Follow the AI tokens

Many core enterprise systems—enterprise resource planning, human resources, customer relationship management and marketing platforms—are rapidly embedding agentic functionality. Historically priced by seat or transaction, these products are increasingly driven by token-based consumption. As agent usage expands and demand patterns mature, token-driven costs will surface—either explicitly or buried deep in contract language. CTOs should have in mind that opaque token exposure will become a defining feature of enterprise software negotiations and that new cost models will continue to emerge, with agent usage and task complexity becoming the basis for token-based chargebacks.

For application programming interface (API)-based solutions, most organizations begin AI experimentation on cloud for pragmatic reasons: cloud-resident data, cloud-native applications, and rapid access to innovation through both open and proprietary models. But model choice is no longer a technical preference—it can also be an economic decision. Cost per million tokens varies widely by model class, and those differences compound quickly at scale. Model selection also has direct implications for hosting strategy, data gravity and long-term flexibility as new providers enter the market and incumbents adjust pricing and positioning.

Self-hosted options may give CTOs the greatest control over the full AI-native stack—and over token economics—but can require deliberate capital investment to build or retrofit AI-ready infrastructure. For predictable, high-volume workloads, this control can unlock materially better unit economics. For others, it may introduce new trade-offs between flexibility, capital intensity and operational complexity.

Model selection can have a direct impact on the cost per million tokens and also implications on the hosting strategy.

The market is abuzz with discussions over whether to use open, closed, derivative or proprietary models to power AI builds. As new providers enter the space and established leaders develop and shift their strategies open models are fast closing the gap to frontier or closed models that have been the predominant choices for AI programs to date.²

The way to lose control of tokenomics is to treat all prompts equally. AI systems operate under a three-way trade-off between accuracy, latency and cost, and many enterprise workloads do not require frontier-scale reasoning.

Large language model (LLM) routing is emerging as a core enterprise control mechanism. By dynamically matching prompt complexity to the minimum viable model, CTOs can directly shape demand, latency and token burn in real time. High-cost reasoning models can be reserved for high-value tasks, while routine requests can be handled by smaller or fine-tuned models. The result is a tiered intelligence stack that aligns cost with business value—rather than burning tokens indiscriminately.

Beyond routing, CTOs can also systematically apply efficiency levers such as prompt optimization, caching, batching and context window constraints. These techniques rarely change the user experience but can materially reduce token consumption at scale.

Tokenomics are governable—but only if token spend is treated as a managed financial asset, not a byproduct of experimentation. Continuous monitoring of token usage, GPU hours, storage, egress and energy consumption is now common.

To add to that, CTOs should consider also pairing technical telemetry with financial controls: chargeback mechanisms, usage caps, and scenario-based forecasting grounded in real workloads. Left ungoverned, token spend escalates silently. But governed properly, it can become predictable and defensible—even in the face of volatility.

Contracts deserve equal scrutiny. Per-token escalators, bundled minimums, data residency fees and restrictive licensing tiers can quickly erode flexibility. Token economics should therefore be reflected in procurement and vendor governance strategies, reinforcing the importance of hybrid and modular architectures that avoid one-way-door decisions.

Infrastructure choices directly shape tokenomics at scale. CTOs and infrastructure leaders can consider sourcing GPUs from multiple locations to drive resiliency and flexibility. Neocloud providers offer alternatives to hyperscalers through high-performance GPU-as-a-service models, while on-premises AI factories provide control and data proximity for specific use cases.

Architecturally, if you decide to build an AI factory, several decision points arise: liquid cooling approaches; power and cooling sourcing; and networking choices such as InfiniBand versus Ethernet. Storage and networking can both become bottlenecks to GPU utilization.

There is also the choice of software required to enable a hybrid AI strategy across hyperscalers, neoclouds and on-premises environments.

Capital expense versus operating expense directly shapes AI token unit economics: For predictable, high-volume workloads, self-hosted AI infrastructure can deliver sharply better long-term unit economics—but requires upfront capital. Flat IT budgets will likely not suffice; capital reallocation is often necessary to unlock efficiency.

Token volume thresholds determine when ownership economics outperform consumption: Deloitte analysis shows that, for many enterprises, there exist tipping points on ownership economics vs. consumption and those should be understood. Translating technical thresholds into financial break-even milestones is often essential to informed decision-making.

Architecture choices determine financial risk and flexibility in AI token economics: Well-architected hybrid models mitigate both spend risk and strategic rigidity, enabling adaptation as technology pricing and business demand evolve.

AI tokens are a direct driver of productivity and measurable business value: Tokens are not overhead. Each one directly enables productivity, automation or customer impact. Govern them rigorously—but invest in them deliberately, with FinOps discipline to ensure measurable ROI.