Put a bank’s disputed payment on the table. A customer paid once. The account shows two entries. The person who takes the call can see both, but only the payments team can tell whether the second is a temporary authorisation, a duplicate posting or a transaction the customer has forgotten. In the meantime the customer has money missing and the bank has a file that somebody must own. Imagine offering the bank a price for returning that file with a supported explanation, the necessary correction approved and a record that will stand up when the customer calls back. The procurement question is which department used to pay for all of that.
I would start there if I were trying to understand Anthropic’s prospects after its flotation. Start with one job that already has a cost, even though nobody has ever seen that cost on a single invoice. Look at who pays for the screen in front of the employee, the database behind it, the time spent looking for the second entry and the engineer who maintains the connection to the ledger. Then ask which of those expenses the buyer can actually stop paying. A machine capable of following the entire dispute through the bank acquires a claim on a much larger budget than a machine that rents out a few seconds of processing. Collecting on that claim is a commercial achievement in its own right.
The cloud provider has an unusual seat at this table. It can quote the bank for a finished service. It can collect a payment on a model developer’s behalf. It can also sell the developer the capacity on which the service runs. Those roles give it distinct interests in the same decision. The bank might move its contract to an AI company on Monday and send more business to its old cloud supplier on Tuesday, without anyone in procurement changing their mind. The second purchase takes place upstream, where the bank’s accounts payable team seldom sees it.
What arrives on the bill
Microsoft gives us a particularly clear view of the distinction. Under its Claude Consumption Unit arrangement in Foundry, model use is converted into billing units after applicable discounts and appears on the customer’s Azure invoice. The model detail is available elsewhere. Eligible consumption counts towards the customer’s Microsoft Azure Consumption Commitment. A finance officer studying the invoice can see Microsoft, while the application owner choosing the model and negotiating its price may be dealing with Anthropic. Two people in the same company can give different, perfectly sensible answers when asked who sold them the intelligence. 35
Now put yourself in that finance officer’s position near the end of the contract year. A team proposes to move its model calls to a cheaper direct agreement. The team has timed its own workflow, checked the answers and secured a price below the one it sees in the cloud console. Finance wants one more piece of information. How much of the existing commitment will expire unused if those calls move? Spending the money with a new supplier could leave the bank paying the old one as well. If another department has work ready to consume the released commitment, the decision changes again. A comparison of prices per token cannot settle a question about two departments’ capacity to use the same prepaid or committed pound.
A second complication appears when the bank crosses a border. Google documents regional and multi-region options for Claude that let a buyer choose where processing takes place. A bank can value the ability to keep processing within an approved geography while distributing work across eligible locations. To evaluate a competing offer, it needs to know whether that offer covers the same locations when a site fails. A cheaper supplier can win easily on an ordinary afternoon and prove expensive on the one afternoon that determines the contract. The analyst cutting costs and the officer signing the continuity plan are purchasing different attributes of the same call. 36
Amazon can bring a different argument to the meeting. Its Bedrock catalogue gives a buyer access to competing model families within one managed service. A bank that expects next year’s best model to come from a different lab has a reason to value this arrangement. It can keep parts of its existing operations and change what runs inside them. The lab, in turn, has a reason to earn the loyalty of the bank’s payments staff directly. A product that remembers how that bank records a reversal and which disputes its supervisors reopen is harder to replace than a generic model behind a common interface. The contest begins with a choice about where that knowledge will live. 05
Three different rulers
Amazon reported AWS sales of $45.984 billion for the quarter to September 2026 and operating income of $17.773 billion. Take the revenue over four actual quarters and AWS recorded $161.382 billion. Take the September quarter alone and multiply by four and the pace was almost $184 billion. Both figures have appeared in serious conversations about the size of the cloud. The difference between them exceeds $22 billion, simply because one looks back over the previous year and the other assumes that the September pace persists. An essay predicting a change in leadership can move the finish line a considerable distance by switching rulers mid-race. 05
Alphabet disclosed $30.709 billion in Google Cloud revenue and $11.121 billion in segment operating income for that same quarter. Its account of the period includes sales of TPU systems as a contributor to growth. Microsoft says Azure revenue exceeded $100 billion in the fiscal year ended in June. The word “exceeded” matters. So does the word “year”. Plotting a bar at exactly $100 billion beside a September-quarter number would quietly introduce an error in both magnitude and time. Microsoft Cloud encompasses other businesses as well. A credible test for a new market leader must retain all three companies’ reporting boundaries, even where that leaves an empty cell on the page. 06 07
Three clouds, three disclosures
AWS and Google Cloud disclose segment sales and operating income for the same quarter. Azure gives us a different measure.
Source data and method
Amazon’s supplemental AWS segment table and Alphabet’s segment results present amounts in USD millions. Divide by 1,000 for the displayed billions. Microsoft reports the threshold in its FY2026 annual report. The CSV includes reporting windows, scope and source URLs.
