What Is Adjusted EBITDA? Read Anthropic's Profit Like a Buyer

Adjusted EBITDA is a company's profit before interest, taxes, depreciation and amortisation — and then before whatever further costs management chooses to set aside as one-off, non-cash or outside the ordinary course of business. No rulebook fixes that second list. The company writes it. Which means the measure is only ever as honest as the costs that were taken out of it.

That definition is being tested in public right now. The Financial Times reported in September 2026 that Anthropic had told a small group of shareholders it expected a second consecutive quarter of positive "adjusted operating income"¹ — a close cousin of adjusted EBITDA — with the measure stripping out costs including stock-based pay.¹ Other outlets described the excluded items as one-time or exceptional. Reporting also put gross margins above 80% before revenue shared with distribution partners such as Amazon and before the cost of training the models,² with quarterly revenue past $11.5 billion¹ and annualised revenue of $65 billion in July.³ The company submitted a draft registration statement confidentially,⁴ so no public filing exists: no audited results, no profit figure under ordinary accounting rules, no cash-flow statement. Every number above belongs to the press, not a prospectus. That makes this a teaching case and nothing else — there is no view here on Anthropic's shares or any future listing.

The habit is what transfers. Owners meet adjusted numbers in every private-fund deck, every portfolio-company board pack, and — with the roles reversed — in the "add-backs" a buyer's accountants propose when a founder sells. These are the questions that separate a reasonable adjustment from a convenient one.

Can we see the bridge from reported profit to this number, line by line?

A good answer is a table, not a label: profit or loss under ordinary accounting rules on the top line, each adjustment on its own line with an amount and a one-sentence reason, the adjusted figure at the bottom. Where audited statements do not yet exist — as at a private company — a good answer says so plainly and bridges from whatever unaudited accounts do exist.

A weak answer leads with the adjusted figure alone, or waves at "certain one-time and non-cash items" without listing them or pricing them. "One-time" is a claim; the list is the evidence.

The adjusted number is only as trustworthy as the worst line in that bridge. Without it, a defensible exclusion and an aggressive one look identical on the page.

Which of these excluded costs will be paid in cash again next year?

A good answer sorts the exclusions into three buckets and says which is which: genuinely non-cash items, genuinely singular cash events (a site closed, a lawsuit settled), and recurring cash costs that have been relabelled as extraordinary. The first two are ordinary practice. The third is the problem.

A weak answer blurs the buckets. The reference case is WeWork, which for 2017 reported a net loss of $933.5 million⁵ while presenting a measure it called "community adjusted EBITDA" of positive $233.1 million⁶ — positive only after marketing, pre-opening costs and general overhead had been removed.⁷ Those are the ordinary cash costs of running the business.

This is where the long-term stake sits. What compounds across decades is free cash flow — the cash left after a business has paid for everything it needs to keep running — and earnings already diverge from cash for perfectly innocent accounting reasons. An adjusted measure that removes recurring cash costs widens that gap on purpose, and describes a business that does not exist.

If the excluded spending stopped, would the business still be there in three years?

A good answer names the spending that has to be repeated simply to stay in the market, and refuses to park it outside the operating business. Where a cost is genuinely discretionary — a campaign that can be switched off without the revenue following it — the answer explains why, and shows the line separately so a reader can put it back.

A weak answer reclassifies the cost of being in business as investment in some calmer future version of the company. Groupon's 2011 registration statement presented "adjusted consolidated segment operating income" of positive $60.6 million for 2010 against a GAAP operating loss of $420.3 million; online marketing was excluded from the non-GAAP profit⁸ — the cost of acquiring the subscribers who were the business. SEC staff called that exclusion potentially misleading, and the measure was removed from the filing.⁹

The principle is now written down: the SEC's Compliance and Disclosure Interpretation 100.01 states that adjustments can make a non-GAAP measure misleading, and its own example is the exclusion of normal, recurring cash operating expenses.¹⁰ That is the sharp and fair question hanging over any AI lab's adjusted profit. Excluding stock-based pay is conventional. Training the next model is a different kind of exclusion — for a frontier lab, training is not a project with an end date, it is how the product gets renewed. Owners need not be regulators to apply the test: if the spending cannot stop without the business stopping, it is a cost.

When people are paid in shares, what does that cost the owners who are already here?

A good answer grants the ordinary practice and then prices it: how many new shares the pay creates, what percentage of the company existing holders give up, and what cash outlay would be required to retain the same people without equity. Stock-based compensation is excluded across listed technology companies, and flagging it is fair — provided the dilution is disclosed alongside.

A weak answer stops at "it isn't cash, so it isn't a cost." That treats a recurring claim on the business as though nobody owned it.

Talent is never free. Paying engineers in shares rather than cash makes the adjusted number look better while shrinking every existing owner's slice of the free cash flow the business will eventually produce. Removing a non-cash line does not create cash that can compound.

What did the same period look like in cash?

A good answer sets three figures side by side for the identical period: the adjusted measure, profit under ordinary accounting rules, and cash from operations less the spending required to keep the business competitive. Where free cash flow is negative because the business is building something, a good answer says so in plain terms and explains what the money is buying — without promising what it will return.

A weak answer argues that cash flow is "not the right way to look at a growth business" and leaves it there. That may be true for a period. A manager who cannot say when it stops being true has not finished the thought.

Adjusted EBITDA is a lens on operations. Free cash flow is what can actually be distributed, reinvested or used to buy back shares, and in a healthy business the two converge over time. A persistent gap between them is itself information — the subject of the three-question test on revenue and cash.

If this number sets the price, which add-backs would we reject from the buyer's chair?

A good answer prices the business both ways — the multiple on the reported figure and the multiple on the adjusted one — and names the add-backs a disciplined buyer will refuse to pay for. A market-rate salary in place of an owner's outsized one can be legitimate. A personal expense that leaves with the seller can be legitimate. Marketing that produces the customers, maintenance the buyer must keep funding, and a "one-time" restructuring that recurs are not.

A weak position is a single headline number, called normalised, with the add-backs summarised as a percentage rather than listed. The percentage hides the argument; the names are the argument. It is also worth asking, without accusation, who is paid more when the adjusted figure is higher — the adviser, the banker, anyone whose fee moves with a mark. That does not make an adjustment wrong. It tells the room whose interest it serves.

The same logic runs in reverse at exit. A quality-of-earnings review — the buyer's examination of which profit genuinely recurs — proposes add-backs and, just as often, take-backs: under-market rent from a related party, deferred maintenance, a key hire not yet on the payroll. Sellers who arrive with a documented bridge already built keep control of the conversation. And because earnouts are frequently written against adjusted metrics, the definition keeps working long after closing.

How to use the list

This works as a filter, not a speech. Three failures are enough to slow a meeting down: no line-by-line bridge, a recurring cash cost labelled one-time, and no cash figure set beside the adjusted one. Every adjustment is a claim that a cost will not return, and it earns belief only when the cost genuinely sits outside the business. The company chooses its definition; owners choose whether to underwrite it.

Disclosure: The Stark Fund has no position in Anthropic.