The table has five lines. Google Services earned $39.5 billion in operating income in the second quarter. Google Cloud earned $8.8 billion. Other Bets, the collection of Alphabet's smaller ventures, lost $1.8 billion. Then comes a line that belongs to none of them, "Alphabet-level activities," at minus $5.8 billion, and below it the total for the whole company: $40.8 billion.1 The eye goes to the first two lines. The fourth is the one that tells you what the first two leave out.
Alphabet published those figures in its earnings release last Wednesday, July 22, and repeated them in its quarterly report on Form 10-Q, filed with the SEC on July 23.12 The 10-Q explains the fourth line in one footnote: Alphabet-level activities "primarily reflect expenses related to our shared AI research and development."2 Nothing is concealed here. The line is printed, labeled and explained, and the arithmetic closes: 39.544 plus 8.814, minus 1.799, minus 5.789, is 40.770 billion dollars. The statements are interim and unaudited, as quarterly statements usually are.
My argument is that a segment operating result and a consolidated operating result answer different questions, and that the difference lies in where the boundary is drawn around costs. The segment figure says how a business performs under the company's rules for which costs it carries. The consolidated figure says what the whole enterprise earned once every cost is counted somewhere. Neither replaces the other, and the reconciliation between them is what makes both usable. This is not a claim that Alphabet's segments are overstated, nor an estimate of what any segment would earn under a different rule. It is a claim about how to read a segment number, and about what to demand of the ones we produce inside our own companies.
Operating income (loss) · Q2 2026
three months ended June 30 · USD billions · as reported
Each segment carries the costs its rules assign it.
Printed and explained, not hidden: mostly shared AI research.
The whole company, every cost counted somewhere.
Where the boundary runs
The interesting part of the 10-Q is not the table but the note that explains how it was built. Some costs, it says, are directly attributable to a segment: traffic acquisition, certain engineering work, inventory. Others, such as technical infrastructure and office space, are managed centrally and then charged to the segments as a service cost, generally based on usage, headcount or revenue. A third group is not charged to any segment at all. It includes certain shared AI research and development, including the pay and the infrastructure used to develop Alphabet's general AI models; corporate initiatives such as philanthropy; and corporate costs such as parts of finance, human resources and legal, including certain fines and settlements. Severance, office reductions and revenue hedging gains and losses also stay outside.2
That description corrects a tempting misreading. Google Cloud's segment result is not free of infrastructure costs; the note says Cloud's operating income rose $6.0 billion from a year earlier, mainly on higher revenue, partly offset by higher usage costs for technical infrastructure and higher employee compensation.2 What stays outside the segment is narrower: mainly the development of the general models that, by the company's own description elsewhere in the filing, serve several of its businesses.
The filing also says who uses the segment figure and for what. Alphabet's chief executive, as the company's chief operating decision maker, uses segment operating income to allocate resources in annual planning and to judge each segment against its plan.2 That is the purpose the boundary serves. A segment result is designed to answer a management question about a business, under rules management chose. It is not designed to answer every question a reader might bring to it.
Why shared work resists division
Shared research is hard to assign for a reason that has nothing to do with accounting tricks. A general model trained this quarter may improve a search product, a cloud service and a product that does not exist yet. There is no fact of the matter about how much of its cost belongs to each, only rules that are more or less reasonable for a given purpose. Charging it all to the segment that sells it most directly would load one business with work that serves several; spreading it by revenue could charge the largest business for research it uses less than others do. Leaving it at the company level, as Alphabet does, avoids inventing a division and puts the whole amount in plain view.
The size of that line has grown. In the second quarter of 2025, Alphabet-level activities came to $3.4 billion; a year later, $5.8 billion.2 A reader who tracks only the segment lines would miss that growth entirely. A reader who subtracts the whole amount from one segment and calls the result its "true" profit would invent a number the company never reported and no rule supports. The responsible reading keeps both lines in view and the bridge between them intact.
The same table, inside a company
Every large organization has its own version of this table, and the people who build it are often in technology. A shared data platform, an identity service, a model team: each serves several business units, and each eventually needs a number in somebody's investment case. A small example, invented for the purpose, shows what the boundary does. Two units earn $30 and $20 before a shared platform that costs $10. The group earns $40. Assign the platform by usage, by headcount or to its heaviest user, and the two units' results change. The group's does not. Figure 2 lets you move the cost.
Allocation worksheet · illustrative
Held at group level, the units show 30 and 20, and the 10 sits on its own line. Unit B clears its 15 hurdle. The group earns 40.
The total is the same under every rule, and that is the point. Allocation does not create or destroy profit; it decides which unit's story absorbs a cost. But stories have consequences. If Unit B needs to show 15 to keep its budget, it passes when the platform is held centrally, passes by usage, sits exactly on the line by headcount, and fails when it carries the whole cost. The platform cost the same in all four cases. Only the decision changed.
That suggests a discipline for internal cases. Say which costs are directly attributable to the unit, which are shared, and what rule assigns the shared ones. Say which decision the number is meant to inform. Then test whether that decision survives the other reasonable rules. If it does, the rule hardly matters. If it does not, the decision depends on the rule, and the rule is what deserves the board's debate.
Average is not incremental
Even a perfect reconciliation describes the past and the average. A fully allocated margin says how a business performed with all its costs counted; it does not say what the next dollar of investment will return. A platform already built is a cost the next project shares without causing, and a new one may cost far more, or far less, than the average suggests. The question a board faces about a new investment is incremental: what changes, in cost and in revenue, if we do this and not otherwise? A supplier's segment margin, however carefully read, cannot answer that for us. Neither can our own historical allocation.
Objections, considered
Segment reporting exists to aid understanding. It does, and this argument depends on it. The segment table, the note on allocation and the reconciliation together tell a reader more about Alphabet than a single total could. The point is to use all three, not to discard the segments for the consolidated figure.
Shared research is an investment in future products, not a cost of this quarter. That may well describe its economic purpose. It is still reported as a current expense under the accounting that applies, and a reader who mentally moves it onto a balance sheet is substituting a judgment for a reported fact. Both views can be worth holding, as long as they are labeled.
Allocation is subjective anyway. Often it is, which is why the rule should be written down and why the decision should be tested under the alternatives. Subjectivity is a reason for disclosure, not for pretending a single number settles the matter.
A supplier's margin says little about our project. Exactly. What transfers from Alphabet's table to an internal proposal is not a margin assumption but a habit: name the perimeter, keep the bridge, and know which question the number was built to answer.
A profit number needs a boundary before it can support a decision.
Before the next investment case
- Draw the perimeter. Which costs sit inside the unit's result, which sit elsewhere, and where is the line that reconciles them to the total?
- Name the rule. How are shared platforms and teams assigned, and who chose that rule, for what purpose?
- Try the alternatives. Does the decision in front of you survive a different reasonable allocation, or does it depend on this one?
- Ask the incremental question. Is the case built on an average margin, or on what changes if this investment is made?
- Keep the bridge. Can someone get from the unit's figure to the company's total, line by line?
Back to the five lines. Read in order, the fourth one no longer looks like an interruption between the segments and the total. It is the reason the total can be trusted and the reason the segments can be compared: a disclosed boundary, with a name, a size and a footnote. The profit was there all along. So was its perimeter.