The Vanished Wages

"Firing on all cylinders." "Industry-leading growth." "A record quarter." "An earnings beat."

Those were the headlines after Stagwell reported their Q2 on Thursday last week. Surprised, I took the opportunity to review their results this weekend.

Here is what the same release says. The company lost money. Loss per share of $(0.03). Net loss attributable to shareholders of $8 million, wider than the $5 million it lost a year ago. And operating income halved, from $23.2 million to $11.5 million.

So why the headlines? I want to be careful here, because the easy version of this piece is a swipe at the trade press, and the easy version is wrong.

Every one of those headlines is accurate, but also incomplete. Adjusted EBITDA did rise 15% to $109 million. Net new business of $171 million was a company record, 45% up on last year, and included taking IBM from Ogilvy after thirty-two years. Adjusted earnings per share grew 39%. All of this is near the top, where any journalist filing to a deadline would write exactly what they wrote.

(And to be fair, MediaPost was the only outlet I could find that added anything, carrying Brian Wieser's Madison and Wall observation that Stagwell did not disclose political revenues and that they were likely worth around two points of the reported organic growth. Everyone else ran the release. More on the political point later.)

The number that explains all of it is not in the income statement. It is in the investing section of the cash flow statement, roughly twenty pages deep. In The Standing Gallop last week I took apart the three mechanisms that let a truly flat Omnicom report 6.1% 'growth'. Same discipline this week.

Earnings releases are designed documents, and this one was designed to be read exactly the way it was read.

Here it is. It seems that Stagwell is developing some pretty serious proprietary software.

In the first half of 2026, Stagwell capitalised $62.4 million of software. In the first half of 2025, it capitalised $29.2 million. Stagwell capitalised almost as much software in six months as it did in the whole of last year.

And the acceleration is the story. $28.2 million in 2023. $35.1 million in 2024. $67.5 million in 2025. Then $62.4 million in six months.

If "capitalised" makes you think of school and proper nouns, and not about accounting principles, stay with me for a minute while I explain why it matters.

Nadia runs a bakery. She employs eight bakers, pays them £40,000 a year each, and every penny of that comes off her profit. Wages are a cost. Bread is sold, wages are paid, the year closes.

Then Nadia takes two of her bakers off the bread line for six months and has them build a new oven in the back. Same two people. Same wages. Same money leaving her bank account.

But now she doesn't call it wages. She calls it an oven. Forty thousand pounds of salary stops being a cost and becomes an asset worth forty thousand pounds, sitting on her balance sheet. Her profit for the year goes up by forty grand, and she has an oven. She'll charge an amount each year from her profit based on how long the over will be useful.

Nothing about this is a trick. If there is a real oven, that is exactly the right way to account for it. You don't charge the whole cost of an oven against one year's bread.

But notice what now depends on judgement. Is there an oven? And is Nadia baking anything with it, or is it standing in the back looking impressive?

Now the last turn, and it is the one that matters. Nadia asks to be judged on profit before the oven charge, because an oven is an investment, not an operating cost. So the £40,000 never appears anywhere in the number she is measured on. Not when she spent it, because it became an asset. Not over the following decade, because it is excluded.

Two bakers worked for six months and it vanished from the story.

Every holding company currently claims to be building proprietary software. So let's compare their books.

Stagwell builds its own. $84.1 million of capital expenditure and capitalised software in the first half, against $1,490 million of revenue. Around 5.6%.

Publicis buys. Net capital expenditure in fixed assets of €91 million in the first half, down from €115 million a year earlier, or roughly 1.3% of net revenue. The half-year release doesn't break out software development separately, so the comparison is directional rather than exact. But the group this industry most associates with proprietary technology, with Epsilon and CoreAI, spent less on building than it did last year. It has been buying instead: LiveRamp, Lotame, Captiv8, Adge, Moov AI.

Omnicom rents and expenses. First-half capital expenditure of $115.1 million against revenue of $12,805.4 million, a fraction under 1%. Their intangibles note shows purchased and internally developed software rising from $275.6 million gross to $289.0 million over six months, so about $13 million of additions, on a $49.9 billion balance sheet.

Omnicom is roughly ten times Stagwell's size and spent $115 million to Stagwell's $84 million. On the same measure, gross revenue against gross revenue, Stagwell is running at six times the intensity.

One explanation you can rule out is tax. American software development did go through a period of forced capitalisation (the 2017 tax act required development costs to be spread over five years) until that was reversed by the "One Big Beautiful Bill" Act (so funny), signed in July 2025, which restored immediate deduction for domestic development. But tax treatment and accounting treatment are separate systems, neither compels the other, and both of these companies file under identical American rules. Whatever explains the gap between them, it is not Washington.

Three objections, and they are all good ones.

