AI Tokenomics Personas and Operating Model

• Owns the platform layer and the Total Cost of AI (TCA): orchestration, routing, gateways, observability. Platform-layer spend can exceed raw token cost; tokens are 8–27% of total agent run costs ^[9].
• GPU utilization and model serving: KV cache grows with context length and concurrent users until, on long-context workloads, it can cost more than serving the model itself.
• Model lifecycle: frontier models replaced within months; infrastructure commitments outlive model shelf life.
• Effective cost is a systems property. IBM Research: Claude Sonnet 4.6 cost roughly half of GPT-4.1 across 417 tasks despite higher list prices, because cache-read pricing interacted with context reuse ^[10]. Engineering owns the serving conditions that produce results like this.

• Pricing is non-linear: tokenizers, caching, context windows, and routing move effective price by large multiples. Model choice alone swings cost 9.1x for identical workloads ^[9].
• Anomaly detection shifts from spend spikes to complexity shifts: agent loops, runaway sessions, prompt injection attacks.
• Visibility into opaque spend: SaaS vendors bundling AI costs, multi-provider routing, credit-based pricing. FOCUS 1.2 added virtual-currency columns for this ^[11], and 1.5 goes considerably further: model developer, family, ID, and version are merged into the working draft, so AI spend can be broken out by model without vendor-specific parsing, and a principal identifier is in flight to attribute a charge to a person, team, or automated agent ^[12]. One thing 1.5 does not settle: separating cache-served tokens from freshly processed ones is still under consideration rather than committed, so the cache hit rate that drives much of the cost may not appear in billing data this release.
• Change management: the questions the rest of the organization needs asked about why AI does not behave like traditional cloud compute. Forecast variance is expected when consumption moves this fast; the answer is agility, re-forecasting frequently as usage patterns shift, not a more accurate single forecast.

The subtypes are consumption surfaces, not departments. The same person moves between surfaces in a day:

• Interactive prompt: a human in the loop on every turn: drafting, Q&A, review. Spend scales with seats and habits.
• Agentic: the user initiates a task and the agent runs, sometimes in parallel sub-agents, sometimes for extended periods. Spend scales with task depth rather than headcount, and the failure mode is a loop or a runaway session, not a bad prompt.
• Embedded: AI baked into tools the user already uses: autocomplete, drafts, search. Invisible to the user, and through bundled SaaS pricing often to the organization as well.

Two further axes split the same user: who provisions the surface (a platform or IT-managed seat, a department-managed tool, a personal subscription the organization never saw), and whether the AI is Internal (augmenting the user’s own work) or Product (embedded in what the user delivers to customers). The tools, the providers, and the spend profiles differ across all three axes.

• Demand: adoption, usage patterns, and the interactive vs. agentic mix are the demand signal that every forecast, reservation, and capacity plan on this map runs against.
• Quality floors in use: the floor Product or Engineering sets is only as real as whether the user’s actual task clears it. Re-prompts, corrections, discarded outputs, and workarounds are the cheapest quality data in the stack, and they sit with the user.
• Off-platform spend: personal keys, personal subscriptions, consumer tools pointed at work data. The spend that FOCUS-conformant data never sees, and where a material share of the data risk sits.
• Value realization: intelligence becomes business value at the point of use, so usage is the first value signal you get. When a tool sits unused, the value case, the quality floor, or the tool itself is wrong somewhere.

• Unit-cost forecasting does not apply, but not for the reason usually given. Frontier list prices are broadly flat year over year; what falls is the cost of clearing a given quality bar, as cheaper tiers become good enough and caching and routing improve. Cheaper useful output then pulls in more demand, and the total climbs. A forecast built on falling per-token prices will be wrong twice over.
• Product AI (growth and margin) vs. Internal AI (cost reduction): budgets, hurdle rates, and review cadences should differ by posture.
• Value chain: energy and capital to intelligence, to business value, with loss at each stage. Finance owns quantifying the last conversion.
• Reasoning and agentic workloads shift the consumption mix underneath any forecast; the same feature can consume an order of magnitude more tokens after a model change, with no procurement event to flag it.
• The token line is not the whole bill. Platform tooling, evaluation, retrieval infrastructure, and the engineering and review effort required to make output usable all belong in the AI business case, and a case built on model pricing alone will understate cost and overstate return.

• Each use case maps to a spectrum: smaller models with short context at modest speed through larger models with long context generating quickly. Over- or under-specifying is waste in both directions.
• Quality floor per task type: define the minimum acceptable output, then optimize cost within that constraint. Agree that floor with Engineering and hold it per task, not on average.
• Different metrics per initiative: growth for product-facing AI, efficiency for internal tools.
• Pricing the output is now a product decision. The market is converging on outcome-denominated pricing for agents precisely because per-token and per-action pricing pushed forecasting risk onto buyers ^{[7] [8]}.

• Multiple sourcing channels: direct from model labs, cloud marketplaces, specialized inference providers, bundled credits. No single hyperscaler offers reservations across all top-tier models, so multi-source is structural.
• Physical scarcity: HBM and DRAM effectively sold out through 2026, fabrication capacity short, data center power as the binding constraint ^{[1] [2] [6]}.
• Model expiry does not match commitment terms: you cannot commit for three years to something with a six-month shelf life. Provisioned capacity can cost more per token than pay-as-you-go even at full utilization. See the FinOps Navigating GenAI Capacity Options paper for vendor breakdown ^[13].
• Credit and virtual-currency billing complicates comparison shopping; demand FOCUS-conformant billing data as a contract term where possible. FOCUS 1.5 is adding a SKU price dataset that carries prices independently of consumption, which is what lets you compare model rates across vendors before committing rather than reconstructing the comparison from invoices afterward ^[12].

• AI spend is reaching capital-allocation scale. Hyperscaler capex guidance of ~$725B for 2026 signals that compute posture is now a board-level topic even for consumers of AI ^[5].
• Leadership owns the Product AI vs. Internal AI split at the portfolio level: how much of the AI budget is a growth bet vs. an efficiency program, and what evidence would change the mix.
• The value mandate: the market is broadly “doing AI” rather than solving named business problems. Leadership forces the naming, because you cannot measure value until you name the business problem.
• Commitment horizons stretch beyond normal technology governance: multi-year capacity, power, and vendor agreements need executive and Legal engagement early, not at signature.

• Capacity roadmaps vs. demand waves. Demand: token volumes growing 7x YoY at the largest operators, enterprise generative-AI spend roughly tripling YoY in 2025 ^{[3] [4]}.
• GPU memory supply constraints and fabrication timelines. HBM and DRAM allocation reserved years ahead ^{[1] [2]}.
• Data center power availability as the binding constraint on growth. Data center electricity is projected to roughly double to ~945 TWh by 2030, and grid-connection queues are already delaying projects ^[6].
• Supply forecasting that accounts for physical scarcity, region, and hardware generation. Pinterest estimates Hopper-to-Blackwell at roughly 30x in token economics ^[14].
• Energy procurement and legal agreements locking in capacity, including power purchase agreements. Technology companies are now among the largest corporate buyers of renewable power.