Cisco is betting that token consumption, not compute raw capacity, becomes the budget constraint enterprises can’t ignore as AI agents replace intermittent chatbot queries with continuous, machine-to-machine workloads. Jeetu Patel, Cisco’s president and CPO, pegs current agent deployment at under 1% of potential users, which he reads as a structural supply shortage ahead, not a bubble. Cisco’s response is a unified management plane called Cloud Control<\/a> that routes inference across cloud, private data center, and endpoint, with AMD handling compute and Cisco owning the network, security, and token visibility layer.<\/p>\n The story is really about who owns the economic control plane of agentic AI, and Cisco is making a direct play for that position. If your inference workloads are already fragmented across a hyperscaler, a colocation facility, and a growing fleet of endpoint devices, the tokenomics problem, meaning visibility into which agents are burning tokens at what rate and whether that spend is justified, is real and largely unsolved today. Organizations running primarily on a single cloud with a handful of tightly scoped agents are largely insulated for now. Those operating multi-environment infrastructure with ambitions for broad agent deployment are the ones this story is actually about.<\/p>\n The small language model argument here is sharper than it first appears. Cisco’s Antares release, a purpose-built open-weight model for code vulnerability detection that runs without exfiltrating proprietary data, isn’t a product announcement dressed as strategy. It’s a proof of concept for a routing principle: frontier models like those from Anthropic or OpenAI are expensive, general-purpose, and carry data residency risk when the query involves sensitive code. A task-specific model that catches 70% of vulnerabilities locally, with a frontier model handling the residual 30%, is a meaningful cost and risk architecture, not a theoretical one. The pressure on CTOs to build that kind of tiered routing logic, rather than defaulting every agent task to the most capable model available, will only intensify as token costs accumulate at scale.<\/p>\n The claim that ease of use, not compute, is the binding constraint on adoption is where Patel’s argument is most interesting and most worth interrogating. He’s right that eight billion people activating thousands of agents each is nowhere close, but for the enterprise CTO, the relevant ceiling isn’t mass consumer adoption, it’s whether the orchestration and governance tooling can keep pace with the agents already being deployed inside the firewall. The falsification condition for Cisco’s whole thesis is straightforward: if hyperscalers build credible token telemetry and routing natively into their own control planes, Cisco’s management layer becomes redundant before the market matures. That’s the vendor renewal you should be stress-testing now.<\/p>\n In AI infrastructure, tokenomics refers to the economics of token consumption, where a token is roughly a word fragment processed by a model during inference. Because agents run continuously and chain calls between multiple models, token usage compounds in ways chatbot deployments never did. Think of it like metered water pressure: a single tap running occasionally is manageable, but dozens of taps open simultaneously, feeding each other, creates a volume problem that requires measurement before it can be managed. Tokenomics is the discipline of tracking that consumption against business value.<\/p>\nWhat this means for your business<\/h2>\n
Concept deep-dive: Tokenomics<\/h2>\n