The Toll Road Theory of Agentic AI

For the last three years, the enterprise AI conversation has been a model conversation. Which foundation model is best? Which fine-tune, which retrieval system, which agent framework? The agentic-AI pitch from Cisco President and Chief Product Officer Jeetu Patel is that this framing is already obsolete. Agents act, reason, and loop continuously at machine speed, replacing the click-and-wait cadence of traditional software — and every one of those loops is a network event.

"Every agentic action is a routing challenge, a trust decision, and a telemetry event," Patel wrote in a June 2026 Cisco blog post. If that is even roughly right, the agentic-AI value chain will be settled at the network and control-plane layer, not only at the model layer — and the network team, not the data science team, may end up holding the keys to whether enterprise AI works at scale.

The $2.1 billion tell

The financial signal is hard to ignore. In Cisco's Q2 fiscal 2026 earnings call, CEO Chuck Robbins and CFO Mark Patterson reported that AI infrastructure orders from hyperscalers hit $2.1 billion in the quarter alone — equal to the entirety of fiscal 2025. The company guided to more than $5 billion in AI orders and more than $3 billion in AI infrastructure revenue from hyperscalers for the full fiscal year, with networking product orders growing more than 20% year over year.

Those are Cisco-reported numbers from a vendor with skin in the game, but the trajectory is consistent with a thesis Patel has been pushing for months: that "tokens per dollar per watt is the new currency" of the data center, and that networking sits at the fulcrum of that equation. Treat that slogan as Patel framing, not industry consensus — but the orders growth is harder to dismiss as marketing.

The toll-road thesis

The strategic posture behind the numbers is what one analyst called the most coherent platform argument Cisco has made in years. Mike Leone of Moor Insights & Strategy used that phrase at Cisco Live US in June 2026, where the company unveiled Cisco Cloud Control — an umbrella platform that replaces Security Cloud Control and stitches together AI Canvas, MCP tools, third-party agent management, and a Cloud Control Studio with an Agent Builder and OpenAI Codex integration slated for later in 2026.

The structure is familiar. Just as Cisco collected rent on routing, switching, and security for two decades, the bet is that agent operations — the orchestration of AI agents across compute, data, workplace, and security domains — will sit on a control plane, and that control plane is the new toll booth. Bob Laliberte of TheCube Research summed up the industry mood with a quip that has aged fast: "Networking is cool again."

What changes for operators

If the toll-road theory holds, four things that have been treated as background plumbing become first-class planning problems for IT and finance.

Capacity. Agentic workloads are not request-response. A single user query can fan out into dozens of model calls, tool invocations, and data retrievals. Cisco's own research — which has not been independently corroborated and should be treated as vendor data — points to multi-x traffic growth on enterprise WANs and large per-task traffic lifts as agents chain together. Whether those numbers replicate under independent measurement is an open question, but the directional pressure on WAN and egress capacity is real.

Observability. Continuous agent loops produce continuous telemetry. Cisco's bet is that a Splunk-based Cisco Data Fabric provides the cross-domain telemetry plane needed to make agent operations legible. Whether the Splunk integration has matured past what Omdia analyst Jim Frey called a "honeymoon period" is one of the more important open questions for the platform-cohesion claim.

Trust. Cisco's February 2026 security redefinition for the agentic era — AI Defense, multi-turn red teaming, real-time agentic guardrails, AI-aware SASE, secure routing, and post-quantum crypto in IOS XE 26 — is the trust layer of the thesis. If agents act on behalf of users and tools, the network becomes a natural enforcement point. Smart switches that fuse packet forwarding with on-box inspection extend that enforcement to the access layer, while Hypershield and Secure Access extend zero-trust to AI agents and the tool calls they make.

Egress economics. Patel's "tokens per dollar per watt" framing is a network framing. If per-task AI traffic grows an order of magnitude, the egress bill — long an afterthought in enterprise architecture — becomes a line item finance notices.

The counter-position

The toll-road theory is not settled. Cisco is a vendor with skin in the game, the research underpinning its traffic-growth claims is internal, and incumbents benefit when "the future" is defined as their core domain. Hewlett Packard Enterprise and Arista are pushing alternative control-plane stories; Arista's data-center switching posture in particular overlaps directly with the hyperscaler AI infrastructure orders driving Cisco's numbers. Mauricio Sanchez of Dell'Oro, while noting that Cisco SASE share has climbed roughly 20% since 2023, is measuring share within a market Cisco already owned; whether share gains translate into platform lock-in at the agent operations layer is a different question.

Cisco's claim of 50-plus partners for cross-domain AI agent support is also a marker, not a finish line. Omdia's Frey noted that this brings Cisco into a partner environment competitors — including the hyperscalers whose orders are powering the current growth — already operate in. The bet is that the network and control-plane layer remains separable from hyperscaler stacks. That is not obvious.

What to watch

Network Actions, the network-native AgenticOps product Cisco put into beta in June 2026, is the operational test of the thesis. It ships with telemetry, root-cause analysis, and a digital-twin alpha for safe experimentation on switches and routers — all early-stage, all caveated. If Network Actions produces credible customer deployments before the next fiscal year, the platform argument hardens. If it stays a beta with Cisco-only deployment data, the toll-road framing stays a thesis.

The question for enterprise IT and finance planners is not which model wins. It is whether the network is architected to handle continuous, autonomous agent loops, what the egress bill looks like when per-task AI traffic grows, and who controls the control plane. Cisco is betting the answers favor its core domain. The data so far is suggestive. The data through the rest of fiscal 2026 will be dispositive.