There is another reason for keeping profit on the page. A cloud operator has to fill a site through the useful life of the equipment, and a model supplier can buy capacity on a schedule that begins long before its end customer signs up. One can report rising sales while granting progressively deeper discounts to keep machines busy. The other can announce a large enterprise contract whose early years are spent integrating systems for a fee that barely covers the engineers. The bank keeps its budget in cash. Both suppliers report revenue. Their shareholders ultimately need a return on what each had to commit to earn it. Those are three separate vantage points on the same growth.
The comparison I propose for 2031 starts with four consecutive quarters of recognised sales for each business and then repeats the exercise for operating profit wherever the disclosures allow it. If a lab overtakes one cloud segment while reselling a great deal of that segment’s infrastructure, it will have passed an important test of commercial reach. We will need the second comparison to establish how much economic power travelled with the customer contract. A distributor can become very large by passing expensive inputs through its accounts. The customer’s willingness to pay a premium for what the distributor adds is a more demanding achievement.
The channel that shrank and grew
The most revealing figure in Anthropic’s prospectus is tucked into the financial statement notes. In the first nine months of 2025, roughly 48 per cent of its revenue came from sales through cloud marketplaces. A year later the corresponding share was about 44 per cent. Put those percentages beside nine-month revenues of $2.545 billion and $33.242 billion. The implied marketplace amounts rise from about $1.22 billion to $14.63 billion. The percentages are rounded and the derived amounts are estimates. These are sales of Anthropic’s service through a channel. The amount retained by the hosts of those marketplaces requires a different disclosure. 01
Consider what an executive could conclude from either half of that comparison. A sales director who sees only the percentage decline might instruct the team to arrest a deterioration in the channel. A partner manager who sees only the dollar increase might argue that the marketplace’s position is now impregnable. Both would be reaching beyond their information. A customer may have increased its purchases through its existing channel, bought a second service directly, or moved existing work between the two. The public numbers combine all three behaviours. Only matched customer histories could tell us how often one supplier took a contract away from another.
The filing describes the cloud providers as collection agents on these marketplace sales. That observation settles a question about whose revenue the filing is measuring. It leaves open the terms of each infrastructure contract and the identity of the party that will control renewal. The cloud that collected the bank’s payment need not even be the cloud that supplied all the computing behind it. For clarity the instrument below initially puts both roles in one company and prices a case at a round hundred dollars. The bank pays once. Open the instrument and one payment separates into the fee for distribution, the rent for computing and the balance available to deliver the work.
The marketplace beneath the model
Follow one $100 payment from a bank to the parties that deliver the result. Select a ring to inspect a transfer. Change the contract to see who keeps the fee and who earns the compute bill.
The ledger below records the transfer represented by each ring.
The lab holds the renewal. The cloud collects $100, passes $94 to the lab and receives $24 back for compute.
| From | To | Purpose | Paid |
|---|---|---|---|
| Bank | Cloud | Customer payment | $100 |
| Cloud | Lab | Remittance after fee | $94 |
| Lab | Cloud | Wholesale compute | $24 |
Assumptions and controls
Direct sales have no channel fee. Under the cloud prime arrangement the cloud buys the lab’s service for $65, then sells compute back to the lab. In each case the bank pays $100 once. Each ring uses 100 equal angular divisions. Compare the fraction of a full turn on each ring. The highlighted ring and ledger row identify the selected transfer. Drag to orbit. Scroll, pinch or use the buttons to zoom.
At the default settings the marketplace keeps six dollars for the channel and receives twenty-four for infrastructure. Seventy remains with the lab before its other costs. Calling that seventy profit would erase the people who installed the workflow, the support contract and the cost of a difficult case that has to be run again. Change to a direct sale and hold the underlying machine usage steady. The cloud still receives twenty-four. The bank still pays a hundred. Six dollars changes hands because the route to the customer has changed. A lab may spend more than six dollars acquiring and supporting that customer itself. Direct distribution buys more control even on occasions when it earns less money.
Now give the cloud the primary contract. In this hypothetical bid it promises the result to the bank and buys the lab’s service for sixty-five dollars. The lab in turn buys the same twenty-four dollars of computing from the cloud. Follow the transfers in sequence and the cloud ends with fifty-nine dollars, the lab with forty-one. The cloud has a larger balance and a different obligation. If the bank refuses to accept the case, it knows which counterparty it can pursue. The fee attached to taking that call is one reason the enterprise software industry has remained more durable than generations of technically superior substitutes might suggest.
What a bank has already done
There is a real bank behind a nearby, more modest proposition. Trust Bank in Singapore built a customer service application on Amazon Bedrock. The architecture described by AWS uses a containerised agent, retrieved customer and product context, and a database of material against which to ground an answer. AWS reports that more than 52 per cent of customer enquiries reached resolution without human intervention, and that some service agents were retrained for AI operations. These are results reported in the supplier’s case study. They concern customer enquiries, a wider and less specialised category than the disputed payments used in our thought experiment. 37
The interesting piece of evidence is the list of things that had to work together. Somebody had to maintain the bank’s approved information, decide when an answer counted as satisfactory and feed troublesome conversations back into the evaluation process. A resolved enquiry created demand for several technical services as well as for the model call. An incumbent with a working deployment has a chance to sell the next improvement to a buyer with whom it already shares an operating history. A new supplier asking to replace it would have to demonstrate that it could carry over the hard-won lessons embedded in that history.
