Customer demand

Do useful tools become durable paid businesses?

Yes in important observed respects—but not at one uniform margin or for every supplier. Customer studies establish useful outcomes in selected settings. Purchasing and company disclosures establish recurring commercial activity. The latest evidence also gives a stronger favorable case that sellers can retain value, rather than assuming all gains are competed away. None of those findings establishes the return on every infrastructure project or the price of every ownership claim.

This guide connects two different questions: what makes a purchase worthwhile for the buyer, and what remains after the provider's full costs. The September 19 customer-value investigation supplies the first set of studies and buyer examples. The September 29 payment-and-capture investigation follows recurring revenue, full expenses, cash and embedded advertising outcomes. They retain separate evidence dates and methods.

For a buyer or worker, a financial disappointment need not mean useful service disappears. A buyer can switch vendors, an application can change its upstream model, or a new owner can operate existing assets at a lower cost. Those possibilities do not guarantee continuity; they explain why usefulness, vendor durability and investment return require separate evidence.

Start with the complete useful outcome

A resolved support issue, accepted software change or better advertising outcome is more informative than a token or a seat. The buyer's benefit depends on quality, integration, review and rework. Time saved can create more output or less drudgery without reducing payroll. Neither an outcome label nor a workload equivalent automatically measures avoided cash expense.

The support evidence contains a credible productivity improvement from assisting human agents. The coding evidence includes a positive multi-company field study and a negative experiment among experienced maintainers; different tasks, vintages and estimands make averaging their headlines meaningless. Office assistance can save email time without establishing an equivalent employer-profit gain. Current paid-seat disclosures do not silently update those older study results. Study designs and buyer economics.

The latest named-buyer case follows Seagate's reported platform replacement and Freddy deployment. It is stronger than an anonymous adoption claim, but the undated vendor/customer story is not an invoice audit or causal saving study. Seagate's operating resources demonstrate a possible buying budget, not the identified source of its software payment. Buyer-to-provider chain.

Ask which customer and which dollars persisted

Freshworks' company-wide recurring-revenue retention is evidence of persistence, not an AI-product renewal ledger. Reported retention declined from 106% to 104% across the June 2025–2026 comparison; constant-currency retention rose from 104% to 105%. Both are needed to interpret the change. They include expansion and contraction within the metric's customer base; neither means every customer renewed or every dollar was collected. Definition, currency comparison and revenue bridge.

The quarterly revenue reconstruction attributes growth to both existing and new customers. That supports commercial substance. Paid AI accounts and a reported Freddy ARR threshold separately support a paid-product case, but their different dates and lower bounds cannot supply an average invoice, full-year retention rate or standalone cash margin.

Earlier transaction-panel evidence also showed repeat buying and multi-provider purchasing. It did not identify all buyer funding, net out every credit or create a representative national cohort. Continued use of AI can persist while spending shifts between vendors. The relevant question is the same customer's net payment through renewal—not whether a recognizable logo appears on a customer list. Earlier purchasing evidence and its limits.

Look below gross margin, then reconcile cash

Freshworks' April–June 2026 revenue was $237.377 million. Its gross margin was about 84.8%, but after operating expenses its operating profit was $6.060 million, a 2.55% margin. That is genuine progress from a quarterly operating loss, not evidence that 85% of each subscription is distributable AI cash. The first half remained loss-making on the operating measure. Full expense account.

Stock-compensation expense fell $11.432 million, an amount equal to 77.7% of the $14.716 million quarterly operating-income improvement arithmetically. This is not a causal estimate of AI efficiency. The compensation change is already in expenses; it must not be counted again as a separate saving. A noncash expense is not economically free to shareholders.

Cash is a further test, not another name for operating profit. Freshworks' first-half operating cash less cash property/equipment and capitalized software was $107.120 million in 2026, compared with $107.440 million in 2025. Its adjusted free-cash-flow measure rose because it added back more specified costs that were actually paid. Both observations can be true. Cash reconciliation, including the signed working-capital correction.

