AI is already reshaping your network. Why isn’t it reshaping your revenue?

As AI workloads demand deterministic performance, traditional network economics are changing. This article explores why telcos must rethink their role in the AI value chain.

Traffic is trending symmetrical, latency budgets are tightening, and a new kind of SLA is emerging. AI is creating new sources of value for telcos, but monetizing them will require defining where and why they belong in the AI value chain.

Will telcos merely enable the AI economy, or will they participate directly in the value it creates? As AI workloads move from centralized training to distributed inferencing, the answer will increasingly depend on where providers choose to compete. From deterministic connectivity and AI-grid interconnection to inference hosting and infrastructure services, AI is creating new opportunities to move beyond traditional transport economics. As the first post in a multipart series, this blog will focus on the connectivity layer because it is where AI’s impact on the network is already visible. Future posts will examine the broader opportunity across AI-grid interconnection, inference hosting, and infrastructure services.

Sixteen milliseconds

That is roughly the total time to render a single frame of video at 60 frames per second, the window of time in which an AI agent must capture information, infer, and update its interactive digital human image, or the user sees the failure. Conversational voice stops feeling live at the slightly slower rate of a few tens of milliseconds, impacting AI voice agents. And for AI-enabled robotics, industrial automation, and even self-driving cars, similar delays stop being a quality problem and become a safety one.

These are the operating budgets of workloads already running across service-provider networks for a customer who sees deterministic network performance as a functional dependency, not simply a connectivity line item. That dependency, not traffic volume, is what makes the AI connectivity experience different. And, while not every AI workload operates inside these windows, a growing class of interactive, real-time, and machine-oriented applications is beginning to expose latency requirements that traditional network service-level agreements (SLAs) were never designed to underwrite.

For twenty years our industry has anticipated the next critical app, which would finally reverse the commoditization of connectivity. We rode the successive waves of broadband, mobile data, video, and cloud. Each created more traffic, but none of these changed what enterprises were fundamentally willing to pay for. AI appears to be not only different, but quite possibly the most significant tailwind the industry has had in over a decade. But the benefit will accrue only to operators willing to reconsider what they are selling.

Why this wave behaves differently on the network

For three decades, the industry engineered for north–south, asymmetric flows. However, as AI adoption continues to grow exponentially, AI factories are being distributed, inference models are being deployed closer to the edge, and agents are being deployed in applications sitting closer to the end-user, resulting in a shift in traffic flows from north–south to east–west and from asymmetric toward symmetric.

This does not, by itself, end oversubscription, but it challenges the directional assumption that shaped a generation of network design. As enterprises train, customize, replicate, and synchronize models across cloud, data center, edge, and sovereign environments, latency is shifting from a preference to a functional requirement for an increasing number of workloads.

The challenge is not simply to lower average latency. It is the delivery of consistent performance across defined traffic classes, paths, locations, and workload contexts, with enough telemetry to prove when the service envelope was met. This is the difference between a forecast and an observation, and it is being asked for now. According to Omdia research commissioned by HPE Juniper Networking, 87% of enterprises using AI require at least one premium network performance characteristic. Compared with conventional applications, the gap in premium performance requirements represents a significant opportunity for service providers.

A guarantee is not an SLA

Yet, the industry has sold performance SLAs for twenty years. For example, the Metro Ethernet Forum (MEF)’s Carrier Ethernet specification already provides metrics for frame delay, jitter, and loss. So, why is SLA performance for AI services new?

The difference is the reference point. A traditional SLA measures the network in isolation. It addresses basic connectivity questions like, is the port up, and is the mean latency within the agreed-upon objective? As an example, traditional carrier SLAs commonly define latency as round-trip network delay measured through periodic probes and averaged over a calendar month. These measurements are typically limited to the provider’s managed network domain and often exclude local access facilities, customer equipment, and application performance.1, 2, 3 If the provider fails to meet the SLA target, the remedy is generally a service credit tied to a portion of the monthly recurring charge rather than the business impact of the workload.4

But AI workloads are highly dependent upon individual transactions occurring deterministically. So, while an operator can hit five-nines availability over the period of 30 days, it can still fail every latency-critical AI workload on that network, because uptime and mean monthly latency say nothing about whether, say, transaction number four million met its 16-millisecond budget. The guarantee that AI requires is measured against the application’s outcome, not the network’s health. It couples the operator to the workload’s success rather than the link’s condition. This is exactly the scope carriers have historically kept outside the boundary of standard connectivity contracts. This new form of SLA no longer focuses on certifying the pipe, it now underwrites the workload.

Yet, it does not mean operators must guarantee every element of the application stack. It means they must define the service envelope where they can observe, control, and enforce the performance characteristics AI workloads increasingly require.

Why this time may actually be different

Four times in twenty years, whether the shift from dial-up to broadband, pivot to mobile data, introduction of streaming video, or transition to the cloud, our industry expected a traffic wave to reverse commoditization, and four times the value accrued elsewhere. Each asked the network to carry more, but none asked it to guarantee anything. A workload that needs more capacity rewards whoever is cheapest, which is why each prior wave has further commoditized the network as opposed to reversing the trend.

But a workload that needs a bounded, enforceable, per-flow, and application-level guarantee rewards whoever can underwrite it, which is a fundamentally different game. However, this is not a promise that telcos will automatically win. Others are already working to capture the same value by keeping the network commoditized so the value will be associated with the software, cloud, or application-layer services.

  • Hyperscalers will push inference to their own edge in order to minimize round-trip distance and delays.

  • SASE vendors will look to bond multiple links, apply forward error correction and intelligent path selection to approximate deterministic behavior over lower-cost, nondeterministic transport.

  • Inference model providers, like streaming video content delivery networks (CDNs) before them, will look to leverage edge caching, at telco locations, with creative solutions, like predictive prefetch, to minimize the impact of jitter, so the workload no longer needs a premium deterministic guarantee.

For telcos to win, they must look beyond simply improving their network infrastructure. They need to think strategically about how and where they will play in the value chain, otherwise they will end up spending capital to remain a commodity with better capacity and slightly lower latency. The AI delivery wave creates the opening. It does not assign the winner.

The bottom line

AI has handed the industry an uncommon circumstance: a workload that structurally requires what purpose-built networks are designed to deliver. But proximity to the opportunity is not the same as a claim on the value. Telcos that simply add capacity and wait for revenue to follow risk repeating the pattern of broadband, mobile data, video, and cloud: more traffic, more investment, and limited value capture.

The pipe is no longer the point. The value is in the guarantee, measured against the workload, not just the link.

This is the first post in a six-part series on the AI-native telco. We begin with deterministic connectivity because it is the first visible pressure point. Future posts will explore the broader opportunity across AI-grid interconnection, inference hosting, GPUaaS/CPUaaS, and infrastructure services, all aimed at answering one question: how can telcos participate directly in the AI value chain instead of watching others monetize what their networks make possible?

Learn more at 

The AI-Native telco: Capturing revenue opportunities in the AI value chain.

1Verizon Business Monthly IP Latency Statistics

2Verizon XO High Speed IP Transit Service SLA

3Latency Clause Examples including Windstream MPLS SLA

4Ziply Fiber Ethernet SLA

Meet the author:

Steve Daigle, Global Head of Telco Systems Engineers, HPE
@sjdaig, X (Twitter)
https://www.linkedin.com/in/stevejdaigle/ LinkedIn

This content extract was originally sourced from an external website (HPE) and is the copyright of the external website owner. TelecomTV is not responsible for the content of external websites. Legal Notices