Token consumption is becoming a material input to enterprise work. The strategic question is what that spending builds: useful workflows, proprietary evaluation knowledge, durable output, and the capability to improve the next investment.
Microsoft CEO Satya Nadella's June 2026 discussion of human capital and token capital gives that question a name. Managing the two together means measuring what people and AI accomplish, the full cost of that work, and which capabilities remain useful after a model call ends.
What Is Token Capital, and Why Does It Matter?
Token capital is shorthand here for the repeatable, measurable capability an organization builds through token-powered work. Tokens are an operating input. Reusable workflows, evaluations, proprietary context, effective oversight, and durable results are what the investment may produce.
Token spend is what the organization pays. Token capital is the business capability that spending helps develop. A larger token bill does not establish that capability, and generating more responses is not the same as creating more lasting value.
The balance-sheet framing is strategic, not an accounting classification. Calling something token capital does not establish an asset's recognized value or make AI expenditure eligible for capitalization. Finance should assess individual costs under the applicable accounting rules and policies, separately from this management concept.
Why the Enterprise AI Spend Picture Is Hard to Explain
AI costs can appear in subscriptions, provider invoices, cloud services, gateways, agents, and departmental expenses. Supporting infrastructure, implementation, governance, review, and rework add costs that a token invoice does not show.
An aggregate bill can grow because adoption expands, models change, agent sessions run longer, retries increase, or a temporary credit ends. None of those explanations establishes whether the organization received additional value.
Establish a reconciled cost scope before evaluating token capital. Show billed spending and observed consumption separately, normalize periods and currencies, document shared costs, and retain unknown attribution rather than assigning it arbitrarily.
Measure What the Tokens Produce
The value-per-token guide extends cost visibility into outcome economics. Two workflows can consume the same token volume while producing different quality, review burden, and business results.
- Attribute consumption and cost. Connect the source, tool, model, team, agent, workflow, and use case where records support it.
- Define an accepted result. Choose a unit that reflects the work, such as a verified review, resolved customer issue, completed transaction, or durable engineering change.
- Measure the outcome and guardrails. Include quality, cycle time, rework, human review, and relevant business measures.
- Compare the relationship over time. Use comparable work, consistent denominators, and clearly defined cost scope.
Tokens per feature measures consumption relative to a defined engineering deliverable. Cost per accepted feature measures the associated dollars. They are related but not interchangeable, because model prices, tokenizers, input and output mix, caching, and contract terms can change costs.
There is no universal value-per-token benchmark or single companywide ratio that makes unlike outcomes comparable. Engineering, support, legal, and finance need different units, with shared financial definitions underneath.
Keep the financial terms precise. Benefits divided by cost is a benefit-cost ratio. Net ROI is typically expressed as attributable benefits minus the defined costs, divided by those costs. If benefits are not credibly monetized, report operational results and unit costs rather than labeling them financial ROI.
Tokenmaxxing Is Consumption, Not Capital Formation
Most TokenMaxxing Is Unproductive argues against treating token burn as a proxy for effective AI use. A usage leaderboard can encourage expensive activity without showing whether the result was useful.
That article reports an internal Larridin analysis published June 7, 2026, in which the first third of spending was associated with 85% of the measured productivity gains. The article does not provide enough methodology to treat that relationship as a universal threshold or a causal rule. Use it as a reason to examine your own marginal returns, not as a mandate to cut two-thirds of every team's budget. Source: Larridin internal analysis, June 7, 2026.
Ask what another increment of spending produces. Does it support harder work, better quality, or less total effort? Or does it add retries, repeated context, abandoned sessions, and review? High spending can be productive; low spending can still be poor value.
Set workload-appropriate guardrails with accountable owners. Investigate cost spikes and declining outcome efficiency without punishing heavy users simply for using more AI. Remove incentives that reward consumption rather than accepted work.
Build a Tier-Aware Token-Capital Budget
The frontier-versus-commodity inference analysis explains why one blended inference-price assumption can conceal important differences. Model generations and tiers may follow different cost trajectories, and a stable listed token price does not guarantee a stable task cost.
Segment the budget by the models and workloads actually in use. Forecast frontier, mid-tier, and lower-cost consumption separately, with scenarios for volume, pricing, model changes, caching, credits, and output length. Treat reported market trends as dated research context, not a permanent two-tier law or a guaranteed future price path.
Test the Premium Before Paying It at Scale
Compare candidate models on the same representative task and acceptance criteria. A lower-cost model is a routing candidate only if it meets quality, reliability, latency, privacy, and risk requirements. A premium model can justify its expense where it demonstrably improves the outcome or reduces the work needed to get there.
Include retries, tool calls, human correction, and downstream effort in the comparison. Cheaper input tokens can be offset by longer output, more attempts, or additional review.
Watch Tokenizer and Workload Changes
The source discusses how a tokenizer change can alter native token counts even when the listed rate remains unchanged. The financial effect depends on workload, caching, and completion behavior. Test actual cost per accepted task after a model upgrade rather than translating token-count growth directly into the same percentage cost increase.
Token counts are not a universal measure of equivalent work across models. Preserve provider and model definitions so comparisons do not mistake a change in tokenization for a change in business output.
Evaluate Reserved Capacity as an Investment in Work
The reserved AI capacity guide adds a commitment question to token capital. Capacity can cost money even when it is lightly used, and a highly utilized deployment may still be a poor economic fit.
Separate provisioned throughput from the reservation or pricing commitment that may cover it. The deployment supplies capacity; the reservation changes eligible billing under the provider's terms. They can require separate management and reconciliation.
