The Denominator Problem | RLK Consulting

Profit margins across the S&P 500 just hit roughly 15.7 percent, the highest FactSet has recorded since it began tracking the measure. The full-year estimate sits near 13.9 percent, the best annual number since 2008. Margins widened straight through a stretch of higher energy and input costs, and the chart is already showing up in board decks as proof that the AI payoff has arrived.

Before anyone frames that chart, two questions. Which line actually moved? And who, exactly, is pocketing the improvement?

I do technology strategy, not finance. But you can’t advise on AI spend without reading income statements, and a year of reading them for mid-market clients turned up answers less flattering than the chart. Also more useful.

Where the gain actually sits

Revenue per employee has become the favorite way to describe AI’s business impact, and the metric has genuinely moved. Median revenue per employee at public software companies hit roughly $395,000 in 2025, up from about $327,000 three years earlier.

A ratio has two ends, and everyone quoting this one is assuming the flattering end did the work. It mostly didn’t. Software companies spent those three years cutting staff on purpose. The go-to-market teams I see are hitting targets at headcounts that would have looked understaffed in 2023, and customer-success groups a quarter leaner are holding their satisfaction scores. The numerator grew at an ordinary pace. The denominator shrank by design.

That distinction is the whole ballgame for a $40M or $150M company deciding what to expect from its own AI program. A denominator gain is a one-time step change. You can delete a department once. There is no sequel. The second year produces nothing further from the same move, and the company arrives at a lower cost base with the same revenue and the same customers it had before.

The largest study of enterprise deployment says the same thing with a bigger sample. MIT’s GenAI Divide research examined around three hundred public AI deployments alongside executive interviews and found that about 95 percent of pilots produced no measurable P&L impact at all. The gains that did show up concentrated in individual productivity, which is speed felt by the person doing the task and invisible in the quarterly financials.

Meanwhile, the AI revenue that does exist is easy to find. Microsoft’s AI business passed a $37 billion annual run rate, growing 123 percent, with Microsoft 365 Copilot past twenty million paid seats, up 250 percent in a year. Google Cloud grew 63 percent last quarter, to a $20 billion quarter, with another $462 billion in signed contracts it hasn’t even billed yet. AWS is running at a $169 billion yearly pace, with its AI business alone past $25 billion. ServiceNow’s AI products crossed a billion dollars of annual contract value last quarter.

Every one of those dollars appears on some other company’s income statement as an expense. The AI economy has a clear, well-documented winner category so far, and it is, as tradition requires, the people selling the shovels.

The revenue-per-employee legends sit in the same category. Lovable hit $400 million of annual recurring revenue with 146 employees, about $2.77 million per head. Midjourney runs hundreds of millions in revenue on a headcount around a hundred, without ever taking outside money. Anthropic and OpenAI post revenue per employee above anything in the Forbes Global 2000. These numbers get cited constantly as evidence of what AI does for a business. They are evidence of what AI does for a business that sells AI. Benchmarking your distribution company against Midjourney is astrology with spreadsheets.

Before I go further, the best argument against everything I just said. Strip Alphabet and Amazon out of the S&P 500 and the margin still sits near 15 percent, still a record. The cost reduction across the broader index is real, it is reaching actual bottom lines, and anyone waving it away is overcorrecting. The open question is tenure. How long does the company that captured the saving get to keep it?

The takeaway: the record margins are a cost story, and the surest AI money so far belongs to the companies selling the tools.

The savings have two predators

The mechanism doing all the work is symmetry. A capability you rent is symmetric: every competitor can rent the same one, on the same terms, from the same vendors, next quarter. Symmetric advantages get competed into price. Asymmetric advantages, the ones built on something rivals cannot buy, are the only kind that stay. Almost everything sold as “AI transformation” today is symmetric, which is why almost all of it lands in the cost line and why the cost line is where the erosion starts.