Imagine interviewing the person who funded the rollout. I would ask for the number of customers who contacted the bank again about the same issue within seven days. I would ask to see overtime, outsourced support bills and the size of the backlog before and after the change. At a growing bank, keeping headcount steady can represent a considerable economic gain. Quantifying it means constructing a credible account of how many people would otherwise have been hired. I would also want a cloud invoice for the same cohort of enquiries. The reported resolution rate alone cannot tell a reader whether the resulting gain went to the customer, the bank, a model developer or Amazon.
One can learn a great deal from the customer who declines to switch. Suppose the payments team has a year of decisions linked to the ledger, and supervisors can reproduce the reasoning behind any adjustment in an inspection. The incoming bidder can read the same records and write a better explanation. Can it preserve the links when a case moves across systems? Will a supervisor still be able to recover the evidence after a model is retired? Somebody must pay to qualify the replacement against old cases that customers can still contest. As that stock of prior work grows, portability of the decision record becomes a substantive term of the next tender. A bank that secured the right to export that record in usable form has already bought itself a more competitive renewal.
How much cheaper must the work become?
A cloud operator can receive fewer dollars per case and more dollars in total. Assume a thousand accepted cases at a hundred dollars each. Forty dollars from every case reaches the cloud, giving it forty thousand dollars over the period. In a second period its receipts fall to thirty dollars per case. If the bank buys sixteen hundred completed cases, the cloud now receives forty-eight thousand dollars. Its share of the buyer’s expenditures falls ten percentage points and its receipts rise by a fifth. The volume increase is an assumption. Changing it is the quickest way to find out how much of a growth story rests on new demand.
A cloud can grow as its share falls
At $100 per accepted case, the cloud can collect less from each sale and still finish the period with more money.
$100,000 paid · $40,000 to cloud
$160,000 paid · $48,000 to cloud
At $30 per case, 1,334 accepted cases are needed to exceed Period A’s cloud receipts.
Method
Period A has 1,000 cases at $100, with $40 per case going to the cloud. For either period, cloud receipts equal cases × cloud receipts per case. Its share is receipts divided by total buyer spend. The threshold is the smallest whole number of cases that strictly beats $40,000.
Where would those extra six hundred cases come from? A bank may have a backlog of low-value discrepancies that it has historically written off because investigation costs more than recovery. Cheaper investigation can make some of them worth opening. In that event the appropriate baseline for an additional case is the loss the bank previously accepted, together with the cost of any customer redress it avoided or deferred. There might be no consultant invoice to displace. There might instead be a new service whose expenditure is justified by previously forgone recoveries. A claim that the lab won money from consulting would require finding a consultant the bank had actually intended to hire.
The opposite can happen just as easily. The number of payments may be essentially fixed over the life of a contract. A better upstream validation rule may eliminate many exceptions altogether. An AI supplier paid per resolved case has a commercial reason to expand its scope as the original queue shrinks. A cloud supplier paid by usage has a similar exposure. The bank, by contrast, might earn its largest return by fixing a reference-data error that prevents a thousand disputes from being opened in the first place. A procurement system that rewards every closed file and assigns no value to prevention can make the supplier’s dashboard look excellent while keeping the bank unnecessarily busy.
This is where I would press a team promising to remove waste. Compare its proposal with the best affordable change the bank could make without buying a new AI service. Try the validation rule, remove a duplicate data feed or renegotiate an idle reservation. Keep the required quality, capacity for busy periods and time to correct an error constant. Count the staff hours released separately from expenditure the finance office can remove. A team that saves an hour every morning has achieved something real. A budget that pays the same staff to clear an overdue queue has also achieved something real. It takes an additional piece of evidence to say that an outside company has captured an hour’s wages as revenue.
The terms of the next bid
My prediction depends on the tender that comes after the pilot. In the first year the lab may help the bank design a service at an introductory price, while the cloud provider supplies discounted capacity and a consultancy puts its best people on the project. At renewal the buyer will know the actual case mix, the probability of reopening and the cost of an unusually difficult month. It can ask bidders to quote for the same accepted work and name the party responsible for repairs. The firm that can make a credible price under those conditions has a plausible route into the bank’s operating budget.
If the bank invites only its cloud account team to that negotiation, the cloud has preserved a valuable advantage. If the payments director calls the lab first and asks it to obtain competing infrastructure prices, something significant has moved even if the bank’s cloud bill rises in absolute terms. If the bank owns the evidence and can run a fair competition between both, it may keep most of the benefit itself. The question I would put to a board is which of those three positions its contracts are creating today. By the time the next invoice arrives, the answer may already have been written into the right to take the work elsewhere.