AI Profitability Scorecard: Which AI-Native Companies Earn an Operating Profit

Of 7 AI-native companies with financials you can read, 1 reports a GAAP operating profit: Palantir, at a 47.1% margin in Q2 2026. The other 6 lose money before interest and tax. AI profitability is real, but it is concentrated in one company, and 3 of the loss-makers still publish a positive adjusted figure that removes the cost that matters most.

The short answer: among AI-native companies with public numbers, only Palantir reports a GAAP operating profit, at 47.1% of Q2 2026 revenue. CoreWeave, Nebius and Tempus AI report positive adjusted EBITDA but GAAP operating losses. SoundHound, C3.ai and Anthropic's reported draft prospectus show operating losses on every basis that matters.

What counts as AI profitability in this scorecard

Revenue leaderboards for AI are everywhere. Profit leaderboards are rare, because most of the numbers needed for one are either private or adjusted. This scorecard uses a single line that every US-listed company has to report and cannot redefine: GAAP income from operations.

Operating income is revenue minus the cost of running the business. That includes compute, salaries, stock compensation and depreciation. It excludes interest, tax and gains or losses on investments, which are financing and portfolio outcomes rather than operating ones.

The sample is 7 companies whose primary business is selling AI infrastructure or AI software, and whose figures are available in a filing or a widely reported draft. Each is measured on its most recent quarter, except Anthropic, where only full-year 2025 figures have been reported. The revenue side of this question is covered in the AI revenue leaderboard ranked by revenue quality, and this post does not repeat it.

Why net income and adjusted EBITDA are both excluded

Net income is too noisy for this purpose. In this sample alone it is distorted by a roughly $34 billion convertible remeasurement at Anthropic, $98.5 million of securities gains at Tempus and $640 million of interest at CoreWeave. None of those tells you whether the product makes money.

Adjusted EBITDA has the opposite problem. Each company defines it, and for an infrastructure business it removes depreciation, which is the main cost of the asset being sold. My position is that adjusted EBITDA is a fair measure of cash generation before reinvestment, and a poor measure of whether a GPU cloud is a profitable business. Both are reported below so you can see the gap.

The AI operating margin scorecard, company by company

Here is the whole sample on one line each. Stock compensation and depreciation are expressed as a share of revenue, because those are the 2 costs that adjusted figures most often remove.

Sources: company earnings releases and 8-K exhibits listed in the references; Anthropic from Reuters reporting of a confidential draft, via Capital Brief. Margins and ratios are Zan Digital's arithmetic on reported figures. * Includes related employer payroll taxes, because the release does not split them out. n/a means the figure was not in the documents read for this post. Periods differ, which is the scorecard's main comparability limit.

Palantir: the only GAAP operating profit, and what stock pay takes from it

According to its 8-K earnings release dated 3 August 2026, Palantir reported Q2 2026 revenue of $1.935 billion. That is up from $1.004 billion a year earlier. GAAP income from operations was $912.0 million, against $269.3 million in Q2 2025. The margin moved from 26.8% to 47.1% in 12 months.

That is not a modest result. A software company at that scale with a 47% GAAP operating margin would be exceptional in any cycle, and it is the only clean operating profit in this sample.

47% after a 14% stock bill

Palantir's adjusted operating margin was 62%. The difference is almost entirely stock. The release adds back $265.2 million of stock-based compensation and $17.3 million of related payroll taxes, which together take operating income to $1.194 billion.

Stock compensation rose 66% year on year, but revenue rose faster, so it fell from 15.9% of revenue to 13.7%. That is the direction you want. Depreciation is negligible at $8.2 million for the quarter, because Palantir sells software and buys very little hardware.

Cash and accounting agree here. Cash from operations was $1.216 billion and adjusted free cash flow was $1.220 billion, a 63% margin. When the GAAP margin, the adjusted margin and the cash margin all land within 16 points of each other, the profit is real. I would treat Palantir as the benchmark for what an AI software profit looks like, while noting that it took 2 decades to get there.

CoreWeave and Nebius: profitable before depreciation, loss-making after

The GPU clouds, often called neoclouds, sell rented compute. Their main cost is the hardware itself, which shows up as depreciation rather than as cost of revenue paid in cash. That single fact explains why the same quarter can carry a 59% margin and an operating loss.

