SAP reported second-quarter results Thursday that looked good where the company most needed them to. Current cloud backlog, the metric that leads revenue, grew 26% at constant currency to nearly €23 billion, an acceleration from the first quarter that reversed two quarters of lagging behind cloud revenue growth. Chief executive Christian Klein called it a welcome trend reversal, which it was.

The more consequential number came out of an analyst question. UBS asked why research and development costs jumped 14% against a headcount increase of only 3%. Klein's answer was that token consumption in AI development and the hiring of higher-cost AI experts drove the gap, and that he expects headcount growth to moderate as productivity gains offset the spending.

That is a small exchange describing something structurally new about how software companies spend money.

Why R&D decoupling from headcount is a change

For as long as enterprise software has existed as a category, R&D expense has been approximately equal to people. Engineers were the input, salaries were the cost, and you could model a software company's development spending off headcount and compensation with reasonable accuracy. It is why R&D as a percentage of revenue was a stable and comparable metric across companies, and part of why software carried such attractive economics: the marginal cost of serving one more customer approached zero, and the marginal cost of building was a person you chose to hire.

Token consumption breaks that relationship. AI-assisted development introduces a variable, usage-metered input into the act of building software, one that scales with how much work gets done rather than with how many people are employed. SAP's numbers are an early quantification of the effect: a 14-to-3 ratio between cost growth and headcount growth is a large gap to attribute to compute and specialist salaries.

The implications run in two directions at once, which is what makes it worth watching rather than simply worrying about.

The favorable reading is Klein's. If tokens substitute for engineering hours, output per euro rises, headcount growth moderates, and a company eventually builds more software for less. That is the entire productivity thesis for AI in software development, and if it holds, the current cost bulge is a transition rather than a new baseline.

The unfavorable reading is about the shape of the cost rather than its size. Headcount is a fixed cost you control through hiring decisions, adjustable in steps and forecastable a year out. Metered compute is variable and scales with activity, which makes development spending more responsive to usage and less predictable. And if every competitor adopts the same tools, the productivity advantage competes away in pricing while the cost line remains. You would be left with a new expense category and no durable edge.

Which reading is right will not be visible for several quarters. What is visible now is that the expense arrived first.

AI is showing up as cost on both sides of the income statement

The R&D line is only half of it, and the other half appears in gross margin.

SAP's non-IFRS cloud gross margin came in at 74.6%, down 0.7 percentage points year over year. Separately, Goldman Sachs had already cut its second-half 2026 gross margin forecast for the company, from 73.3% to 72.8%, citing higher hardware costs alongside dilution from recent acquisitions.

This is the part that distinguishes AI features from ordinary software features. A traditional SaaS capability, once built, costs almost nothing incremental to deliver to another user. An AI capability consumes compute every time it runs. Serving it has a cost of goods sold in a way that serving a database query does not, and that cost scales with adoption rather than being amortized away by it.

So the same technology is appearing as a cost in R&D, where it is used to build, and in COGS, where it is used to serve. The revenue benefit, meanwhile, is largely prospective. Management said AI and SAP Business Data Cloud featured in more than 90% of its 50 largest deals, which is a striking demand signal, and it is a statement about what closes deals rather than about what those features are separately monetized for.

That gap between certain present cost and expected future revenue is the actual investment question in enterprise software right now, and SAP is a clean instance of it.

Bookings accelerated while profit growth slowed

The quarter had a consistent internal shape once you separate the leading and lagging measures.

Cloud revenue rose 24% to €6.3 billion and total revenue 11% to €9.9 billion, while non-IFRS operating profit grew 9% to €2.7 billion, a deceleration from the prior quarter. Management attributed the slowdown to lower revenue growth than Q1, an unusually low stock-compensation expense in Q1 that flattered the comparison, accelerated R&D investment, higher marketing spend around the autonomous enterprise launch, and acquisition dilution.

Those are all real explanations rather than evasions, and most are timing effects. The composite picture is a company whose order book is speeding up while its profit growth slows, because it is spending into a product transition. That is a defensible posture and it is also indistinguishable, from the outside and in any single quarter, from a company buying growth. The difference shows up later, in whether the backlog converts at the margins the model assumes.

SAP also trimmed its 2026 non-IFRS operating profit guidance to €11.8 to €12.2 billion to reflect more than €100 million of expected dilution from the Dremio and Prior Labs acquisitions, which closed in early and mid-July, while reiterating cloud revenue guidance and the roughly €10 billion free cash flow target. Cutting profit guidance for a disclosed acquisition effect while holding revenue and cash flow is about as benign as a guidance reduction gets.

The question the analysts pressed

The sharpest exchange on the call was not about AI. BNP Paribas noted that first-half cloud revenue was running ahead of guidance while the implied second-half deceleration looked drastic, and asked why. Klein attributed the caution to macro uncertainty, particularly in the Middle East, while saying the post-Sapphire pipeline was stronger than a year earlier and that the company was increasingly confident of hitting guidance.

That answer is genuinely ambiguous and both readings are reasonable. Conservative guidance after a strong first half is standard practice and sets up a beat. It is also what a management team says when it sees something in the pipeline it does not want to underwrite publicly. Management separately said it expects current cloud backlog growth to decelerate slightly exiting the year, which is consistent with deliberate conservatism and also with an actual expected slowdown.

The third quarter resolves it, and there is no way to resolve it now.

The valuation gap is the loudest signal

The strangest fact about SAP is not in the earnings report. The stock has lost roughly a third of its value over twelve months and trades more than 47% below its 2026 peak, near its 52-week floor, while Morningstar carries a fair value estimate close to double the current price and consensus targets imply upside above 50%.

That is an unusually wide divergence between what analysts model and what the market pays, and it means the market is pricing something the backlog does not capture. Several candidates are plausible: that AI compute costs permanently compress the margin structure that made enterprise software valuable; that cloud growth is decelerating on a schedule the backlog has not yet revealed; or simply that European software multiples have compressed for reasons that have little to do with SAP's execution. Distinguishing among those from outside is not possible, and anyone claiming certainty about which one is operating is guessing.

What can be said is that the divergence is a testable proposition rather than a permanent condition. If margins stabilize while backlog converts at guided rates, the analysts are right and the price is wrong. If gross margin keeps eroding as AI adoption scales, the market has been early rather than mistaken.

What to watch

Three things, in order of how much they would change the picture.

Cloud gross margin trajectory, because it is the direct measure of whether AI delivery costs are a transitional bump or a structural reset. A margin that stabilizes near current levels supports the entire bull case. Continued erosion as AI usage scales undermines it regardless of how fast revenue grows.

The relationship between R&D spending and headcount over the next several quarters, because Klein made a specific and falsifiable prediction that headcount growth moderates as productivity gains offset token costs. If R&D keeps growing at four or five times the rate of headcount without a corresponding acceleration in output, the productivity thesis is not working as described.

And whether the second-half deceleration management guided toward materializes. A beat confirms conservatism. A miss confirms that the caution was informed.

The bookings number was the good news this quarter and it is the one most likely to be quoted. The R&D line is the one that describes something new, and it is the number worth carrying into every other software earnings report this season.

Further reading