TL;DR
Anthropic's November ANT filings show a much larger September quarter and a lower operating margin. In the unaudited quarterly tables, revenue rose from $11.555 billion in June to $16.956 billion in September. GAAP operating income swung from $974 million to a $567 million loss. Adjusted operating income increased from $2.235 billion to $2.463 billion, while the adjusted margin fell from 19 to 15 percent as rounded in the filing. The sharper question for the next quarter is how growing usage, short-term compute, and larger capacity commitments change incremental operating costs.
The analyst arrives with a June workbook and two numbers from a November headline. Her colleague has read the new prospectus and wants to mark Anthropic both more profitable and less profitable. Both of his numbers are real. Each sits in a different row of the filing and tells a different story. The two decide to reconcile the September quarter before drawing a view about the business Claude supports.
The conversation is fictional. The operating results, expense categories, management descriptions, and non-GAAP reconciliation come from Anthropic's November 17 S-1 amendment. The quarter comparisons are made from the company's unaudited quarterly consolidated tables. The prior public October version disclosed the June period. An interim quarter supplies dated financial evidence inside the offering process. November 17 S-1 amendment and October 26 S-1
The first question is what grew by five billion dollars
The analyst divides $16.956 billion by $11.555 billion and subtracts one. Sequential recognized revenue growth is approximately 46.7 percent. The numerator and denominator each cover three months. The result answers how the income statement changed from the quarter ending June 30 to the quarter ending September 30. It does not describe revenue per unit, year-over-year growth, or billed usage since the end of September.
Her colleague opens the nine-month table. It reports $33.242 billion of revenue through September 30, up from $2.545 billion for the same nine months of 2025. Checking the quarterly sum reinforces the framing. Add March's $4.731 billion, June's $11.555 billion, and September's $16.956 billion to reproduce $33.242 billion. The latest quarter accounts for slightly more than half of that nine-month figure. The equality is a useful check on periods in a spreadsheet with more than one window.
The growth description in management's discussion identifies successive product and model releases, greater usage, and broader adoption as part of the explanation. It also describes roughly $600 million of negative price effects between June and September associated with new-model prices and contractual discounts, which partially offset higher usage. That disclosure supports an important possibility. Revenue can rise sharply in the same period that some realized unit prices fall. The way those changes enter the income statement depends on the customers, products, and units involved.
A colleague proposes adding $600 million back to September revenue and calling the result what the company would have earned at June prices. The analyst labels the sum a simple arithmetic illustration under an all-else-equal assumption. A different price can change consumption behavior, composition, and contract terms. The filing's reported price effect helps explain the comparison. It leaves the business response to a hypothetical alternative price outside the observed results.
The same care applies to a quarterly number multiplied by four. Four times the reported September quarter is $67.824 billion. That arithmetic expresses a constant-quarter annualization of past recognized sales under a stated assumption. A reported run-rate metric can have a different definition, and a forecast needs a view about future usage, discounts, availability, and recognition terms. The analyst stores the quarterly observation in its own cell and attaches a definition to any number derived from it.
A variance walk for the operating result
The operating statement supplies enough detail to rebuild the $1.541 billion change in GAAP operating income. Start at $974 million. Increase by the $5.401 billion difference in revenue. Then carry in each change in operating expense. The accounting periods use the same millions units, so the arithmetic has no hidden translation to dollars or annualized values.
| Quarterly operating item in USD millions | June 2026 | September 2026 | September less June |
|---|---|---|---|
| Revenue | 11,555 | 16,956 | 5,401 |
| Compute and infrastructure | 5,558 | 9,863 | 4,305 |
| Personnel | 1,857 | 1,711 | -146 |
| Sales, marketing, and partnerships | 2,135 | 2,361 | 226 |
| Other operating and administrative | 1,031 | 3,588 | 2,557 |
| Total operating expenses | 10,581 | 17,523 | 6,942 |
| GAAP operating income or loss | 974 | -567 | -1,541 |
The Bartz settlement row is zero in both quarters. The listed expense movements sum to $6.942 billion. The company generated $5.401 billion more revenue and reported $6.942 billion more operating expense. Subtracting the latter from the former reconciles the change exactly at the published precision. The two largest increases belong to compute and infrastructure and to other operating and administrative. Those are the rows that deserve investigation before an analyst generalizes about labor or sales efficiency. Quarterly Consolidated Statements of Operations
For a small scale test, the analyst imagines a company starting with $100 of sales and $92 of operating expenses. It then grows sales to $147 and expenses to $152. Sales growth brings $47 into the new period, and $60 more expense absorbs it and cuts into the prior margin. The premise is broadly analogous to the shape of the reported quarter, with simpler invented inputs. The comparison is useful precisely because it keeps the absolute change in expense visible alongside expense as a share of sales.
