EBITDA

EBITDA exists to answer one specific question: putting aside how a company is financed, taxed and
how it accounts for old spending, how well do the operations themselves earn money? That framing
is genuinely useful for comparing two businesses that might have very different capital structures
or tax situations but sell a similar product.

The trouble starts because EBITDA is also the number multiples get quoted against, which gives
every seller a direct incentive to make it as large as possible. Adjustments that are defensible
in isolation — removing a one-time legal settlement, a founder's above-market salary, a
non-recurring relocation cost — accumulate into a figure that can drift a long way from the cash
the business actually generates. A buyer expects some adjustment. What they check for is whether
the adjustments are one-time in fact, not just in name.

For an AI company specifically, EBITDA needs unusually careful reading because two of its
biggest costs — model inference and cloud compute — behave differently from ordinary cost of
goods. They scale with usage in ways that are easy to describe as "temporary" or "one-time"
right up until the next quarter, when they show up again at a similar size.

EBITDA starts from operating earnings and adds back four categories of charge.

Interest — the cost of debt, removed because a comparison of operating performance should not
depend on how much a company happened to borrow, or on what terms.

Taxes — removed for the same reason: tax rates and structures differ by jurisdiction and
ownership, and a comparison of operations should not be a comparison of tax planning.

Depreciation — the non-cash charge that spreads the cost of tangible assets, like servers or
equipment, over their useful life. Adding it back reflects that the cash was already spent when
the asset was bought, not when it depreciates on the books.

Amortisation — the equivalent charge for intangible assets, including acquired technology or
capitalised software development. Same logic, different asset class.

Adjusted EBITDA goes further, adding back items the seller argues will not recur: one-time
legal costs, restructuring charges, a founder's salary normalised to a market rate. This is where
the real negotiation happens, because every add-back needs its own justification, and a buyer's
diligence team will ask for the evidence behind each one rather than accept the label
"non-recurring" on its own.

EBITDA is not cash. It excludes capital expenditure entirely, so a business with high EBITDA
and heavy ongoing infrastructure spend can still be consuming cash faster than it earns it. This
is a live issue for AI companies, where compute and inference costs are a real and recurring cost
of serving customers, not a one-time investment — treating them as an add-back is one of the more
aggressive adjustments a buyer will push back on.

Every add-back needs a receipt, not just a label. "Non-recurring" is a claim, not a fact.
Ask for the specific line items behind each adjustment and whether anything similar has appeared
in a prior period — a cost that shows up every year is not one-time no matter what it is called.

Founder and related-party costs cut both ways. Normalising an owner's below-market salary up
to a market rate is standard and fair. Adjusting it down, or removing other owner-benefit
expenses entirely, needs the same scrutiny in the other direction.

A margin without the context behind it says little. EBITDA margin varies enormously by
business model, growth stage and how much a company is investing in growth versus harvesting
profit — reading it against a single external benchmark is a common way to draw the wrong
conclusion from an otherwise honest number.

A listing's Deep Dive tab presents a startup's EBITDA and margin figures to buyers and investors
alongside the underlying numbers they are built from, and the AI analysis report reads those same
figures — from what the founder uploads plus public data — when it produces its valuation range,
score and per-section confidence. The report reads the numbers as given; it does not itself audit
which add-backs are genuine. A Verified listing is where those key figures have been checked
against source documents, which is the layer that actually answers whether an adjustment holds up.
The free AI valuation tool at /valuation is a useful first check on how a reported EBITDA
translates into a range before a full listing goes live.