First: a platform costs roughly the same to build whether you are a $2.4 billion company or a $15 billion one.

Intensity ratios systematically flatter the big. Stagwell running at six times Omnicom's percentage may be arithmetic about size rather than evidence of anything.

Second, the products exist.

BERA.ai grew 28% last quarter and the Harris Quest family grew 19%. There is $16 million of committed Enterprise Tech Solutions revenue booked in the first half and a $60 million multi-year government contract signed. Stagwell has 1,370 engineers inside Digital Transformation and is expanding that base into Latin America, Egypt, India and South East Asia. This is not a company pretending to have a technology business.

Third, spending less is not automatically virtue.

Publicis buying its platforms means renting your future from companies you acquired, with its own integration risk and its own amortisation. Omnicom at 0.9% has simply decided its future is scale rather than code.

Three bets, three risks. Not one villain and two saints.

So: is Stagwell building proprietary software that genuinely belongs on a balance sheet, more so than anyone else in the industry? Or is it managing a number?

Both explanations fit every fact I can see. And I think the honest answer is that they are not mutually exclusive: that Stagwell is building something real, and that the marginal judgement calls have been made in the direction that flatters the figure the company is judged on. That is not a scandal. It is how most companies behave under pressure.

What would settle it is software revenue. A licence line. Recurring subscriptions, disclosed, growing.

Put plainly: is she baking anything with it?

Show us the software revenue and the question answers itself. It isn't there.

And there is a second question underneath it.

The platform at the centre of Stagwell's AI story is built with Palantir. The partnership, announced last November, pairs Palantir's Foundry with Code and Theory's orchestration software and the Marketing Cloud's proprietary data. Stagwell's own description of the architecture is precise and worth reading twice: its software layer sits on top of Foundry.

That architecture contains two different kinds of cost, and they are not accounted for the same way. Software Stagwell writes and controls is an asset it owns. Capitalise it, amortise it, and nobody should object. Foundry is not Stagwell's. The right to use somebody else's hosted platform is an operating expense, and the work of configuring one is treated differently again and lands in a different part of the cash flow statement. All $62.4 million appears in investing activities. Stagwell does not disclose the split.

Which returns us to the bakery. Nadia didn't build the whole oven. She built the door and the controls, and bolted them onto an oven she rents from the firm across town.

Let's run it the other way, to see what the results of the real business are being managed by this treatment. Using Stagwell's own figures and its own 26.5% normalised tax rate:

40% earnings growth in the headline and widely reported. 0% based on backing out the software treatment and based on the same number of shares in both periods. Quite a big difference. But it gets them through this quarter before the political feeding trough arrives this quarter.

In holding their full year guidance on earnings, Stagwell has deliver almost all of it in the second half of the year.

Management gave its reason on the call. Operating cash flow improves from here because of seasonality and anticipated election-related communications activity. Communications grew 12% organically last quarter, from $97.6 million to $112.0 million, and CEO Mark Penn was explicit that this is before the political supercycle properly begins: the midterms, then the primaries, then 2028. Stagwell's own risk factors state that political activity concentrates revenue in the third and fourth quarters of even years.

When I wrote Stagwell's Crusade in November, and again in March when the advocacy line halved in an off-year, the argument was that Stagwell is the only major holding company whose financials visibly swing with the American political calendar. I framed it then as a client risk question.

This is starkly different to Publicis in both ways, based on my time on the Management Committee there. Publicis eschewed all political work, in any country, and even when we cancelled all awards investments and entries to prioritise the build of the Marcel AI platform it was always, to my recollection, expensed and not capitalised.

Stagwell builds. Publicis buys. Omnicom consolidates and rents. Those are three genuinely different answers to the same question: how do you survive in an industry the platforms are steadily eating? The divergence is strategic before it is ever accounting. These companies are no longer doing the same thing.

The same activity, building the technology every one of them describes as existential, now lands in three different places in three different sets of books. In one it lifts reported profit. In another it arrives at a price somebody else set, with an invoice to prove it. In the third it comes straight off the bottom line.

We have spent this year arguing about who is winning. The new business tables cannot be verified, as I set out in The Retention Trophy. Organic growth depends on where the perimeter is drawn. Net and gross mean different things at different companies.

On Thursday this week, CEO Cindy Rose presents WPP's first half, giving us a fourth company, with a platform of its own to account for, reporting under a different framework again. WPP and Publicis file under international standards, where development costs are tested against six criteria. Omnicom and Stagwell file under American ones, where most research and development is expensed outright and software is the carve-out. Four companies, two rulebooks, and every one of them applying its own standard faithfully.

Because the harder question is no longer which of these companies is winning.

It is whether we can still tell.

Is there a single number left that lets you compare these companies fairly?

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 3 August 2026 in the Brandflow newsletter on LinkedIn.