Economists watching the macro picture describe the same sequence from altitude. The New York Fed’s Liberty Street Economics puts the near-term effect on the inflationary side, driven by the cost of the buildout, with the productivity payoff arriving late in the decade. Economists file that payoff under disinflation, which is a polite word for your efficiency gain leaving your P&L and reappearing in your customer’s. Robert Solow noted in 1987 that the computer age was visible everywhere except the productivity statistics. The gains eventually arrived, and they arrived mostly in prices. Buyers got them.

For a mid-market operator that turns into a forecast with a date on it. A distributor that pulls 8 percent out of order-processing cost with tools its four competitors can license next quarter holds the advantage for as long as it takes those competitors to notice. Then the saving shows up in bids, and the industry settles at a permanently lower cost structure with margins roughly where they started. Anyone who lived through ERP, offshoring, ecommerce, or the cloud migration has watched this movie. Each wave produced enormous real efficiency. Almost none of it stayed with the adopters, because adoption became table stakes within a few years, and the companies that came out ahead were the ones that used the window to change what they sold.

That’s predator one: your competitors, arriving on a schedule set by their procurement calendars.

Predator two is your vendors, and their schedule is written in bond documents.

What the cloud builders are spending

The bill behind your discount.

Sources: 2025 combined actuals as reported by the five companies; 2026 estimate from CreditSights and FactSet. The group used to self-fund this spending. New borrowing covered 9% of it in fiscal 2024 and 32% by mid-2026, per FactSet.

Today’s AI pricing is a land grab paid for with other people’s money. The five biggest cloud builders spent about $405 billion on data centers and chips in 2025 and plan more than $690 billion for 2026. They used to pay for this out of pocket. FactSet reports new borrowing covered 9 percent of that spending in fiscal 2024 and 32 percent by mid-2026. The gap is being borrowed at speed: about $108 billion raised in the bond market in 2025, then $194 billion in the first half of 2026 alone. Goldman Sachs expects more than a third of Big Tech’s AI investment to run on borrowed money by 2027, and Fortune counts another $1.65 trillion in spending promises that sit outside the official balance sheets.

Then there’s the money going in circles. Nvidia invests in OpenAI, OpenAI commits hundreds of billions to cloud providers including Oracle, and those providers buy Nvidia silicon to deliver the capacity. The same dollars orbit the group and get booked as somebody’s growth at every stop. Bloomberg mapped the deals; the map is a circle, and 2026 tallies put it above $800 billion. OpenAI’s own internal forecast has it losing about $14 billion this year. The bond market noticed. The price of insuring Nvidia’s debt against default just posted its biggest one-day jump on record.

The accounting stretches the fuse without defusing anything. A server’s cost gets spread across five or six years of income statements, a building’s across twenty-five to forty, and the loans behind both stretch decades. So today’s income statements show a sliver of what this buildout actually costs. The rest arrives on a schedule, and it arrives whether or not demand grows into it.

Two ways this resolves. Either demand grows fast enough to pay for all of it at today’s prices, or the lenders come looking for their return and the price of intelligence goes up. The second path is already underway, and you can see it in the fine print. AI capability keeps migrating into premium tiers. Credits and usage meters are showing up inside contracts that used to be flat. Upgrades get structured so that declining them means losing things you already had. Microsoft attached Copilot to Microsoft 365 alongside a subscription increase, which is a price hike wearing a keynote. Enterprise renewal quotes now routinely carry a material premium for AI-enabled editions of software the customer already licensed.

So a mid-market CFO should treat current AI pricing as promotional, because it is promotional in the strict casino sense: the first night is cheap and the house remembers. Model every AI-dependent workflow at two and three times today’s unit cost and see what survives. A business case that only clears at present pricing is a wager on somebody else’s balance sheet staying patient. The defensive work is unglamorous and it works: caps on annual price increases, real exit rights, your data in a format you can leave with, and a second provider kept warm enough to matter at renewal.

The takeaway: savings from rented tools leak twice, to competitors who copy them and to vendors who reprice them. Model every AI business case at two to three times today’s prices and see what survives.

