Making AI an Asset, Not an Expense

Token prices are the wrong first question

Wrong first question: On 29 September 2026, MIT Technology Review Insights published Making AI an asset, not an expense. The page states the piece was produced by HPE and was not written by MIT Technology Review editorial staff. Treat it as a sponsored operating brief, not as an editorial finding. Even so, the opening diagnosis is the one boards already hear in vendor rooms: cost talk starts with token prices and ends with access to the newest cloud model. The sponsored brief asks whether that level of capability is always required. It is not. The useful move is to keep the question and drop the infrastructure pitch that follows it.

Consumption is a starting regime: The same brief argues that as AI moves from isolated pilots into production portfolios (assistants, retrieval systems, and agentic workflows), a consumption-only approach can turn spend into a variable monthly line that is hard to forecast. That is a real finance problem. It is not an argument to buy dedicated capacity by default. Ownership only makes sense when demand is steady, predictable, and large enough to keep capacity productive. The brief itself says there is no universal crossover number. It depends on the workload, the mix of retrieval and generation, energy, and the operating model. Generic cost benchmarks are marketing. Your actual volume is the model.

Activity is still not P&L: MIT Project NANDA, in The GenAI Divide: State of AI in Business 2025, reported that 95% of custom or integrated GenAI pilots had no measurable P&L impact. McKinsey / QuantumBlack, in The State of AI: Global Survey 2025, found that nearly two-thirds of respondents said their organisations had not yet begun scaling AI across the enterprise. Those two findings sit next to the sponsored cost brief on purpose. Buying a different economic model does not invent a P&L result. If the workload is still an experiment, you are not capitalizing an asset. You are pre-paying for unused capacity.

Leadership still owns the economic model: In The Decision Maker's Handbook to Data Science, Dr. Stylianos Kampakis argues that executives own the standard of evidence, not the vendor. The same rule applies to AI cost. Token price lists and model-brand bake-offs are procurement theatre until someone can name the workload, the expected demand, and the owner who will review utilization after go-live. Capital follows a control system. It does not follow a price card.

Five rows before you capitalize AI

Write these five operational criteria on one page before you leave consumption pricing, buy dedicated capacity, or renew a frontier-model contract for a production workflow. If any row is blank, you are not making a capital decision. You are making a hope.

• Demand: what 12 to 18 month volume, by workload family, justifies leaving pay-per-request pricing?
• Fit: which job is this, and which class of capability does that job need? Name why frontier access is required, or why it is not.
• Crossover: at what utilization does ownership beat consumption for this workload? There is no universal number. Show the assumptions.
• Utilization owner: which named executive will review idle capacity after go-live and stop or reallocate it?
• Stake: which operating roles gain if the tool works, and what do they lose if it replaces their current routine?

Unused capacity is a people problem

Fear of loss is not resistance: On 29 September 2026, McKinsey Talks Talent published Leading when change feels threatening, with McKinsey partner Johanne Lavoie. The episode states that leaders love to talk about resistance to change, and that McKinsey research shows resistance is not the full story. Fear of loss is a large part of transformation: roles, routines, and relationships. That framing matters for AI capacity. If people believe a new workflow takes their judgement, their status, or the work they know how to defend, they will not bring volume onto the platform. Idle GPUs and unused seats then look like a technology problem. They are an ownership and loss problem.

Give them a stake, not a seat: On 28 September 2026, Harvard Business Review published a short with Harvard Business School associate professor Zoe Cullen: How Can CEOs Motivate Employees to Participate in AI Innovation? The claim is operational, not motivational: if you want employees to help build and improve AI tools, give them a reason to have a stake in their success. Seat rollouts without a stake create traffic and then silence. A utilization owner who cannot name who gains when the tool works is not an owner. They are a reporter of unused licences.