Now the colleague asks if compute cost grew faster than the business. Compute and infrastructure expense increased by about 77.5 percent quarter on quarter, using $9.863 billion versus $5.558 billion. That exceeds the approximately 46.7 percent revenue growth calculated from the same periods. As a share of revenue, the disclosed expense moved from approximately 48.1 percent to 58.2 percent using the underlying amounts. The company rounds these values to 48 and 58 percent. At that point the pair can discuss the possible causes.
The personnel line moved in the opposite direction even as the company describes stronger operating demand. The quarterly trend discussion attributes much of the decrease to lower share-based compensation as certain option awards reached full vesting, with increases in cash compensation partly offsetting the decline. A lower reported personnel figure could therefore coexist with higher cash payroll. A headcount narrative built from the reported expense alone would skip the disclosure that explains the movement. Quarterly operating trends in the amended prospectus
Give the special expense its own label
Other operating and administrative expense grew from $1.031 billion to $3.588 billion. The prospectus's reconciliation identifies $1.960 billion of non-cash charitable contributions in the September quarter related to its donation matching program. They were settled in Class N shares according to the note. The filing records the charges in that administrative expense category. The movement in that category and the contribution are evidence at different levels of detail, so the contribution belongs as a component rather than an additional row added to total expenses.
The remaining difference between the administrative expense change and the disclosed contribution is $597 million using those two rounded-to-million figures. That is a residual across the rest of the category and any reporting precision. It supplies a bounded follow-up question rather than an explanation for each dollar. If an analyst wants to attribute the $597 million to a named cause such as office growth, she needs specific evidence for that cause. The source can support a residual while leaving the detailed attribution open.
The donation matching program's non-cash form matters for understanding the accounting. Issuing shares and making a bank payment are different settlement mechanisms. The company's cost presentation and number of claims on equity are both economically relevant. A financial reader can use the stated add-back to compare one adjusted operating measure and retain the expense and dilution questions in the appropriate records. This treatment does not depend on a view of the giving program itself.
Compare both operating margins using the same revenue
The adjusted measure's definition appears alongside the quarterly reconciliation. Anthropic starts from GAAP operating income or loss and adds back share-based compensation, non-cash charitable contributions, and the Bartz settlement where relevant. June adds $1.261 billion of share-based compensation to $974 million of operating income, giving $2.235 billion. September adds $1.070 billion of share-based compensation and $1.960 billion of non-cash charitable contributions to the $567 million operating loss. The resulting adjusted figure is $2.463 billion.
The adjustment total moved from $1.261 billion to $3.030 billion. Its increase of $1.769 billion, combined with a $1.541 billion decrease in GAAP operating income, produces the $228 million improvement in adjusted operating income. The reconciliation shows how the directions can diverge. Comparison with the prior quarter therefore requires all three displayed columns: the starting GAAP amount, the changes the company excludes, and the adjusted amount those exclusions produce.
The margin question uses September's $16.956 billion and June's $11.555 billion as separate denominators. June's adjusted margin is roughly 19.3 percent and September's roughly 14.5 percent, compatible with Anthropic's rounded 19 and 15 percent presentation. Adjusted operating income rose roughly 10.2 percent. The pre-adjustment margin, measured by dividing respective GAAP results by revenue, went from approximately 8.4 percent to negative 3.3 percent. The change in the dollars and the changes in the two ratios all carry different useful information about the quarter.
A colleague asks whether the adjusted measure removes the short-term compute pressure as unusual. The company's stated adjustments for this measure do not include such compute. The September margin therefore still reflects the reported impact of additional compute expenses. That gives the analyst a concrete reason to compare the two margins in the same quarter. She can read the adjusted view while seeing the operating effect of capacity procured to meet demand.