A separate AI invoice is not the only way to earn

Meta's advertising business provides a different route. Better recommendations can improve an existing paid service. A provider-authored production experiment reports 4.3% more conversions in Facebook Feed and Reels. That supports a bounded monetization mechanism, not a 4.3% increase in worldwide advertiser profit or Meta's total cash. The publication date is January 27, 2026; the experiment's full fieldwork calendar and design details were not established for independent replication. Paper and limits.

An illustration holds that 4.3% outcome gain fixed while assuming 2% more buyer spending. Gross platform receipts rise, yet buyer cost per conversion is about 2.21% lower. This is not an observed budget change or a profit calculation: it shows why customers and providers can both gain. Extra service costs, auction responses and the value of each conversion remain separate. Shared-gain calculation.

The whole company's accounts cannot attribute all additional advertising revenue—or every cost and capital purchase—to that experiment. Meta's second-quarter operating income fell while its half-year operating income rose; cash after the stated capital payments fell in both comparisons. Real embedded monetization and a demanding investment bill can coexist. Matched company accounts.

What does this establish for the build-out?

It strengthens paid persistence and the possibility of retained provider value. It does not supply a universal provider margin or demonstrate sufficient cash for all physical investment. Application hosting costs are already in application expenses; adding the entire upstream investment bill again would overstate costs. Omitting the provider's development, distribution or replacement needs would understate them.

A protected capacity contract can keep receiving payments even when a retail application's pricing changes. A well-funded customer can honor an expensive purchase while earning a poor return. A lower future equipment price can benefit a new buyer while impairing an earlier owner's recovery. The funding guide, capacity guide and investment-return guide follow these different claims.

The favorable test is continuing net payment together with full-cost cash capture and adequate renewal economics. The adverse test is continued use with weaker retained dollars, higher cost or financing that absorbs the gain. A change in any one of these needs its own evidence; the absence of a disclosed AI margin is not evidence that the margin is zero.

The infrastructure continuation evidence now supplies a complementary favorable example: CoreWeave reports a newly signed contract for an older A100 design into 2029. Continued demand for adequate service need not require every buyer to choose the newest hardware. It is still not a same-customer renewal rate, a physical-age record or measured lifetime cash. The contract observation and its limits.

A posted cheaper node is not necessarily a cheaper completed task: quality, latency, storage, actual speed and migration determine the comparison. The supplied equal-output tariff calculation solves a required speedup, not observed performance or achieved rental yield. This strengthens the reason to measure useful service and complete costs rather than transport application retention percentages into a GPU investment model. Comparable service before comparable prices.

A real customer payment is one party’s receipt and another’s cost

The new joint-cash case follows CoreWeave’s hosting payment to Core Scientific. It does not assign any particular Microsoft, Meta or OpenAI workload to those sites. Whole-company operating cash is already net of operating payments; it is not a second pot of gross receipts available to pay the same supplier again. The connected perimeter.

The issuer’s combined tenant-credit/other-cost row supports a range for the net payment, not a complete customer budget. In 2027 it is $1.045–1.346 billion before separately borne construction, GPUs and other tenant uses. The unknown credit split is not a discount the buyer can choose. Future resources allocated to this exact purchase remain unmeasured. Payment allocation and its limits.

This preserves the favorable demand findings above while limiting what they can prove. Freshworks’ company-wide persistence and Meta’s provider-reported conversion gains are not measurements of CoreWeave’s site-specific future cash margin. A negotiated or legally earned payment reduction can help the buyer while reducing host receipts; it is not automatically a new joint saving. Genuine delivery-cost improvement would be a different route to shared gain. A worked transfer example.

Choose the depth you need

For complete study designs, buyer illustrations and the earlier pricing evidence, read Real customers, uneven capture, dated September 19. For company-wide retention, cost and cash bridges and the embedded-monetization countercase, read Paid persistence is real; the division of the gain is not uniform, dated September 29. Each investigation carries its own sources, downloads and limits.

This maintained guide is not an additional research investigation or a new observation of every underlying fact. The complete catalogue keeps all current dated and historical reports accessible. The current answer explains how the combined evidence bears on the bubble question.