Edgecore bets the next AI data center is a federation of smaller sites, stitched by fiber

A 100-megawatt substation takes years to permit. A cooling plant can swallow a city block. And a single hyperscale campus, the kind that trains the largest AI models, can draw more power than some small countries. Those are the physical limits now shaping how the next wave of AI infrastructure gets built, and they are pushing the industry away from ever-bigger single sites toward distributed clusters stitched together by optical fiber.

Edgecore Networks, a Taiwan-based subsidiary of white-box hardware maker Accton, is making its clearest public bet on that shift. The company argues that distributed AI infrastructure will become more important as data centers run into limits on power, space, and cooling, according to a Digitimes report on the company's positioning. Its chosen tool is all-optical networking, in plain terms a design that replaces the mix of copper and optical cables that links servers and switches inside a data center with end-to-end fiber, so that bandwidth and latency stop being bottlenecks when a single training job spans buildings or cities.

The pace of its recent moves is unusual. At Computex 2026 in Taipei, the company and its partners organized their showcase around the theme of an "all-photonics AI era," bringing together silicon, optics, and software vendors under one roof. Earlier, at the OCP Global Summit 2025 in San Jose and Gitex Global 2025 in Dubai, the company pushed a similar message to the two largest audiences for open hardware buyers: US hyperscalers and Middle Eastern operators building AI capacity from scratch.

The partnerships matter because Edgecore is not a chipmaker or a cloud provider. It sits one layer down, selling open-standard switches and networking gear to operators that want hyperscaler-style economics without buying proprietary stacks. In November 2025, Accton's Edgecore arm joined 1Finity and Liqid to deliver "seamless all-photonic datacenter connectivity," a marketing phrase for a real engineering problem: keeping a cluster running as if it were one machine when it is physically split across rooms, buildings, or subsea cables. And on the cross-border side, Edgecore has joined NTT and Chunghwa Telecom on a project to advance AI data centers that span Taiwan and Japan, an early example of the kind of geography-distributed AI buildout the company says is coming.

Read together, the announcements sketch a specific bet: that the next phase of AI infrastructure growth will look less like a single 1-gigawatt campus and more like a federation of mid-sized sites, with optical fabric as the load-bearing layer between them. For Accton, that is a route into the hyperscaler supply chain through open hardware rather than proprietary boxes; for operators, it is one answer to the question of what to do when the local substation says no.

Behind the vendor story is a broader shift in how AI infrastructure gets bought. The pitch targets buyers who already use OCP-aligned open hardware, the open-compute standard that lets them mix and match switches, optics, and software from different vendors instead of buying a vertically integrated stack from one supplier. Edgecore sits at the center of that trend as one of the largest sellers of OCP-compliant white-box switches, and its partnership with 1Finity and Liqid is explicitly built around making a multi-vendor optical fabric work as a single system. If distributed AI data centers do become a real share of new buildout, the winners will not just be the operators who deploy them but the supply chain that learns to assemble optical fabrics from commodity parts.

The honest caveat is that most of the public evidence here is company-asserted. Digitimes' full article is paywalled; the Edgecore press releases and partner statements describe intent and architecture, not deployment scale. Whether distributed AI data centers become a real share of new buildout, or remain a niche bet for the next few years, will hinge on the same thing the technology is supposed to fix: whether the optical layer can carry the latency and bandwidth of a training job that no longer fits in one building.