Cloud infrastructure is traditionally where the majority of artificial intelligence (AI) workloads have been processed but as data sovereignty, operational resilience and on-device AI gain importance, cloud infrastructure is not the right home for all AI workloads. Looking to take a lead in the area, Extreme Networks is set to introduce Exchange Zones to bring AI processing to local compute.
Scheduled to launch in the first half of 2027, the system aims to bring enterprise AI to premises, allowing users to choose their level of data sovereignty and control. That means running the same AI applications locally and in the cloud wherever their data originates and workloads run, with control over data residency, processing and cost.
“As AI moves toward the distributed sites where enterprise data is generated, organisations need to run AI applications on-premise to meet data sovereignty requirements and control inferencing cost, yet many AI marketplaces still assume a cloud-only model,” said Brandon Butler, senior research manager of networking and infrastructure services at analyst firm IDC.
Extreme said that Exchange Zones will bring data encompassing network management, security, business operations and logistics on-premise, unlocking use cases that cloud-only architectures can’t yet deliver. Users will be able to pool compute resources, including spare capacity on its access points and switches – for those that have already deployed at branch offices, retail locations and other distributed sites – to run AI applications and plug-ins.
At a conference in Amsterdam, Nabil Bukhari, CTO and president of AI platforms at Extreme Networks, noted that after decades of building networks that reach every corner of the enterprise, these networks hold a lot of idle compute. Exchange Zones will put that compute to work, letting AI run locally on infrastructure customers already own.
“[We have already launched] technology that helps you run the network either ambiently or autonomously,” said Bukhari. “So, [we thought] what else can we do with our networks? Your networking infrastructure produces high density and high richness telemetry data, operational data, analytics on behaviour, traffic – none of those are used or needed to manage a network. Most of the time, all cloud systems drop that data … because carrying [it all] into the cloud is a lot of cost.
“But what if you could use that data – you can come up with use cases, and then what happens? You hit into the wall of reality. Then comes the question of sovereignty: which cloud am I going to put [data] into? Can I send my medical records into X cloud or Y cloud? Then you run into localisation. If you have operational technology networks, good luck saying that I am going to use [cloud] data in real time to make decisions around my OT system. Your data and your inferencing must be together.
“Then you run into a survivability problem. What happens if your cloud connection breaks? And the last part is economics. It’s easy to say, ‘I’m going to take all of this high density, high value data, and I’m going to put it into the cloud’ – every bit that goes into the cloud costs you.”
As a solution, and bringing localisation to the fore, each Exchange Zone will combine an Extreme Universal Compute Platform or a qualifying customer-owned server with spare capacity from access points and switches. That is designed to place workloads to meet automatically customer-defined requirements for latency, data residency and resilience. Models run locally by default, and cloud inference is available only when customer policy allows it and is logged whenever it occurs.
With respect to the issues raised by Bukhari, Exchange Zones will be deigned to run exchange applications and their models inside a boundary the customer controls. That means addressing that some applications are not able to tolerate the delay of a round trip to the cloud, the economics and the need for regulated industries and distributed operations to gain tighter control over where AI runs.
In particular, Exchange Zones will keep AI inferencing local, and each zone is bound to a single site in a single country with observability across residency records, model usage and data life cycle management. This will show what stayed local, what crossed a boundary and why, and which model was resident.
Looking at survivability and localisation, Extreme stressed that critical operations must survive cloud outages. Exchange Zones will run time-sensitive AI locally, keeping retail stores, warehouses and distributed sites operational when connectivity degrades.
As sending high-volume data to the cloud racks up transport, storage and recurring inference charges. By running AI on compute customers have already deployed, Exchange Zones will look to reduce data movement and cloud inference costs for the workloads that matter most to companies.
Looking at the marketplace in which the new product will operate, Bukhari stressed that Extreme was not looking to compete with edge AI companies which typically have their own operating model.
“We are solving problems or we are accelerating use cases that are related to high-density network data,” he said. “We gave [the system] an AI runtime so that it can actually run agentic AI applications – it’ nothing to do with managing your network, but everything to do with utilising the data from your network. So, that is a zone and you can run a zone locally. It could be in every building – [whether] school, store, medical office.”
Extreme Exchange Zones is part of Extreme’s enterprise stack which comprises: network infrastructure, where data is generated and compute already sits; Extreme Platform ONE for unified management and security; Extreme Exchange to extend AI skills and applications; Extreme AI as the intelligence core; and Extreme Agent ONE as the ambient and autonomous AI interface, offering as a new class of AI agents for enterprise networking that moved beyond generic, prompt-based AI.
As it was making the launch of the Exchange Zones, Extreme also announced “strong early momentum” for the Coworker built-in, ambient, context-aware AI agent for Agent ONE. Specifically, in the first month of availability, customers are said to have collectively saved an estimated 1,000 hours of troubleshooting, and users responded to 98% of critical nudges generated by the system to make contextual recommendations within the workflow.
Early engagement with Agent ONE Coworker was reported to be roughly 10x that of Extreme’s first-generation conversational AI agent, with the company stating this signals that enterprise organisations are ready for AI that proactively surfaces potential network issues.
In addition, 41% of Extreme Platform ONE customers used Agent ONE for daily tasks; 58% of support cases opened through Agent ONE Coworker arrived with all the evidence the support team needed to start solving the problem; and Extreme Agent ONE is claimed to have helped customers resolve issues in minutes, compared with the industry average of two days.”
“Conversational AI only helps when people already know what to ask. That’s a design limitation, not a user problem,” Bukhari said. “We built Extreme Agent ONE Coworker to reason continuously over live network context and act on what it finds, under governance our customers control.”
The next step in Extreme Agent ONE’s evolution is “operator mode”, expected to be available in the first half of 2027. It will work independently within defined governance boundaries, respond to events in real time and run scheduled tasks without constant human direction.