Consumer agent hype is not an enterprise asset: Stratechery on 29 September 2026 argued, in a note on agents and Meta Connect, that Meta has a chance to own the consumer agentic space and that going for enterprise is a mistake. You do not need to adopt that product thesis to use the executive implication. Consumer agent demos are not a reason to pre-pay enterprise capacity. If the workload is not already repeating inside your own systems, you are buying a story about agents, not an asset.

Value is captured when it can be governed: In Business Models in Emerging Technologies, Kampakis with Mourouzis, Cardoso, and Zinopoulou argue that value is captured when a capability can be governed and repeated. In Predicting the Unknown, Kampakis argues that uncertainty stays after the interface becomes fluent. Those two claims close the utilization test. A fluent chat window does not make unused capacity productive. A named owner, a stake for the people who run the workflow, and a kill or reallocation rule do.

Build capability: access free executive courses and clinics

Structured practice over price cards: Sustainable commercial capability cannot be purchased through a token contract or a capacity invoice alone. It requires structured, hands-on practice on demand, fit, and ownership. Rather than committing capital to unverified stacks, leadership teams should first build baseline internal capability on unit economics, verification, and utilization. Tesseract Academy provides open access to foundational executive courses and practical clinics designed for decision-makers who must govern AI initiatives and protect company balance sheets.

Free diagnostic tools: The Free AI Fluency Course gives department heads and controllers frameworks to evaluate model proposals, separate experiments from production workloads, and identify unused capacity before the next capital request. Learners can work those tests inside the AI Coach in the free course, then bring a one-page pack to finance.

More free executive courses: For organisations seeking broader data literacy and strategic governance, the free courses hub covers data modernisation, hiring standards, and enterprise AI roadmaps. Starting with free diagnostic practice lets department heads, commercial directors, and financial controllers share a vocabulary and a standard of evidence before anyone treats AI as a line on the balance sheet.

What to bring to the next capital meeting

One page is enough: Before you leave consumption pricing or expand a frontier-model contract, require sponsors to complete the five rows for every production workload. Keep unanswered questions visible. Blank cells are information. Hidden blanks are risk.

If the utilization owner is missing, stop: If a sponsor cannot name the demand, the fit, the crossover assumptions, the executive who will review idle capacity, and the operators who have a stake if the tool works, the organisation is not buying an asset. It is pre-paying for unused traffic that finance will pay for twice: once as consumption or capacity, again as cleanup. Capital follows evidence, not slogans.

Sources cited

• Making AI an asset, not an expense. MIT Technology Review Insights (produced by HPE). Sponsored Insights piece, 29 Sep 2026. Not MIT TR editorial staff. Opens on token prices vs needed capability; asks when consumption pricing stops fitting.
• The GenAI Divide: State of AI in Business 2025. MIT Project NANDA. 95% of custom or integrated GenAI pilots had no measurable P&L impact.
• The State of AI: Global Survey 2025. McKinsey / QuantumBlack. Nearly two-thirds of respondents said their organisations had not yet begun scaling AI across the enterprise.
• Leading when change feels threatening. McKinsey Talks Talent. Johanne Lavoie, 29 Sep 2026. Resistance is not the full story; fear of loss of roles, routines, and relationships.
• How Can CEOs Motivate Employees to Participate in AI Innovation?. Harvard Business Review. Zoe Cullen, 28 Sep 2026. If you want employees to help build AI tools, give them a stake in their success.
• One More Note on Agents, Meta Connect, Meta Enterprise Platform. Stratechery. 29 Sep 2026 note: Meta has a chance to own consumer agents; going for enterprise is a mistake.
• The Decision Maker's Handbook to Data Science. Apress / Springer Nature, 2020 and 2024 editions, sole author. Executives own the standard of evidence and the economic model.
• Business Models in Emerging Technologies. Business Expert Press, 2022, lead author with Mourouzis, Cardoso, and Zinopoulou. Value is captured when a capability can be governed and repeated.
• Predicting the Unknown. Apress (Springer Nature), 2023, sole author. Uncertainty stays after the interface becomes fluent.