAPI 3.0 Interconnect specification developed by AWS and Google Cloud, with a symmetric API for coordinated Layer 3 connectivity. AWS describes Fabric One as building on previous work; Google Cloud will integrate it with Cross-Cloud Network.

This is where the change has second- and third-order implications. An AI agent can request and provision connectivity programmatically: control does not disappear, but it shifts from individual configurations to the policies that govern them. For teams managing self-hosted or hybrid infrastructure, the issue is not only technical: security and compliance become properties of orchestration rather than after-the-fact reviews. The IDC survey cited in the announcement notes that many organisations still balance manual processes with network automation; Fabric One tries to move beyond that coexistence by making the network declarative.

Equinix brings the scale of its footprint to the table: more than 280 data centres across 77 metropolitan areas, over 230 cloud on-ramps and about 3,000 providers. The message to model providers is direct: eight of the top ten AI model providers and nine of the top ten neoclouds already have active deployments inside its data centre network. For cloud providers the relationship is ambivalent: AWS and Google Cloud are Fabric One partners, but the symmetric specification reduces room for proprietary networking lock-in. Vendors selling dedicated GPU capacity and open models gain a new channel; those whose advantage rested on multi-cloud complexity lose an excuse.

For anyone evaluating on-premise or hybrid deployment, the critical variable is the boundary between physical control and contractual control: a dedicated environment in an Equinix data centre is not the same as a self-hosted cluster behind the enterprise perimeter. AI-RADAR offers analytical frameworks to compare the trade-offs without shortcuts. Fabric One is expected in beta in 2026 and general availability in 2027: a timeline that suggests Equinix is building a market, not responding to already mature demand.