What has actually produced revenue

A small number of buyer-side moves have generated real top-line growth, and every one I’ve examined shares the same structural property: the advantage sits in something a competitor cannot purchase. They built their asymmetry on purpose, from four kinds of material.

Proprietary data. A specialty manufacturer with fifteen years of failure records against specific operating conditions can build a predictive-maintenance offering nobody can copy, because the model is only as good as the history behind it. The AI is a commodity. The data is the asset, and the company already owned it.

Distribution. A firm with an installed base and a service relationship can attach a new AI-enabled offering at near-zero acquisition cost. The identical offering built by a startup dies on the price of the sales motion.

Regulatory position and trust. Where the buyer needs an accountable counterparty with insurance and a compliance record, the incumbent can sell an automated service that a technically superior challenger cannot place.

Changing the pricing unit. My favorite, because it’s the most available and the least used. Companies that sold hours and now sell outcomes convert an efficiency gain into margin the customer cannot see and therefore cannot negotiate away. Professional services firms are perfectly positioned for this and mostly won’t do it, because it requires the seller to accept measurement risk, and the billable hour is a warm bath.

Direction is the common thread. All four start from an asset the company already held and ask what AI just made sellable. Programs that start from the tooling and go hunting for applications land in cost reduction nearly every time, because cost reduction is where general-purpose capability points when nothing proprietary is steering it.

Take a $70M industrial services firm running preventive-maintenance contracts. The cost-out version of its AI program routes technicians automatically and trims dispatch overhead, worth a few points of route density. Real money, available to every regional competitor within a year or two, then gone into pricing. The revenue version starts from the twelve years of technician notes sitting in its field-service system. Those notes hold failure signatures for specific equipment under specific duty cycles, and no competitor has that history on that installed base. Turned into a condition-monitoring subscription sold to customers already under contract, the archive becomes a second revenue line at margins the maintenance business never saw. Both versions use identical, widely available technology. The asset under one of them took twelve years to accumulate, which is why it will still be earning in year five when the routing savings are long gone into bids.

The takeaway: durable AI revenue starts from an asset you already own that competitors can’t buy. Start from the asset. Never start from the tool.

Ten minutes with the board

Three questions locate any company on this map.

What has AI changed about what we sell? Accept only answers a customer would recognize. Faster quote turnaround counts if customers chose you for it and the win rate moved. Internal cycle time with a flat win rate is a cost story that wandered into the growth section.

What would we charge for this capability if we sold it on its own? And has anyone tested that? A capability with no standalone price is an internal efficiency. Useful and legitimate, and a different thing.

Which competitors can buy what we bought, and how fast? Under eighteen months means the advantage has an expiration date, which is fine so long as the plan treats the date as real.

Most companies will find they are making the same money faster and cheaper. That has genuine value, and it deserves to be banked deliberately rather than assumed: Sysco just put roughly $100 million of AI-driven cost savings into its official forecast for next year, in writing, where analysts can hold them to it. That is what taking the cost story seriously looks like.

The exposure is stopping there. A cost-only position sits between two predators on two clocks: competitors copying the saving into price on one side, vendors repricing a debt-financed buildout on the other. A company holding a lower cost base and nothing else when both clocks run out has spent three years to arrive back where it started, minus the subscription.

Efficiency buys time. Revenue comes from what you build with the time, and the building has to start from something you already own that nobody else can buy. Start with your field-service system, your failure records, the archive your industry would pay to see. The window is open and the clocks are running. The shovel vendors send their regards.

How RLK Can Help

The AI Strategy and Capital Allocation engagement exists for this fork: which parts of the program belong in the cost column, which assets you already own could carry a revenue line, and what to pay for either. The AI Business Case engagement stress-tests the numbers, including what survives at two and three times today’s AI pricing. And my free AI Diagnostic will show you in about eight minutes whether your current program is a numerator story or a denominator story. Start the conversation.