Navigate the economics of AI

A token is the fundamental unit of AI work. It’s a small chunk of data: characters of text, pieces of an image, or a slice of audio that an AI model processes. Applications run on tokens. Every interaction—whether training, inference or reasoning—is measured in tokens.
Unlike prior waves of technology investment, AI spend is volatile and nonlinear. AI costs rise not just with adoption, but with:

• workload type
• model complexity
• infrastructure intensity.

This new reality is less about total cost of AI ownership and more about navigating AI economics with precision: tracking, predicting and optimising token use. Tokens are the true unit of value. In the AI economy, they are the currency that translates opaque infrastructure decisions into tangible economic terms: what it costs in AI consumption to generate a dollar of revenue, margin or productivity.

Organisations that address these economics will differentiate themselves not just by adopting AI, but by efficiently converting tokens into enterprise value. Those that don’t risk AI costs drifting into endless experiments—with the very real downside of locking in tech debt with no predictable return.

AI cannot be governed by the same cost models that guided past technology waves. Traditional frameworks—total cost of ownership, per-user licensing or static virtual machine pricing—were designed for predictable workloads and stable consumption patterns. AI is different. Workloads scale in nonlinear ways, consuming resources at unpredictable rates, and AI costs are ultimately measured not in licences or cores but in tokens.

Tokens have become the true unit of AI economics, the common denominator that reveals:

• what organisations are paying for
• how efficiently they are consuming it
• where value is (or isn’t) being created.

This requires a paradigm shift—where a CIO may need to think like a CFO, while a CFO may need to think like a CIO.

The implications for C-suite leaders are profound. Without discipline, AI costs drift quietly upward, hidden in SaaS renewals, spiking in unpredictable API bills, or locked into infrastructure commitments that can’t be unwound. And unlike prior technology cycles, where budget overruns were frustrating but manageable, AI overspend directly erodes competitiveness.

• Packaged AI solutions (e.g., Agentforce, ServiceNow Copilot, Workday AI) abstract tokens entirely. Leaders see a predictable subscription or per-seat fee, but little transparency into token consumption efficiency. The risk is overpaying for simplicity.
• API and model consumption (e.g., Anthropic, OpenAI, Together AI) make tokens explicit. Every query is metered, billed and exposed. This brings transparency, but also volatility: costs rise based on workload design, prompt length, and hidden choices of infrastructure providers. AI costs go up due to a token meter running in real time.
• Self-hosted AI factory (internalised token costs). Tokens are the output of decisions about open-source model choices (e.g., Mistral, Llama, etc.), GPU architecture, network interconnects, storage tiers and energy contracts. This path demands high capital expense and skilled teams but grants unmatched control over long-term unit economics and data sovereignty. Large-scale adoption of this archetype is typically referred to as an "AI factory".

There are three major factors to token pricing:

1. the underlying AI tech stack
2. how it is hosted and consumed
3. what type of model and level of customisation is required to power the solution.

Every token processed by a model reflects a cascade of infrastructure decisions. For package buyers, these costs are hidden. Costs are abstracted, bundled into familiar enterprise contracts and vendor managed across every layer of the tech stack, which make unpacking your total cost of ownership challenging.