AI Capex vs Opex: The New CFO Math

Every dollar you spend on AI is either opex (expensed immediately, below the operating line) or capex (capitalized as an asset, depreciated over time). This choice has real consequences for your balance sheet, your cash flow statement, and how investors interpret your financial performance.

The rule is simple: if you are building an asset that will generate value over multiple years, capitalize it. If you are buying a service, expense it. But with AI, the line is blurry. Is a fine-tuned model an asset or a service? Is an inference cluster capex or opex? The CFO needs a framework.

The Case For Capex (Capitalize It)

You should capitalize AI spending if you are building something that:

1. Is owned by your company (not rented from a vendor).
2. Will generate economic benefit for more than one year.
3. Can be reliably measured and tested (you can prove it is working).

Examples:

• GPU clusters for self-hosted inference. If you buy or lease GPUs, this is capex. You are buying hardware. Depreciate over 3 years.
• A fine-tuned model (if you own it). If you spend $100k fine-tuning Llama 2 on your proprietary data, and you own the resulting model, this is capex. Capitalize it and depreciate over 3 years (or the expected useful life before the model is obsolete).
• Custom training data (if it has lasting value). If you spend $50k curating training data for your domain, and that data will be used to improve models for 3 years, you can capitalize it.
• AI infrastructure tooling (if you own it). If you build internal MLOps infrastructure, you can capitalize it under "software" and depreciate over 3-5 years.

Accounting treatment: Capitalize the asset under "Property, plant, and equipment" or "Intangible assets," then depreciate over useful life (typically 3 years for software and models, 5-7 years for hardware). Depreciation goes to COGS or operating expense, depending on the nature of the asset.

Pros of capex treatment:

• You spread the cost over multiple years, which looks better for profitability in year one.
• It signals to investors that you are making a long-term infrastructure bet.
• It matches the asset's economic benefit to the periods it generates value.

Cons of capex treatment:

• You have to defend the asset's useful life. Why 3 years and not 2? Your auditor will ask. If the model becomes obsolete in 18 months, you have to take a write-down.
• Balance sheet gets heavier (more assets). Investors might view heavy capex as lower quality earnings (more accounting judgment, less cash).
• You have to monitor the asset for impairment. If your fine-tuned model stops working, you have to write it down immediately.

The Case For Opex (Expense It)

You should expense AI spending if you are:

1. Renting a service (paying a vendor per token, per seat, per query).
2. Buying ongoing support or maintenance that does not create a lasting asset.
3. Conducting research or experimentation where outcomes are uncertain.

Examples:

• OpenAI API spend. You pay by token. This is opex. Expense when incurred.
• SaaS AI tools (Intercom Fin, Decagon, Sierra). Monthly fees. This is opex. Expense when billed.
• Fine-tuning and training that you do not own. If you hire a vendor to fine-tune a model and the vendor retains the model, that is opex (you are buying a service, not an asset).
• Experimentation and R&D. If you are experimenting with AI to decide whether to invest in a capability, expense it. Once you are confident and rolling it out, you can capitalize future builds.
• Ongoing model maintenance and retraining. If you need to retrain a model quarterly to keep it accurate, those costs are opex.

Accounting treatment: Expense in the period incurred. Depending on the nature, it might go to COGS (if it is customer-facing), R&D (if it is research), or operating expense (if it is general support).

Pros of opex treatment:

• Simpler accounting. You do not have to defend useful life or watch for impairment.
• Looks like "real" business economics to investors (cash in, benefit out, in the same period).
• More conservative. You are not capitalizing uncertain future benefits.

Cons of opex treatment:

• Year-one profitability looks worse. Large AI spending depresses operating margin.
• Does not match economics if the benefit spans multiple years. You incur the cost in year one but the benefit accrues in years one and two.

The AI Workforce P&L Lens

Think of AI the way you think of headcount. A new employee is a capex/opex hybrid:

• The hiring cost and training are sunk (not capitalized).
• The salary and benefits are opex.
• The productive output over 3 years is the "benefit."

The AI Workforce P&L reframes the question: treat AI agents like employees. You do not capitalize the "training" of a new hire; you expense the salary. Similarly:

• Agent acquisition cost (building and fine-tuning): opex (like hiring and onboarding).
• Agent run cost (tokens, infrastructure, compute): opex (like salary).
• Agent depreciation (the agent gets obsolete as models improve): take it as a write-down when it happens, or spread it over the asset's life if capitalized.

This framing makes the capex/opex choice easier: if the AI spending is helping you operate a business function (like payroll), expense it. You do not capitalize payroll.

The Mixed Case: Build-and-Buy

Most companies do both. You might:

• Use OpenAI API for initial deployment (opex).
• After 6 months, fine-tune a custom model because the API is not accurate enough (capitalize the fine-tuning; expense the API).
• After 18 months, self-host because scale makes it cheaper (capitalize the infrastructure; depreciate the API depreciation over time).

Each transition is a CapEx vs. OpEx decision. Your finance team needs to document the rationale:

1. What are we building or buying?
2. Do we own it, or are we renting it?
3. Will it generate value for more than one year?
4. What is the useful life? (If it is an AI model, typically 2-3 years because models improve and become obsolete.)
5. Is there a risk of impairment? (If the model stops working, can we write it down?)

The Tax Angle

One more consideration: tax. Under current U.S. tax law (as of 2026):

• Capex generates depreciation deductions over time, reducing taxable income in multiple years.
• Opex is expensed immediately, generating a deduction in year one.

For high-tax-rate companies, opex is tax-advantaged (you get the deduction sooner). For low-tax-rate companies, the timing does not matter much.

Your tax team should weigh in on whether to capitalize or expense specific AI spending. The CFO and the tax team should align.

What To Do Next

Document your AI spending by category: API (opex), fine-tuning (capex or opex), infrastructure (capex), tooling (capex or opex), labor (opex). For each capitalized item, document the useful life and your impairment testing plan.

Talk to your external auditors before capitalizing anything. They need to bless your useful-life assumption. If you capitalize a model at 3 years and your auditor thinks it is 2 years, you have a problem.

And remember: capex is not inherently better or worse than opex. Investors care about economics, not accounting. Show them the real economics (cash in, benefit out) and justify your accounting treatment. That is what they will audit.