CoreWeave's Q2 2026 release of 11 August 2026 reported revenue of $2.575 billion, more than double the $1.212 billion of a year earlier. Adjusted EBITDA was $1.510 billion, a 59% margin. GAAP operating income was a loss of $49 million, against a $19 million profit a year earlier.

Depreciation consumes over half of CoreWeave's revenue

Depreciation and amortisation was $1.393 billion in the quarter, 54.1% of revenue. Stock compensation was a further $165 million, or 6.4%. Below the operating line, net interest expense was $640 million, up from $267 million, taking the net loss to $626 million.

So CoreWeave's quarter reads in 3 layers. Before depreciation, the business throws off a large margin. After depreciation, it roughly breaks even. After the cost of the debt that bought the hardware, it loses about 24 cents per dollar of revenue. The debt layer is examined in detail in the breakdown of CoreWeave's debt stack and stress points.

Nebius shows the same shape at an earlier stage. Its Q2 2026 release, published through Business Wire on 12 August 2026, reported revenue of $582.3 million, up from $105.1 million. Adjusted EBITDA swung to a positive $236.2 million. The income statement shows a loss from operations of $175.9 million, after $259.7 million of depreciation and $102.5 million of share-based compensation.

Nebius's stock compensation is the bigger surprise. At 17.6% of revenue it is proportionally almost 3 times CoreWeave's, and it rose from $14.7 million a year earlier. Strip it out and the operating loss would be about $73 million, a margin of minus 12.6%.

Useful life is the quiet lever

Depreciation is an estimate. A company decides how many years a GPU server will earn revenue and spreads its cost across that period. A longer life means a smaller annual charge and a better operating margin, with no change in the business.

CoreWeave depreciates its computing equipment over 6 years, according to trackers that cite its 2025 10-K. Nebius extended the life of its servers from 4 to 5 years from January 2026, according to secondary coverage of its 2025 20-F. I could not reach the notes containing either policy in the filings themselves, so treat both as reported rather than verified.

The direction matters more than the exact figure. If either company used a 4-year life on its current fleet, its operating loss would be materially larger. The hyperscaler version of this argument, and the case on each side, is set out in the analysis of GPU depreciation schedules and AI earnings. I disagree with the view that a 6-year life is indefensible. I do think any neocloud margin quoted without its useful-life assumption is incomplete.

AI software margins: Tempus, SoundHound and C3.ai

The 3 AI software companies in the sample have little depreciation to argue about. Their swing cost is people, and particularly the people paid in stock.

Tempus: a net profit made of investment gains

Tempus AI, which sells AI-driven diagnostics and clinical data, reported Q2 2026 revenue of $382.5 million, up 22%, in a release dated 30 July 2026. It also reported GAAP net income of $5.6 million, against a $42.8 million loss a year earlier. That headline is the one most likely to be repeated as "Tempus turned profitable."

It did not, on an operating basis. The loss from operations was $75.9 million, wider than the $61.8 million a year earlier. Net income turned positive because the quarter included $98.5 million of unrealised gains on marketable securities. Adjusted EBITDA was a positive $8.0 million, after excluding $55.6 million of stock compensation and related payroll taxes.

Tempus is improving on its own adjusted measure. It is not yet profitable on the line this scorecard uses, and its reported net profit is a market outcome rather than an operating one.

C3.ai: cash positive, GAAP deeply negative

C3.ai's quarter to 31 July 2026 is the sharpest example of cash and accounting profit pointing in opposite directions. Revenue fell to $52.4 million from $70.3 million a year earlier, per its 8-K exhibit of 2 September 2026. GAAP loss from operations was $98.3 million, a margin of minus 187.6%.

Stock-based compensation was $59.2 million, more than the company's entire revenue for the quarter. Remove it and the non-GAAP operating loss narrows to $36.2 million. Free cash flow was a positive $2.1 million, because stock pay costs no cash and working capital moved in the company's favour.

A positive free cash flow alongside a minus 188% GAAP margin is not a profitable business. It is a business paying a large share of its workforce in equity, which dilutes shareholders instead of draining the bank account. The cost is real, it just lands on the share count.

SoundHound AI, the voice AI company, completes the group. Its Q2 2026 8-K exhibit reported revenue of $61.9 million, up 45%, and a GAAP operating loss of $43.3 million, much narrower than the $78.1 million of a year earlier. Stock compensation and related payroll taxes were $21.5 million, or 34.7% of revenue. The adjusted EBITDA loss was $9.6 million, so SoundHound loses money on its own preferred measure too.

