TLDR: Wix’s Base44 entered 2026 at near-zero gross margin, and filers book inference on different lines, so buyers must diligence cost to serve per unit before pricing a software target’s revenue.
Base44 Shows Why AI Gross Margin Due Diligence Comes Before the Price
An artificial intelligence (AI) product can grow its revenue at full speed while its gross margin stays close to zero. That gap is what AI gross margin due diligence exists to measure, and a public filer has now put it on the record. Wix.com, the listed website-building group, said in its second-quarter 2026 results that it expects its AI app builder Base44 to reach “~60% non-GAAP gross margin in 2H, a significant improvement from near-zero non-GAAP gross margin entering the year”1. Non-GAAP means adjusted away from US generally accepted accounting principles (GAAP); 2H is the second half of the year. The words “near-zero” are Wix’s own, and the release quantifies only the second-half target. The revenue line of an AI-native target can look like software while its cost line looks like a utility bill.
The acquisition terms make the case sharper. Wix’s annual report on Form 20-F for 2025 records that on 13 June 2025 it acquired 100% of Base44 for total purchase consideration of $92.158 million, made up of $18.058 million in cash and contingent consideration with an acquisition-date fair value of $74.1 million2. The contingent part is an earnout: the founder is eligible for further payments “based on certain revenue-related metrics for the calendar years 2025 through 2028.” Revenue drives the payout: Wix’s description of the formula names revenue-related metrics only.
The earnout liability then grew. The same filing puts its fair value at $74 million at the acquisition date and $89.5 million at 31 December 2025, estimated with a Monte Carlo model built on “projected revenue-related metrics and revenue volatility”2. In roughly six and a half months the liability rose by about $15.4 million, during the very period in which, by Wix’s own later description, the product carried almost no gross margin. In its first-quarter 2026 results, Wix reported that Base44 had reached about $150 million of annual recurring revenue (ARR), in its words “~$150 million of ARR as of May”3.
The earnout may well have suited Wix’s own strategy. Wix bought distribution, a product and a founder, and it fixed the cost base later with its own model, Base 1; the $15.4 million rise in the liability shows the seller’s side of that bet improving while the margin sat near zero. The lesson for other buyers sits in the mechanism. A revenue earnout pays for revenue whether or not that revenue leaves anything behind once the compute bill is settled, so the buyer carries the full cost-to-serve risk while the seller is paid on the top line.
Wix’s Segment Margin Fell Five Points While Its Core Product Held Steady
The parent’s reported numbers showed the pressure before the Base44 margin was named. Wix reports two segments, and its 20-F defines Creative Subscriptions revenue to include “other subscription based offerings, such as Base44”2. The same filing explains that the segment’s cost of revenue “increased during 2025, mainly due to hosting and revenue related costs (including AI costs)”2.
The reconciliation tables in the quarterly releases then give the movement. In the first quarter of 2026, Creative Subscriptions non-GAAP gross margin was 80% against 84% a year earlier, and total non-GAAP gross margin was 66% against 69%3. In the second quarter, Creative Subscriptions stood at 80% against 85%, and the total at 67% against 70%1. Wix’s chief financial officer (CFO), Lior Shemesh, said in May that “Core Wix Creative Subscriptions gross margin was stable as AI costs remained minimal”3. Read together, the statements locate the decline: the core product held its margin, the segment that contains Base44 lost four to five points, and Wix then disclosed the AI product’s near-zero starting margin.
Wix also named the cause in plain terms. It described its in-house large language model, Base 1, as giving it “greater direct control over the largest cost driver in AI-native businesses, compute and inference spend”1. Shemesh called its deployment “a turning point in lowering our AI inference and compute costs”, adding that “we expect our AI costs to decrease significantly going forward”1. Wix’s outlook translates the change into about “two points of total non-GAAP gross margin improvement in 2H vs. 1H for the consolidated business”1. One AI product, inside a group with more than $560 million of quarterly revenue, was large enough to move the consolidated margin by two points.
