Equinix turns the network into the control plane for enterprise AI with Fabric One and Inference Exchange

Equinix has used its first Horizon event to pose a question that, according to the company, is redefining enterprise AI architecture: where should inference run? The article, an opinion column by Zeus Kerravala (ZK Research) published on SiliconANGLE, argues that after years of conversation centered on GPU scarcity and cost, training clusters and 'AI factories', the immediate operational challenge has become something else: distribution. Data lives spread across multiple clouds and corporate systems, users and sensors are geographically dispersed, models may be proprietary, open or served as a service, and AI agents will increasingly call other agents, tools and APIs across organizational and infrastructure boundaries. In that scenario, compute is just one piece of a system that needs to be connected, governed and continuously optimized.

Equinix's two Horizon announcements rest on that premise. The first, Equinix Fabric One, is an intent-driven managed 'any-to-any' connectivity service: instead of the customer manually designing and provisioning a collection of different network services, they describe the business or application outcome they need (for example, a secure, resilient, low-latency connection between a data source in one city, an inference provider in another, a cloud application and a factory or branch office) and the platform takes care of composing routing, cloud connectivity, encryption, redundancy and failover as a single managed service. Chris Audie, Equinix's chief product officer, summed up the paradigm shift by saying that "for decades, enterprise networks have been built connection by connection, for every partner and provider they depend on", a model that no longer scales in a world of distributed AI that demands dynamic, real-time connectivity. Fabric One will build on open connectivity specifications developed together with Amazon Web Services and Google Cloud, its reference integration partners; it will enter beta this year and general availability is expected in 2027, initially in North America. Customers will be able to interact with the service via portal, APIs, automation workflows, agents or natural language, while the platform continuously validates the network and manages policies, encryption, resilience and failover.

The second announcement, Equinix Inference Exchange, tackles the complementary problem: how to bring production inference capacity closer to a company's data, applications and users without that company having to build the entire stack itself. It combines Nvidia's enterprise and AI infrastructure reference architectures, the Together AI platform (which provides access to more than 200 open-source models) and Equinix's global network of data centers and interconnection. The service will support multi-tenant deployments for shared efficiency and single-customer environments for workloads that require isolated capacity, and it is expected to be available in the first quarter of 2027. The underlying reasoning is that training can be concentrated in massive, centralized clusters because it is a batch, long-running process, whereas inference lives in the active path of an application, a customer interaction or an agentic workflow, which makes it sensitive to latency, to the cost of moving data, to sovereignty requirements and to the cumulative cost of every token generated.

Equinix CEO Adaire Fox-Martin articulated that idea by pointing to four pressures converging on enterprise IT: fragmented compute, the rise of machine-to-machine traffic, the challenge of achieving predictable returns from AI, and increasingly strict data sovereignty requirements. Her key line was: "where compute runs is becoming as important as the compute itself". Nvidia CEO Jensen Huang joined by video from the G20 summit to take part in a conversation with Fox-Martin and reinforced the same architectural argument: "the world of AI is going to be decentralized, fundamentally decentralized", describing applications that will orchestrate agents with access to data and services spread across multiple clouds and locations. Huang also stressed the importance of proximity: compute must be close to the 'action' (sensors, customers, factories, cell towers, airports and other high-data-volume places where response time matters), while prior knowledge and long-term memory can reside further away. According to the article, Equinix's distributed footprint and interconnection fabric are designed to support both realities at once.

Fox-Martin described Inference Exchange as a service designed to be "production-ready, fully connected and built to scale from day one", with Nvidia compute "preinstalled and preconnected" within the Equinix ecosystem, and with Together AI contributing model variety to avoid lock-in at the model layer. Huang added that companies will use both closed models, important for frontier capabilities, and open models, increasingly relevant for fine-tuning, incorporating domain-specific knowledge, regulatory or sovereignty compliance, and control over strategically important AI capabilities.

Kerravala's analysis emphasizes that Equinix's differentiator is not eliminating the complexity of AI —something, he writes, no provider can achieve— but packaging some of the hardest infrastructure decisions into a more neutral, consumable operating model. The company backs that ambition with figures about its size: it operates more than 280 data centers in 77 metropolitan areas, offers 230 cloud on-ramps, interconnects more than 10,500 companies and claims that eight of the top ten AI model providers and nine of the top ten AI clouds have active deployments on its platform. The author himself warns that those figures alone do not guarantee that Fabric One or Inference Exchange will become indispensable services for enterprises: the proof will come with execution, that is, with the depth of the integrations with clouds and AI providers, the maturity of automated orchestration, pricing transparency, service-level commitments, geographic availability and the real ability to simplify intent-driven connectivity without hiding critical controls.

Taken together, the article reads Horizon as a meaningful shift in Equinix's narrative: connectivity is no longer presented as a support service sitting beneath cloud and AI, and is instead framed as the enabler of a distributed, multicloud and multimodel AI strategy. The author's conclusion is that the winning AI architecture will not be defined solely by who owns the most compute, but by who is able to place the right compute, model and data path in the right location, and to adjust that decision as performance, cost, sovereignty and business needs change.