Equinix Inference Exchange: Nvidia and Together AI Bet the Enterprise Edge on Open Models

At its inaugural Horizon customer event on September 2, 2026, Equinix unveiled the Inference Exchange — a distributed AI inference platform built with NVIDIA and Together AI that aims to pull enterprise AI workloads out of centralized clouds and into the colocation giant’s global footprint of more than 280 data centers across 77 metros. Announced alongside a new interconnection product called Fabric One, the offering is a coordinated bet on a single thesis: as AI moves from experimentation to production, where inference runs becomes a strategic decision that determines performance, cost, and governance — and Equinix intends to own the real estate where the answer lands.

The market’s first reaction was a 2% lift in Equinix shares. The deeper reaction will take longer to read, because the Inference Exchange is not a product that ships today — general availability is pegged for Q1 2027. What Equinix is really announcing is a positioning move in the AI infrastructure stack, three layers deep, aimed squarely at the gap between hyperscaler clouds and enterprises’ own data.

Three layers, one stack

The architecture is a clean division of labor. Equinix provides the physical foundation: power, advanced cooling, and day-two operations, all connected through Equinix Fabric — the company’s software-defined interconnection platform — to the clouds, networks, and AI providers that inference workloads depend on. NVIDIA anchors the compute layer with its Enterprise Reference Architectures, the validated blueprints for AI factory builds that the company has been refining with Equinix for years. Together AI runs the platform on top, supporting both multitenant deployments for shared efficiency and dedicated single-tenant environments for workloads that need guaranteed capacity.

The choice of Together AI as the software layer is the telling part. Together’s inference platform supports more than 200 open-source models, and its inclusion signals who this product is for: enterprises that want to escape the gravitational pull of proprietary model APIs without giving up managed infrastructure. CEO Vipul Ved Prakash framed it as proof that “model choice and performance are not trade-offs” — a direct answer to the lock-in argument that has kept many enterprises on single-vendor AI stacks.

NVIDIA’s Raj Mirpuri, VP of global AI clouds and infrastructure ecosystem, cast the deal in asset-class terms: “As accelerated compute becomes a strategic asset class,” he said, the combination gives enterprises “a powerful, distributed foundation to bring intelligence closer to their data, applications and customers.” The phrase is doing real work — it positions inference infrastructure the way enterprises think about real estate and interconnection, which is precisely Equinix’s home turf.

Why “where” beats “how”

The press release leans on a claim that has quietly become consensus among infrastructure strategists: as enterprise AI scales across models, providers, and geographies, location determines everything downstream. Run inference in a distant centralized region and you accept higher latency, egress costs, and a governance surface that spans jurisdictions. Run it closer to your data and users, and performance improves, costs become predictable, and data-residency requirements become tractable.

Equinix’s unfair advantage is ecosystem density. The company operates 230 cloud on-ramps and interconnects more than 10,500 businesses on its neutral exchange. Eight of the top 10 AI model providers and nine of the top 10 AI clouds are already deployed inside Equinix facilities. That means an enterprise placing inference workloads in an Equinix metro isn’t building new connections from scratch — it’s plugging into a fabric where its data warehouses, its cloud providers, its networks, and now its inference platform already meet. The company claims this cuts time-to-first-token and reduces deployment complexity, which is the operational translation of that topology.

Three deployment scenarios

The Inference Exchange is explicitly designed around three enterprise use cases, and each maps to a live pain point in 2026’s AI market:

Metro edge inference. For latency-sensitive applications — customer-facing agents, real-time document processing, copilots that employees actually use — running inference in the same metro as users and data sources shaves round-trips that a centralized cloud cannot. This is the classic Equinix value proposition, extended from networks to tokens.

Open-model migration. Enterprises moving workloads from closed, proprietary models to open-source alternatives get “a direct, low-friction path” to run that migration in production. With frontier API prices re-anchoring downward — and open-weight models closing capability gaps — cost-driven migration is one of the fastest-growing workloads in the inference market, and Together AI’s 200-model catalog is the on-ramp.

Sovereign AI. Regulated industries and specific geographies increasingly require that both data and inference processing stay within jurisdictional boundaries. The Equinix footprint spans 77 metros across dozens of countries, which lets a multinational bank or health system place inference in locations that satisfy residency rules without abandoning a managed platform.

The competitive read

The announcement lands in a market where the hyperscalers have spent two years convincing enterprises that AI belongs inside their clouds. Equinix’s counter is neutrality: “architectures that are neutral by design, open by default,” in CEO Adaire Fox-Martin’s words. The company has spent nearly three decades as the Switzerland of interconnection — the place competitors colocate because everyone else does — and it is now applying that playbook to AI inference.

The gaps are real, though. GA is five months out, and the Q1 2027 date means the platform will arrive after competitors have had another two quarters of entrenchment. Hyperscalers can bundle inference with training pipelines, data lakes, and governance tooling in ways a colocation play cannot match without deep partner integration. And the NVIDIA relationship, while deep, is not exclusive — the same reference architectures are available inside every major cloud.

What Equinix is really selling is an exit from the bundling. For enterprises that have watched their AI costs track a single vendor’s price list, a neutral exchange where open models run on validated hardware, connected to everything they already use, is the first credible alternative architecture at global scale. Analyst Nick Patience of The Futurum Group put it in exactly these terms: organizations are increasingly focused on “where inference runs and how quickly it can be deployed into production.”

What to watch

The markers between now and Q1 2027: whether Together AI’s platform lands on Fabric with genuine multitenant economics; whether enterprises actually use the open-model migration path at scale, or merely price it as leverage against incumbent vendors; and whether Fabric One — the companion product announced at Horizon — simplifies cross-metro AI networking enough to make distributed inference operationally sane for teams that struggled to manage even one region.

The deeper question is structural. If inference truly becomes a distributed, exchange-traded commodity that runs wherever latency, cost, and regulation are optimal, then the companies that own interconnection win the tax on every token. That is the future Equinix is underwriting with this launch — and the reason a colocation company just became one of the more interesting names in AI infrastructure.