Dual-Resilient DIA: Why AI Workflows Need Two Separate Paths Into the Network

Why the Network Decides How Fast AI Runs

For enterprise AI applications the network is the bottleneck, not the compute you buy in. Between a user’s prompt and the finished result sits a chain of stations: the application, the gateway with authentication and routing, the context of the language model, the compute phase at the provider and the return path of the token stream. Each of these stations is its own trip across the network.

Compute is purchased and, as a rule, available. The path to it is the variable a company actually controls. Nikolaus von Johnston, CEO of SAVECALL, put it this way at the VATM Business X-Change in June 2026: in the AI era the token is the raw material. And a raw material needs a reliable supply chain.

If that chain breaks, it is no longer just email that stops. What stops is the share of value creation that now runs through AI. That is the real difference compared to three years ago.

What Makes an AI Workflow So Fragile

An agentic workflow triggers four to eight tool calls per user request. The language model queries the CRM, a pricing database, stored knowledge, a vector database and finally the mail interface. Between each call the model thinks again. In total that produces 30 to 80 network packets per operation, with an end-to-end runtime of five to fifteen seconds.

Each of these hops has its own transit time and its own variance, and the variances accumulate along the chain. This is why the average value of a line is the wrong metric. What matters is the worst percentile of the measurements, known as the P99 value. It describes exactly the minutes in which a workflow breaks.

Three metrics decide the outcome in practice. Jitter, the variation in packet transit time, should stay below ten milliseconds. Above that, real-time AI loses sequence and speech recognition degrades. Time to first token should stay below 500 milliseconds. Above that, users abort and start again, which doubles the load. And P99 latency has to stay stable, because with eight tool calls it hits every single operation at least once.

How Much Bandwidth an AI Workplace Really Needs

Bandwidth demand per workplace rises sharply with the usage profile. Office work with an AI assistant needs 10 to 25 Mbit/s. As soon as voice and video are processed with AI, demand moves to 25 to 50 Mbit/s. Agentic workflows running continuously in the background need 50 to 100 Mbit/s per workplace.

Compared with a plain office workplace today, that is three to fifteen times the demand depending on profile. Anyone sizing a site now should plan for that range rather than extrapolating current consumption. More important than the absolute figure, however, is whether the line holds that figure at peak. Why bandwidth alone is the wrong metric is covered in detail in our article on AI-enabled networks.

What Dual-Resilient DIA Actually Means

Dual-Resilient DIA means two dedicated internet connections are kept separate at every layer. DIA stands for Dedicated Internet Access, a line with guaranteed bandwidth and no shared usage. The duplication applies not just to the contract, but to every physical component along the path.

Separated are the carrier, the transmission medium, the building entry, the main distribution frame and the backbone route. On the customer side this adds two end devices, two routers, an uninterruptible power supply and automatic failover. The switchover has to happen without manual intervention, otherwise the second line is of little use on a Saturday morning.

Different media are not a detail here, they are the core of the matter. A second fiber in the same trench shares the risk of the first. A combination of fiber and microwave radio, or fiber and mobile, genuinely separates it. Which media are available at a given site becomes clear when you look at the underlay, the physical layer beneath all network services.

Where Apparent Redundancy Fails in Practice

The most common gap is the second line from the same provider. On paper it looks like redundancy. In practice it runs through the same main distribution frame, the same cable bundle and the same backbone route. The company pays twice and still gets one path.

Microsoft describes exactly this trap for Azure ExpressRoute in plain terms. If the primary and secondary connections terminate on the same customer device, high availability inside your own network is given up. If both run through the same port, the partner is forced to give up the separation in its network as well. The same logic applies to internet access.

The second gap is the operating mode. A passive standby line that never carries traffic is never tested either. Microsoft therefore recommends running both paths in active mode. If one path then fails, only about half the flows are affected and recovery is faster. With BFD failure detection, detection time drops from around three minutes to under one second.

The third gap is the duct route. The most common cause of fixed-line outages remains construction work. Anyone procuring redundancy should have the separate duct routing confirmed in writing and check the building entry on site. Two fibers arriving through the same conduit are one line with two invoices.

How AI Hyperscalers Enter the Redundancy Chain

Redundancy does not stop at the building wall, it stops at the AI platform. Anyone using language models in Azure, at AWS or with a specialist provider otherwise sends that traffic across the public internet, with all the variance that is normal there. Private direct connections avoid this.

Azure ExpressRoute and AWS Direct Connect each provide a dedicated path into the cloud. AWS ties its highest resiliency model explicitly to separate connections terminating on separate devices in more than one location. That is the same logic as the building entry, just one layer further up.

The market is catching up here. In April 2026 Equinix launched Fabric Intelligence, a private marketplace for AI services through which companies reach inference, training and storage without sending sensitive data across the public internet. Microsoft introduced Multicloud Interconnect, a managed private path between Azure and AWS built on ExpressRoute and Direct Connect with a quad-redundant design. Anyone procuring now should factor these building blocks in from the start.

What an Outage Costs and What It Is Really Worth

For most companies the cost of an outage sits far above the cost of the second line. The Uptime Institute reports in its outage analysis for 2026 that 57 percent of major outages cost more than 100,000 US dollars. In one in five cases the damage exceeded one million US dollars, for the second year running.

On top of that comes the share nobody books as an incident. If thirty managers each wait half an hour a day for AI responses, that adds up to a five-figure sum per month. This waiting time appears in no outage statistic, because technically nothing failed. It is paid for all the same.

Nikolaus von Johnston describes the mismatch with an everyday comparison. A single internet line, he argues, is an inappropriate operational risk for a company. On the leased car of one individual employee, that same company spends more than on protecting the infrastructure the entire business depends on.