Net loss adds financial instruments, interest, and tax items below the operating result. An analyst trying to explain the economics of the September delivery window can keep those other lines accessible and focus first on the operating comparison. If the question becomes what happened to reported GAAP net loss, the quarterly income statement and separate reconciliation for adjusted net income supply the proper starting point. They answer a different question with other components of the filing.
Why the direction of a margin matters to a growth model
An incremental calculation finds that the adjusted operating result increased by $228 million against a $5.401 billion increase in revenue. The implied difference of about 4.2 cents of added adjusted operating income for each added reported revenue dollar describes these two aggregate quarters at their presented figures. It includes timing and changes in cost categories. An analyst can compare it with the level of adjusted operating margin while preserving its scope as a period-to-period result. It supplies a reason to challenge the assumption that the next incremental dollar will earn the previous quarter's average margin.
The exercise can be repeated directly off adjusted operating expense implied by the disclosure. June revenue less adjusted income equals $9.320 billion. September revenue less adjusted income equals $14.493 billion. The implied increase of $5.173 billion leaves the same $228 million after comparing the revenue increase. The two checks help catch a sign error when a GAAP loss goes through a worksheet that expects a positive expense or a negative income item. They do not introduce a new reported line item.
The compute line had a short-term component
Anthropic reports about $5.0 billion of September-quarter compute under short-term commitments in the discussion of expenses. The company says these resources currently cost considerably more than comparable compute sourced under its multi-year commitments. The total reported September compute and infrastructure figure is $9.863 billion. On those numbers the disclosed short-term portion is about half of that expense. Compute capacity has different availability dates, contract prices, commitments, and uses, so an analyst should preserve the basis of the company's description.
Subtracting the short-term amount from total compute expense gives a rough residual of about $4.9 billion for the current quarter. That calculation is not a forecast for expenses in a future quarter without those purchases. Some short-term compute can serve workloads that contribute revenue. A slower or less flexible replacement could change how much revenue can be served, the amount of unused committed capacity, and the timing of new products. The prospectus explains that the company typically procures compute well before it becomes available. A calculation that removes the expense while keeping all outcomes fixed needs a defensible counterfactual.
The colleague frames two questions for management. Which part of the recently used short-term capacity bridges a known commissioning date for longer-term capacity? Which part supplies an option against demand arriving outside the long-term allocation? A bridge might be evaluated by visible milestones in a particular capacity program. A continuing option could be valuable even when it costs more per equivalent unit. The ratio depends on the company's actual mix and terms. The statement of quarterly expense alone does not resolve it.
A miniature capacity example shows why. Imagine an application team expects to serve 100 units of work in October and has long-term capacity to cover 80. It buys twenty more at a high spot cost. In November contracted capacity rises to 110 and demand rises to 130, so it still needs twenty incremental units under the same assumptions. The new contract can improve the average cost and also leave the short-term amount unchanged. In a different demand path the purchased units can fall. Both paths require comparisons over time. Real infrastructure also has constraints on schedules, location, and which equipment can serve specific work.
At the beginning of each forecast, the analyst writes which capacity arrived and what workloads it can serve. New research compute may support later product improvements. Serving compute can support current revenue. Site availability, customer adoption, and training schedules can change the mix. The public expense total combines those economic jobs. The prospectus discloses spending at a useful level for the operating variance and leaves room for future evidence about allocation.
A budget meeting fixes the margin first
The next question comes from a different part of the room. A colleague wants a scale for how much growth would be needed to maintain the displayed September adjusted margin if capacity expenses climb. The analyst can make a transparent sensitivity from the filed expenses while naming a few simplifying choices. She starts from the implied $14.493 billion of adjusted operating expense and adds a hypothetical $2.250 billion. That increment has no attribution to a contract in the offering. It stands for a new resource becoming available in a quarter, with all other adjusted expenses temporarily held flat for the calculation.
