AI is moving from software budget to capital strategy

Intro

The first enterprise AI budget was easy to recognize: pilots, licenses, API calls, and a few ambitious slide decks. The next budget is harder to spot. It sits across cloud commitments, dedicated environments, data preparation, security controls, integration work, operating teams, and contracts that may outlive the use case that justified them. Pilots travel light. Production brings luggage.

Our last issue examined the infrastructure bill. This issue asks the question that follows naturally: who should own the investment logic? The answer is broader than Technology. Finance, Procurement, Operations, Risk, and business leaders now need a shared view of demand, commitment duration, utilization, and exit cost. The cloud may still be elastic. The contract, occasionally, has other plans.

This is an economic argument, not an accounting ruling. A rented service does not automatically become capital expenditure under GAAP. It can still create capital-like economics when the organization prepays, guarantees minimum spend, reserves scarce capacity, finances implementation, or builds workflows that are expensive to move. The chart of accounts may stay calm. Cash flow and switching costs may not.

AI has moved from a software purchasing decision into capital strategy. Leaders must govern how much capacity they commit to, how long they commit, what demand earns that commitment, and how the company gets out if the assumptions change.

Executive summary

The supplier side already looks unmistakably capital-intensive. Microsoft reported 41 billion dollars of capital expenditures in its FY2026 fourth quarter; roughly two thirds went to short-lived assets, primarily CPUs and GPUs, while the remainder supported long-lived assets. Meta expects 2026 capital expenditures of 130 billion dollars to 145 billion dollars. Amazon said a 66.1 billion dollars year-over-year increase in purchases of property and equipment, primarily for AI, helped move trailing-twelve-month free cash flow to a 7.6 billion dollars outflow. AI strategy now comes with depreciation schedules, an exciting development for a very specific audience.

Most enterprises will not own that infrastructure. They will encounter its economics through pricing, capacity reservations, minimum commitments, prepayments, dedicated deployments, implementation programs, and exit terms. AWS, Azure, and Google Cloud all offer lower prices in exchange for one-year or three-year commitments. That can be sensible for proven baseline demand. It can also turn a forecast error into a recurring invoice with excellent documentation.

The executive task is to separate experimentation from durable demand. Keep uncertain workloads flexible. Commit where adoption and unit economics are measured. Value portability where disruption would be expensive. Then review the portfolio as capital strategy, even when Accounting properly records much of it elsewhere. Flexibility is what the company buys while certainty is still earning its bulk discount.

Core areas to watch

1. 1

   The budget is becoming a portfolio of commitments

   FocusAWS Savings Plans · Microsoft Azure savings plans · Google Cloud committed use discounts

   Summary Cloud economics reward confidence. AWS Savings Plans, Azure savings plans, and Google Cloud committed use discounts exchange lower prices for one-year or three-year commitments. Apparently, 'pay as you go' has a one-year and three-year cousin. The bargain works when a predictable base load exists. It works less beautifully when an enthusiastic forecast is promoted to a contractual obligation.

   Key takeaways

   • Separate baseline demand from option demand. Commit only the stable portion; keep pilots and volatile workloads on flexible terms.
   • Give every commitment an owner, a utilization target, and a review date. Savings without accountability have a habit of becoming archaeology.
   • Price the downside before approving the discount. The relevant comparison includes unused capacity, migration cost, and the value of flexibility.

2. 2

   Commitment duration can outrun the business case

   FocusMicrosoft FY2026 Q4 earnings call · Meta Q2 2026 results

   Summary AI investment combines assets and obligations with very different clocks. Microsoft said roughly two thirds of its quarterly capital expenditures supported short-lived CPUs and GPUs, with the balance in long-lived assets, and it recorded 5.6 billion dollars of finance leases primarily for large data-center sites. Enterprise buyers face their own duration mismatch: fast-changing models and uncertain adoption can sit inside multi-year capacity, platform, and implementation commitments. Models sprint. Contracts have discovered endurance sports.