The same pattern appears at the largest scale. Microsoft’s annual report on Form 10-K for the year to 30 June 2026 states that “Microsoft Cloud gross margin percentage decreased to 66% driven by continued investments in AI infrastructure and growing AI product usage”4, offset in part by efficiency gains, while Microsoft Cloud revenue rose 27% to $214.4 billion. The figure covers Microsoft Cloud alone. Usage growth and margin compression arriving together is a disclosed fact at both ends of the size range, from a newly acquired app builder to the largest cloud business on the market.
Five Filers Book the Same Inference Cost on Different Lines
A buyer comparing two AI-native targets on gross margin assumes both put the cost of serving customers in the same place. The filings show three different placements. Inference, the compute consumed each time a model answers a request, appears in cost of revenue at some companies, in research and development (R&D) at others, and in both at a third group.
Doximity, the physician platform, writes in its quarterly report on Form 10-Q that cost of revenue includes “software costs, including generative AI platform usage and inference costs”5, and that R&D “also includes” software-related costs, “including generative AI platform usage and inference costs”. ServiceTitan’s 10-Q states that R&D expense “also includes fees to third-party product development resources, infrastructure and server costs, inference costs, and allocated overhead costs”6, while its description of cost of platform revenue lists “platform infrastructure and server costs” and leaves inference unnamed. Duolingo describes cost of revenues as consisting predominantly of “third-party payment processing fees charged by various distribution channels in addition to hosting fees and third-party AI costs”7, while its R&D also rose on “increased software and third-party AI costs of $3.3 million” in the quarter.
| Filer and filing | Where AI cost sits | Margin signal |
|---|---|---|
| Doximity, 10-Q, Jun 2026 | Cost of revenue and R&D | AI cost split across two lines |
| ServiceTitan, 10-Q, Jul 2026 | Inference named in R&D | Inference unnamed in cost of revenue |
| Duolingo, 10-Q, Jun 2026 | Cost of revenues and R&D | 72.4% to 72.6% on lower unit AI cost |
| Wix, 20-F 2025 and 6-K Q2 2026 | Cost of Creative Subscriptions revenue | Segment 85% to 80%; Base44 near zero |
| Microsoft, 10-K, FY Jun 2026 | AI infrastructure in Cloud gross margin | Microsoft Cloud down to 66% |
Sources: Doximity, ServiceTitan and Duolingo quarterly reports on Form 10-Q; Wix annual report on Form 20-F for 2025 and second-quarter 2026 results on Form 6-K; Microsoft annual report on Form 10-K for fiscal 2026. Wix segment margins are non-GAAP, second quarter 2026 against 2025. See References.
Each choice is defensible and follows the filer’s own judgement about whether a cost serves customers today or builds the product for tomorrow, and the same model call can do either. The consequence for a buyer is mechanical. Two targets with identical economics can report gross margins fifteen points apart because one books production inference in cost of revenue and the other books most of it in R&D. A valuation keyed to gross margin, or to a revenue multiple chosen because gross margin looks like software, is then keyed partly to an accounting policy.
An illustrative case shows the size of the effect. Take a hypothetical target with $40 million of revenue reporting a 78% gross margin, so $31.2 million of gross profit, with $6 million of production inference sitting in R&D. Moving that inference into cost of revenue gives $25.2 million of gross profit and a 63% margin. At an unchanged multiple of gross profit, the restatement removes about 19% of the value before any other diligence finding. Finding the $6 million inside an R&D ledger takes the target’s cost-centre mapping, which is the first item Neumarz requests in test 1 below.
The deploying customer sees the other side of the same bill. Neumarz has set out how falling token prices still broke three in four AI budgets at banks and asset managers8. Every token a customer consumes is also a cost the vendor serves, and on the vendor’s side it lands in whichever line its accounting policy chooses.
AI Gross Margins Move Faster Than a Single Period Can Show
Gross margin in a mature software company changes slowly, so one audited year is a fair proxy for the run-rate. AI cost lines move faster, in both directions, and a single period can mislead either way.
Duolingo shows the favourable direction. Its total gross margin “increased to 72.6% from 72.4%”7 for the quarter to 30 June 2026, and to 72.8% from 71.8% for the six months, which the company attributes to higher subscription gross margin “reflecting continued reductions in per-unit third-party AI costs”. Wix shows the same direction at a steeper slope, planning to take Base44 from near zero to about 60% within a year by moving the work onto its own model1.