For existing capacity, review three questions:
- Are we using what we pay for? Examine typical and peak throughput utilization, and whether any pricing commitment is applied to eligible active capacity.
- Who is using it? Attribute consumption to teams, workflows, applications, agents, and use cases.
- Does the original business case still hold? Compare effective cost with alternatives at actual utilization, or verify the latency, throughput, availability, and data-handling requirements behind the purchase.
Check provider-specific behavior when deployments change or are removed. Stopping execution or deleting a deployment should not be assumed to cancel a separate financial commitment.
There is no universal healthy utilization target. Lower utilization can be justified by a genuine availability requirement; high utilization does not prove better value than pay-as-you-go. Attribution helps evaluate existing commitments, but it does not replace capacity planning or provider-specific purchasing analysis.
Choose Measurement That Connects Cost and Capability
Observability, FinOps, workforce analytics, and engineering measurement provide different evidence. Evaluate their coverage against the question you need answered rather than assuming one category can do everything.
Tools discussed in the original article include Ascerta, Worklytics, and Mavvrik. Their current integrations and scope should be verified against your requirements; a categorical claim that each is limited to one narrow layer is not a substitute for evaluation.
Larridin's measurement approach connects adoption, proficiency, workflows, developer outcomes, and spend across human and agent work. Confirm actual data-source coverage, outcome integrations, allocation methods, and known gaps before calling a view complete.
What Should a Token-Capital Dashboard Show?
- Cost and commitments: Billed spending, observed consumption, fixed commitments, credits, and the defined supporting-cost scope.
- Ownership: Team, department, tool, agent, project, and use case where attribution is reliable.
- Human and agent patterns: Distinct activity and cost views without double counting shared work.
- Model economics: Tier distribution, model changes, effective task costs, and routing-test results.
- Capacity fit: Deployment utilization, commitment application, and the requirement behind reserved capacity.
- Outcomes: Accepted work, quality, cycle time, rework, review effort, and relevant business performance.
- Reusable capability: Validated workflows, evaluation datasets, tested controls, and practices transferred to other use cases.
- Decisions: Material anomalies, unattributed spending, abandoned work, budget risk, and the owner and action for each finding.
Token Spend & Insights supports the spending and attribution record. Developer Productivity and Workflow Intelligence add delivery and workflow context. Session-level diagnosis and provider capacity records may be needed for deeper analysis.
A dashboard shows the evidence used to manage token capital; it is not an accounting valuation of that capability. An observed correlation between AI adoption and outcomes also needs appropriate comparison and attribution before it becomes a causal benefit claim.
Turn Token Spend Into Repeatable Capability
One successful model response is an output. A workflow that reliably achieves a defined outcome, with reusable context, evaluation, oversight, and sound economics, is a stronger candidate for the capability described here as token capital.
Preserve what made the result work: task definitions, data access, tests or review criteria, model-selection evidence, cost assumptions, and intervention rules. Revalidate the workflow when models, prices, business needs, or risk conditions change.
That learning can improve the next investment cycle, but compounding returns are not automatic. A copied workflow may not transfer to a different team or data environment. Measure the transferred result rather than assuming the original gain repeats.
What Should CFOs and CIOs Do Now?
- Establish the cost record. Reconcile subscriptions, usage, cloud charges, agents, and material supporting costs.
- Assign owners and workloads. Document shared and unknown allocations instead of forcing precision.
- Define accepted outcomes. Choose use-case-specific units and quality guardrails before calculating a ratio.
- Set comparison baselines. Use representative historical or matched work, and disclose missing evidence.
- Review model tiers and capacity. Test task fit, inspect effective prices, and confirm commitments still support the original requirement.
- Assess marginal returns. Distinguish valuable expansion from consumption growth that produces little additional accepted work.
- Preserve reusable learning. Record validated practices and test their transfer to another workflow.
Separate the strategic capability conversation from accounting classification and financial realization. Released capacity, projected benefits, and actual cash savings should remain distinct in board reporting.
Frequently Asked Questions
What is token capital?
It is the repeatable capability built through AI-powered work: useful workflows, proprietary context, evaluations, oversight, and durable results. It is a management concept here, not an automatic recognized balance-sheet asset.
How is value per token different from cost per token?
Cost per token describes a billing unit. Outcome-based measures connect consumption or spending with accepted work and quality. There is no universal ratio that makes every department's output comparable.
Is high token usage a sign of productive adoption?
Not independently. It may support valuable work or reflect retries, long context, abandoned sessions, or weak task fit. Evaluate accepted outcomes, total effort, quality, and the return on incremental spending.
Does reserved capacity always reduce cost?
No. Its value may come from throughput, latency, availability, or other requirements. Check deployment utilization, application of any pricing commitment, and alternatives at the actual workload.
Can unchanged model prices still produce a higher bill?
Yes. Tokenization, context length, output, retries, model mix, and caching can change task costs without a listed rate change. Track effective cost per accepted result alongside the pricing table.
How should we express financial ROI?
State the cost scope and the method for valuing attributable benefits. Benefits divided by cost is a benefit-cost ratio; net ROI subtracts costs before dividing by costs. If benefit valuation is uncertain, report operational outcomes and unit economics rather than unsupported financial precision.
Measure What the Investment Builds
Larridin connects AI spending with adoption, proficiency, workflows, and business outcomes so leaders can examine the work behind the bill.
Book a discovery call to discuss your token-spend and outcome measurement.
The objective is not maximum consumption or minimum token price. It is repeatable, useful capability whose economics and quality withstand scrutiny.