   Key takeaways

   • Map the clocks separately. Use-case evidence, contract duration, integration life, and underlying infrastructure do not mature at the same speed.
   • Match commitment length to demand confidence. A strategic use case with measured adoption can support a longer term than an executive-sponsored experiment with immaculate slides.
   • Create expansion gates. New capacity should follow utilization, quality, and business-value targets instead of calendar momentum.

   When a three-year commitment funds a six-month hypothesis, optimism has been given purchasing authority.

3. 3

   Rented infrastructure still creates capital-like economics

   FocusMicrosoft FY2026 Q4 earnings call · Amazon Q2 2026 results

   Summary Ownership is only one part of the economic picture. Microsoft explained that a change in the expected mix of finance and operating leases lowered its reported 2026 capital-expenditure expectation without changing the underlying investment program. Amazon's results showed the other side of the cycle: heavy AI infrastructure purchases materially affected free cash flow. For enterprise buyers, the lesson is simple. Classification can change while cash commitments, utilization risk, and switching costs remain stubbornly real. Calling the bill operating expense does not make it operationally optional.

   Key takeaways

   • Track total committed cost across usage, minimums, prepayments, implementation, data movement, operating staff, and exit work.
   • Show cash timing beside unit economics. A favorable per-unit rate can still create a poor portfolio result if demand arrives late.
   • Keep the accounting conclusion and the investment decision connected but distinct. The line item matters; the economic exposure matters too.

   The line item may move. The obligation, less impressed by presentation, remains.

4. 4

   Optionality is an asset that must earn its keep

   FocusOracle FY2026 Q4 results · AWS Savings Plans · Google Cloud committed use discounts

   Summary Capacity can be secured with on-demand spend, reserved commitments, prepayments, customer-supplied hardware, or owned assets. Oracle disclosed that prepaid and customer-supplied hardware tied to large AI contracts totaled 75 billion dollars at the end of FY2026. The structure reduces Oracle's capital burden while giving customers a direct stake in capacity. Contract architecture is now part of technology architecture. A multi-provider plan can buy resilience; without discipline, it can simply collect invoices.

   Key takeaways

   • Pay for flexibility where demand is uncertain or workloads change quickly. Commit where a stable base load and measurable value justify the discount.
   • Preserve practical exit routes at the business-critical layers: data, identity, workflows, observability, and regional deployment.
   • Value optionality explicitly. Multi-provider designs, portability work, and on-demand capacity cost money; so does discovering that the only available exit requires a multi-quarter program.

   Optionality is expensive. So is discovering that you did not buy any.

Boardroom implications

Executives do not need to develop a personal view on GPU depreciation over dinner - unless the dinner has gone badly. They do need one economic language for AI. These questions belong in portfolio reviews, budget cycles, and major vendor decisions:

1. Demand proof. Which use cases have measured adoption and business value, and how much demand remains experimental?
2. Economic unit. What does one completed, quality-adjusted business outcome cost after infrastructure, data, integration, oversight, and rework?
3. Commitment duration. How long are we obligated, and does the evidence support that level of confidence?
4. Reuse. Can the capacity, data work, and controls serve several workloads, models, or providers?
5. Exit cost. What happens to cost, service, and timing if price, capacity, regulation, or a provider changes?

Case study spotlight

The wrap-up

Moving AI from the software budget into capital strategy calls for disciplined reversibility. Match hard-to-reverse commitments with proven demand, keep uncertain demand flexible, and design reusable capacity around business outcomes. Giving every pilot a five-year program would rather miss the point.

Finance can leave model benchmarks to Technology; Technology can leave accounting policy to Finance. Both need the same view of commitments, utilization, duration, and exit cost. The spreadsheet and the architecture diagram should, at minimum, be on speaking terms. If they meet for the first time during renewal season, introductions are overdue.

Call to action

1. Map. Consolidate AI spend, minimum commitments, reservations, prepayments, implementation costs, and exit obligations into one portfolio view.
2. Segment. Separate proven baseline demand from experiments, seasonal demand, and strategic options.
3. Gate. Tie longer commitments to adoption, utilization, quality, and business-value targets.
4. Negotiate. Treat portability, reallocation rights, service levels, price resets, data movement, and termination terms as capital protections.

AI spend that can outlast the use case belongs in capital strategy, whatever the accounting label says.