Anthropic's draft S-1: a lab-scale operating loss

Anthropic is the only frontier lab with an income statement anywhere near the public record. Its draft prospectus was submitted confidentially in June 2026 and has not been published. The figures below come from Reuters reporting of that draft, summarised by Capital Brief on 29 September 2026.

For 2025, Anthropic reported revenue of $4.59 billion and an operating loss of $8.06 billion, a margin of minus 175.6%. Compute and infrastructure alone cost $7.33 billion. The roughly $34 billion non-cash charge that dominates the net loss sits below the operating line, which is why this scorecard does not use it. That reconciliation is worked through line by line in the analysis of Anthropic's $42 billion S-1 loss.

The more interesting claim concerns 2026. Fortune, relaying the Financial Times, reported that Anthropic expected Q2 2026 revenue of $11.5 billion and an operating profit for a second straight quarter. If the public S-1 confirms that on a GAAP basis, the scorecard gains a second profitable name. I would not count it until the filing shows the quarterly table, and until it states whether stock compensation is inside the figure.

Cash profit versus accounting profit in the AI economy

The AI economy now produces 4 different answers when you ask whether a company is profitable. They come from GAAP operating income, a company-defined adjusted measure, net income, and cash flow. In this sample they agree only once.

Same sources and periods as the scorecard table. Nebius does not print an adjusted EBITDA margin; 40.6% is Zan Digital's arithmetic. Capex multiples are capital spending in the quarter divided by revenue in the quarter.

Notice where the cash goes at the 2 neoclouds. CoreWeave spent $6.42 billion on property and equipment in the quarter. Nebius spent $5.66 billion, nearly 10 times its quarterly revenue. Whatever their adjusted margins say, neither is self-funding, and both depend on capital markets to keep building. The same pattern at hyperscaler scale is traced in the analysis of AI capex and its missing payback math.

Which lens should you use? It depends on the question. If you are asking whether a vendor will survive the next 2 years, cash and funding access matter most. If you are asking whether the business model works at scale, GAAP operating income is the honest line, because it charges for every input. The piece on the AI unit economics margin trap explains why that distinction bites hardest for companies whose main cost is compute.

Where this scorecard is weakest

A scorecard built on 7 companies and 1 quarter carries obvious limits. Here they are, stated plainly.

One quarter is a snapshot. CoreWeave earned an operating profit a year ago and lost money this quarter. Tempus guides to about $65 million of adjusted EBITDA for 2026. Every ranking above could change at the next earnings date, and the periods do not even align: C3.ai's quarter ends in July and Anthropic's figures are a full year.

"AI-native" is a judgement call. Palantir was founded in 2003 and sold data analytics long before large language models. Including it gives the scorecard its only profit. Excluding it would make the headline 0 of 6, and you could argue for either. I kept it in because its current growth is driven by its AI platform, but a reader who drops it is not wrong.

The sample is what is public. OpenAI, xAI, Mistral and most of the AI application layer publish no audited statements. The companies that do report are not a random sample. A profitable private AI company would not appear here, and nor would the many loss-making ones.

Two policies are secondary-sourced. The useful-life figures for CoreWeave and Nebius come from coverage of their annual filings, not from the notes themselves. The Anthropic figures trace to a single Reuters reading of an unpublished draft.

The case that adjusted figures are the honest ones

There is a serious counter-argument. Stock compensation is non-cash, and a growing GPU cloud depreciates hardware it bought in advance of demand. On that view, adjusted EBITDA shows the earning power of the installed base, and GAAP operating income penalises growth.

That argument is strongest for CoreWeave, whose $1.510 billion of adjusted EBITDA is real cash generated by existing contracts. It is weakest for C3.ai, where removing stock pay removes a cost larger than revenue. My view is that adjusted figures are a legitimate second lens and a poor first one. A business is profitable when it covers the full cost of what it sells, and for a neocloud the GPU is what it sells.

Frequently asked questions

How to use this scorecard on your own AI vendors

If you buy from any company in this table, or from a private one like it, ask for 3 numbers in the same conversation. They are GAAP operating margin, stock compensation as a share of revenue, and capex as a multiple of revenue. A vendor who gives you only adjusted EBITDA has told you what it wants you to see.

Then set the next earnings dates for CoreWeave, Nebius and Palantir in your calendar, and read one line in each release before anything else: income from operations. If a second name crosses zero, that matters more for AI pricing than any revenue headline will.

Related reading