Extend your runway.

// So what

Mix up COGS and OpEx and it cuts both ways: leave your AI inference out of COGS and your margins look invisible to investors. Dump every research experiment into COGS instead, and a fast-moving software company starts looking like a money-losing services shop.

// The reality

AI infrastructure doesn’t behave like a flat monthly SaaS bill. Two things push your cost to serve each customer down over time:

Your true cost to serve a customer should trend down over time, rather than stay flat.

// The blunder of generic accounting

A generic bookkeeper looks at your AWS, Anthropic, or OpenAI invoice and dumps the whole thing into software subscriptions.

That single call breaks your numbers in two directions. Leave inference that serves real users out of COGS, and investors can’t see whether serving each customer actually makes money. Dump every training experiment into COGS instead, and your margins can crater from a healthy 80% to a services-shop-looking 35%.

// So what

Prepay for a year of reserved compute to lock in a discount, or spend six figures of engineering time building your core platform, and a generic bookkeeper books the whole thing the day it happens. Your burn looks like it just spiked, even though you bought something that pays off for months or years.

// The reality

Applied AI companies rarely buy physical servers. The two real spends are a reserved compute commitment (a year of capacity paid upfront to lock in a discount) and the engineering payroll that builds your core platform, orchestration layer, or fine-tuning pipeline. Both should have their cost spread across the period they actually deliver value, rather than dumped into the month the invoice cleared. That’s what capitalizing means: record it as an asset on your balance sheet, then recognize its cost gradually as you use it.

// The blunder of generic accounting

Generic bookkeepers take the easy path: cash leaves the bank, so they expense the whole amount that same month. That single call distorts your burn rate and profit and loss, making an efficient, well-capitalized company look like it’s bleeding cash.

It also costs you real money in Canada. Engineering payroll spent building qualifying technology can earn cash back through SR&ED, but only if your books separate that capital R&D investment from ordinary operating expenses.

// So what

The numbers that prove your traction to your board aren’t the numbers that get you a cash refund from the government. Run them as one set of books, and you can lose your valuation story, your tax credits, or both.

// The reality

GAAP and tax law are built for two different audiences, and they’re supposed to disagree:

Run a business across the US and Canada and you need both sets, reconciled against the same underlying numbers, rather than one document trying to do both jobs.

// The blunder of generic accounting

Traditional firms keep one basic set of books and file your taxes straight off it. Since they aren’t tracking your actual R&D work as it happens, by year-end nobody can reconstruct what actually qualifies.

That shows up two ways. In Canada, engineering payroll gets buried in generic wages, so you either miss your SR&ED refund entirely or hand a contingency firm a quarter to a third of it to reconstruct your year after the fact.

Across the border, most Canadian AI startups that raise US venture flip to a Delaware parent with a Canadian operating subsidiary. US tax law then makes that subsidiary’s R&D costs deductible over fifteen years instead of right away, which can create a surprise tax bill even while you’re burning cash.

Treat your venture books and your tax books as one and the same, and you either leave non-dilutive capital behind in Ottawa or create a surprise tax liability in Delaware.