The risk factors show the unfavourable direction. Asana’s 10-Q warns that higher demand for third-party AI services may raise their price, and that “these price increases are difficult to forecast, may be significant, and may lead to increased expenses and lower margins”9. Asana adds that offering AI features “at discounted or no cost to customers” to drive adoption “may contribute to increased third-party AI costs without corresponding revenue increases.” DocuSign’s 10-Q states that ongoing processing and inference costs may reduce profitability, and that customer expectations of speed and performance “could require us to incur higher infrastructure and model costs that we may not be able to recover through pricing”10.
Pricing is the third lever. GitLab, in its fourth-quarter fiscal 2026 results, introduced GitLab Credits, “a usage-based pricing model for GitLab Duo Agent Platform capabilities that provides transparency into AI costs”11. A move to usage pricing ties the revenue line to the cost line, which changes the margin profile a buyer inherits. Read next to Duolingo’s falling per-unit AI costs7, the GitLab change shows the two routes a target can take to protect gross margin: cut the cost of each unit served, or bill each unit consumed. Diligence asks which route the target has taken and how much of the margin plan rests on it.
For diligence, the point is the curve and its drivers. A target at 45% gross margin with a working in-house model, routing to cheaper models and usage-based pricing can be worth more than a target at 70% that depends on one supplier’s price list and gives its AI features away. A single reported year captures one trajectory at best, so the buyer needs the levers and the evidence that each one has been built rather than planned.
The Neumarz Cost-to-Serve Workstream Runs Five Tests Before the Multiple
Standard financial diligence on a software target checks revenue quality, churn and the bridge to earnings before interest, taxes, depreciation and amortisation (EBITDA). For an AI-native target, Neumarz runs cost to serve as a separate workstream with its own data request, because the questions it answers sit upstream of the multiple. Five tests make up the workstream.
1. Map the line. List every inference, hosting, third-party model and vector-storage cost, and record where each sits: cost of revenue, R&D, or sales and marketing. Then restate gross margin with all production serving costs in cost of revenue. The restated figure is the one the multiple should rest on, and the gap between reported and restated margin becomes a quality-of-earnings (QoE) adjustment. Cloud credits granted by a model or infrastructure provider belong in the same review, because a cost base subsidised by credits that expire after closing is a cost base the buyer is still to see.
2. Measure unit cost. Ask for cost to serve per paying customer and per completed task, by cohort and by pricing tier, for at least the last four quarters. A blended margin hides the heavy users who consume the compute; the tier and cohort cut shows whether new customers cost more or less to serve than old ones, and whether the top tier is priced above its own cost.
3. Size the free load. Measure the cost of AI given away through free tiers, trials and bundled features as a share of total serving cost. Asana’s risk factor names the exposure precisely9. A target that has grown on free AI usage carries a cost line that its paying revenue has to absorb once the promotion ends.
4. Test the stop controls. Ask whether the target can cap or halt runaway usage in real time, per customer and per workload, and see it demonstrated. Weak stop controls are documented among enterprises that run agents. A self-selected VentureBeat Pulse survey of 107 organisations, fielded in July 2026 and published on 12 August, found that one in five enterprises (21%) “has no real-time, programmatic way to stop an agent before a budget-breaking bill arrives”12; they learn of the spend from logs afterwards. The respondents are enterprise teams running agents, 53% of them in technology and software, and the survey measures their own spending controls, with vendors’ accounts outside its scope. The test asks whether the target, which serves such customers, can meter and stop its own compute before the invoice lands.
5. Price the lever dependency. Establish how much of the forecast margin path rests on one supplier’s price, on a model the target has yet to build, or on a pricing change customers have yet to accept. Wix’s Base 1 shows the lever can work1. Diligence asks whether the target’s equivalent exists, what it costs to run, and what the margin looks like if it arrives a year late. Take-or-pay compute commitments and reserved capacity belong here too, since they fix part of the cost base whatever usage does.
The five tests produce three numbers a buyer can price: a restated gross margin, a unit-cost trend, and a sensitivity of margin to the one or two levers the plan relies on. Each test also produces a document that the share purchase agreement can refer to.
Deal Terms Should Key Earnouts and Warranties to Gross Profit
The workstream pays off only if its findings reach the contract. Four drafting choices carry them there, and each follows from the filings above.