Under those assumptions expense would total $16.743 billion. To earn a 15 percent margin, the analyst solves for the revenue where the expense represents 85 percent of the total. Dividing $16.743 billion by 0.85 gives approximately $19.698 billion. The step up from $16.956 billion is about $2.742 billion. This required growth is roughly 16.2 percent from the actual September revenue. The worksheet uses 15 percent as a round target. It would deliver a slightly different answer using the unrounded September ratio.
Suppose the commercial team instead models $18.5 billion of revenue on the same expense. The result of $1.757 billion implies about a 9.5 percent margin in the illustration. If the capacity arrived later and the added expense in that quarter were only $1.0 billion, the same $18.5 billion revenue yields $3.007 billion of adjusted income, or about 16.3 percent under the changed assumption. Both paths are simple sensitivities. The alternate timing might also alter serving capacity and revenue, so a real forecast would test where the revenue estimate came from in each path.
This framing allows the team to ask for additional disclosures without assigning a precise probability to either outcome. If equipment arrives too early for a particular set of services, it could have economic value for future products and weigh on this quarter's margin. If capacity is short, higher unit costs on alternate compute could pressure margins even as accepted demand expands. To separate those cases, the analyst would look for procurement changes, availability milestones, and commentary about how much useful service capacity was possible. The amount of expense alone cannot establish all three.
An alternative business question fixes revenue first and asks how much room remains for expenses at the target. At $18.5 billion of revenue and a 15 percent adjusted operating margin, the expense envelope is $15.725 billion. Compared with the September base that is roughly $1.232 billion of additional room. A worksheet built for that question can identify an increasing expense and a potential offset as separate rows. It stops the team assigning every modeled saving to the compute line when another operating category changes independently.
Finally the variance ledger records whether a new workload becomes available during the period or on the last day. The impact of a one-quarter operating expense may differ from an annualized amount inferred from quarter-end availability. The next period can include a full quarter of cost from a commitment reflected only partly in September. The first budget adds the costs on their assumed effective dates and recomputes the sensitivity as those dates move. This turns a conversation about a broad margin target into a smaller set of period assumptions that can be traced to facts.
A separate audit note for the table
The financial reader saves the label attached to each comparison. The quarterly results are consolidated and unaudited, in millions of US dollars. The adjusted margin is a company-defined non-GAAP measure reconciled to its nearest GAAP measure in the same filing. The margin sensitivity uses made-up future expenses and recognized revenue. A later quarter that restates a prior figure can be entered as a revised series with its own source. The original value stays attached to the November analysis.
This note catches a particular error in a spreadsheet assembled from multiple materials. A graph titled September may show a figure for nine months ending on the same date. The computations of sequential growth and incremental cost require the three-month column. If the next filing format changes, the reader checks the dates and basis again rather than binding the spreadsheet to a table position. That habit can protect a sensitive comparison even when a later quarterly result includes a new operating expense label.
What the next reporting period can settle
The pair closes its June workbook and puts the September base in a separate file. It records one reported fact per line and leaves two open issues. First, does the rate of short-term compute use decline when more committed capacity becomes available, and how does compute expense scale relative to revenue? Second, after accounting for compensation and donation-related changes, does adjusted income grow more quickly or more slowly than sales? Later quarterly reports can answer those questions with comparable definitions and periods.
The analysts also leave room for competing outcomes. If revenue continues to expand rapidly and compute expense consumes a smaller share, the next margin comparison can support an improvement in operating leverage. If more compute arrives ahead of monetized demand, expense levels can rise even when usage remains strong. If reported operating performance improves because a non-cash item changes, the adjustment bridge will show how much of the improvement sits there. A testable forecast names the observation that would make the assumed explanation less convincing.
Their data request to the team following the offering is narrow. Preserve the exact quarter-end dates, expense unit, current definitions of the adjusted measures, and the names of any categories discussed in the filing. Revisit the amounts only alongside authoritative reporting. In the meantime the September quarter is already visible in the preliminary offering documents. It tells the analysts something substantial about the economics of serving growing Claude demand even while terms of the public offering were pending at the November 18 check.
The responsible reading of the monthly news cycle begins with a dated period. Revenue, GAAP operating result, adjusted operating result, and expenses can each be compared under a consistent quarterly lens. Once those relationships are reproducible, an analyst can ask useful questions about the next capacity and pricing decisions without pretending the outcome has already been disclosed.