Capitalizing Software Just Got Easier. Your AI Costs Didn't.

A founder handed me a board deck last month with a line that read "EBITDA improved $600K quarter over quarter." Growth was flat. Headcount hadn't moved. I asked what changed operationally. Nothing had. What changed was an accounting policy, and the $600K was real money that still left the bank. It had moved from one financial statement to another.

The old gate is gone

For twenty years, capitalizing internal software development meant proving your project had cleared specific stages. That framework predated agile, and nobody could say cleanly when the "application development stage" started on a product that ships every two weeks. So in practice a lot of engineering spend got expensed by default, because the stage test was too fuzzy to defend.

FASB rewrote it. ASU 2025-06 replaces the stage gate with two conditions: leadership has committed to funding the work, and completion is probable. That's it. You can adopt it early right now, and it's mandatory for annual periods starting after December 15, 2027.

Here's what that does to the numbers. Say you spend $2M a year on engineering. Under the old test you might have capitalized $600K and expensed the rest. Under the new one, more of that work clears the bar earlier, so you capitalize $1.2M. Your P&L now shows $600K less operating expense. EBITDA jumps by exactly that. The cash out the door is identical. You've moved spend onto the balance sheet, where it amortizes over years instead of hitting this quarter's margin.

The AI part won't cooperate

Now the catch that matters if you're building an AI product. The new rule explicitly does not cover the costs of training AI models or converting data. Those stay as expense. One of the accountants who worked on it put it plainly: refining a model behaves more like fixing bugs than building an asset, so it doesn't get to sit on the balance sheet.

Run two companies building the same feature. One writes it in conventional code. The other builds it around a model it's tuning, labeling data, running eval cycles. The first can capitalize most of that engineering. The second capitalizes the plumbing and expenses the actual intelligence. Same product, same cash burned, and the AI-first team reports lower EBITDA and thinner margins because a bigger slice of its work is legally stuck on the income statement.

Capitalization policy is now a lever that moves reported profit without moving a dollar of cash. The more AI-native your build, the less of that lever you get to pull.

What to actually do

Know which line each dollar of development lands on before the quarter closes, not after. If your EBITDA improves and you can't name the operating reason, check whether capitalization did the work, because your lender and your board will eventually ask. And when you benchmark margins against a competitor, find out how much of their build is classic software versus model work. You may be losing a comparison on accounting geography, not economics.

The rule made it easier to make your P&L look better. It quietly made that harder for exactly the companies betting hardest on AI.