It is usually an expense, and finance settled why long before the market started arguing about it. Under U.S. GAAP ASC 350-40, the internal-use software standard updated in June 2025, money spent to run or maintain an existing tool is expensed as it is incurred, with no capitalization. That is the accounting shape of rent: a period cost, not owned value on the books. The build side is treated differently. Spend on building a new internal system can be capitalized, meaning recorded as a long-term asset on the balance sheet and amortized over its useful life rather than expensed. So the line already exists between the two kinds of AI dollar. The tool you run is the rent. The configuration and accumulated context you build is the asset. Almost nobody sorts their AI spend into those two buckets, and the sorting is the whole question.

For most subscriptions, just rent. You pay for access, and the day you stop paying, or the day the tool changes, nothing is left on the books as something you own. The value lived in reach, not in accumulation. An investment is different in one specific way: it leaves behind something the business keeps. If your AI setup would vanish the moment the subscription lapsed, it was rent no matter how useful it felt while it ran. The precise dollar tally of what all that renting costs across discovery, switching, and subscriptions is a separate accounting, laid out in the full AI Tax. The category question underneath the tally is the one that decides everything: is any of the spend building something you would still own.

Because it was rent, and the rent came due on a date you did not set. When OpenAI wound down the ChatGPT plugins beta, users could no longer start new plugin conversations from March 19, 2024, and every existing plugin conversation stopped working on April 9, 2024, with a different feature offered as the replacement. Every workflow a founder had wired up on plugins stopped on that fixed date and had to be rebuilt somewhere else. Nothing you built into that tool carried out of it, because the setup lived inside the tool, not in a layer you held. This is the same evaporation traced date by date in the reset tax: work you paid for once, gone when the product changed.

A feedback loop. Spend compounds when the context it produces feeds back into better work, so each use makes the next output sharper. AI Ireland describes this as the data flywheel: more customers generate richer data, which creates better AI experiences, which attracts more customers, so every interaction compounds the advantage while a competitor starts from zero. A bare subscription has no flywheel, because nothing accrues to you.

The honest limit matters here. Accumulation on its own is not enough. As one analysis of AI moats puts it, data becomes a moat only when it feeds back into a product that gets better, and data that just sits in storage is a cost, not a moat. So the asset is not context that merely piles up. It is context that actively improves the work, and the difference decides whether the spend appreciates or drains.

Run one test on every AI dollar. When the tool changes, does what you built survive, or does it reset to zero. If it resets, it was rent, whatever the invoice called it. If it holds and keeps improving the work, it was an asset. The durable thing is never the model itself, because models are becoming interchangeable. It is the configuration and the accumulated context, the customers, the voice, the history, the corrections, that a competitor cannot buy or borrow. That test cuts cleaner than any feature comparison, and it is the one worth applying before the next tool tempts you into rebuilding from scratch. The same instinct extends to the stack itself, where fewer moving parts hold value better than more, covered in why the stack simplifies over time.

This is where the fix gets concrete. The investment holds only when the durable layer is held apart from the swappable model, so the context accrues to the business rather than to a vendor’s model. Works is built on that separation. The business it learns lives in Notebooks and Work Areas that accumulate the customers, the voice, the pipeline, and the history as a side effect of running the work, and that context survives model and engine upgrades. When a better model ships underneath, the business the system already learned stays learned, so a six-month-old workspace runs on today’s version with no re-setup. The investment compounds instead of resetting. The mechanism of that accumulating memory is its own subject, covered in memory that compounds, and the whole idea sits inside the larger case for compounding AI.

If you want to see that separation running before the public launch, invest where it compounds and sign up for early access. If you would rather read the longer argument first, the complete case for treating AI as an appreciating asset lays it out end to end.

You are not behind on AI. You may just be renting when you meant to invest, and that is a question you can now answer on purpose.

Does AI depreciate like other software, or can it appreciate?

Both, depending on which part you mean. The tool depreciates and is interchangeable, and spend to keep it running is expensed. But qualifying internal-use software cost can be capitalized once past preliminary planning, which moves it to the balance sheet as an amortized durable asset. The configuration and context you build can appreciate. The subscription underneath it will not.

Is my everyday AI subscription treated as an asset on the books?

No. A company-wide AI subscription is booked as general overhead, expensed as incurred, the same accounting treatment as any recurring software subscription. It buys access for the period, not owned value. Only spend that builds a durable, capitalizable system lands on the balance sheet as something the business keeps.

What is the durable asset, the tool or the context I built into it?

The context. Foundation models are commoditizing, so the model itself is the interchangeable part. The durable, ownable asset is the accumulated business context that feeds back into better work and that a competitor cannot recreate. The tool is rent. The context you build around it is the build, and the build is what holds value.

Why does my AI feel expensive without holding any value?

Because the spend has no place to accumulate. When every tool starts your setup from zero, each month of work does not stack on the last, so the cost recurs while nothing compounds. Value holds only when a durable layer keeps the context across tools and model upgrades, so the investment builds rather than resets.

1. Slash, How to Classify AI Token Spending in Accounting, 2026
2. OpenAI Developer Community and Help Center, Winding down the ChatGPT plugins beta, 2024
3. AI Ireland, The New Moat, 2026
4. Valtorian, AI Moats in 2026, 2026
5. Ridgeway Financial Services, Accounting for AI Development Costs, 2026