Key the earnout to gross profit. Base44 is the worked example of a revenue earnout on revenue that, by its buyer’s account, entered 2026 at near-zero gross margin1,2. An earnout measured on gross profit, or on revenue subject to a gross-margin floor, pays the seller for the economics the buyer is acquiring. The definition matters as much as the metric: the agreement should state which costs count as cost of revenue for earnout purposes, so that a later reclassification of inference into R&D cannot inflate the payout.
Adjust the quality of earnings for reclassified inference. Where test 1 moves serving costs into cost of revenue, the restated gross profit becomes the basis for the price and for any locked-box or completion-accounts mechanism, and the adjustment is recorded in the QoE report so both sides price from the same figure.
Warrant the cost-allocation policy. A specific warranty that the target’s allocation of AI, inference and hosting costs across income-statement lines has been applied consistently over the diligence period, as disclosed, gives the buyer a remedy if the policy changed to flatter the margin before the sale.
Make cost-to-serve data a closing deliverable. The unit-cost tables, supplier contracts, credit balances and usage-control evidence from tests 2 to 5 should be delivered and updated at closing, so the buyer starts ownership with the cost base measured rather than estimated.
Sellers gain from the same discipline. A founder who arrives with a restated margin, a unit-cost trend and a demonstrated stop control removes the largest discount a buyer would otherwise apply to AI-native revenue. In the illustrative case above, that preparation is the difference between negotiating over $31.2 million of reported gross profit and defending $25.2 million after the buyer finds the $6 million itself, and it lets the seller propose an earnout on gross profit from a position of evidence.
Neumarz, a Kainjoo SA venture, advises acquirers, investors, family offices and founders on transactions involving AI-native companies, including cost-to-serve diligence and the earnout and warranty terms that follow from it. Parties assessing an AI-native target can request a confidential cost-to-serve review of a target.
This article has been reviewed by Haider Alleg, Managing Partner at Kainjoo SA (Terraxis).
This article is general commentary prepared for professional and qualified contacts. It is general information within the meaning of the Swiss Financial Services Act (FinSA), separate from any offer, solicitation, investment advice or personal recommendation, and advice on individual circumstances remains a matter for a dedicated mandate. Company figures are taken from public filings and are not a view on any security. The cost-to-serve workstream and the worked example are illustrative. Mandate-specific information is shared under a non-disclosure agreement (NDA) only.
References
- Wix.com Ltd. “Wix.com Ltd. Reports Second Quarter 2026 Results,” Form 6-K, Exhibit 99.1, 4 August 2026. https://www.sec.gov/Archives/edgar/data/1576789/000162828026052108/secondquarter2026results.htm
- Wix.com Ltd. Annual Report on Form 20-F for the fiscal year ended 31 December 2025, filed 5 March 2026, Note 3 (Business Combinations) and Item 5. https://www.sec.gov/Archives/edgar/data/1576789/000162828026015222/wix-20251231.htm
- Wix.com Ltd. “Wix.com Ltd. Reports First Quarter 2026 Results,” Form 6-K, Exhibit 99.1, 13 May 2026. https://www.sec.gov/Archives/edgar/data/1576789/000162828026034370/firstquarter2026results.htm
- Microsoft Corporation. Annual Report on Form 10-K for the fiscal year ended 30 June 2026, filed 29 July 2026. https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm
- Doximity, Inc. Quarterly Report on Form 10-Q for the quarter ended 30 June 2026, filed 6 August 2026. https://www.sec.gov/Archives/edgar/data/1516513/000151651326000040/docs-20260630.htm
- ServiceTitan, Inc. Quarterly Report on Form 10-Q for the quarter ended 31 July 2026, filed 8 September 2026. https://www.sec.gov/Archives/edgar/data/1638826/000163882626000094/ttan-20260731.htm
- Duolingo, Inc. Quarterly Report on Form 10-Q for the quarter ended 30 June 2026, filed 6 August 2026. https://www.sec.gov/Archives/edgar/data/1562088/000162828026053603/duol-20260630.htm
- Neumarz. “Token Prices Fell 67%. Three in Four AI Budgets Broke Anyway,” 19 July 2026. https://neumarz.com/ai-token-budgets-banks-asset-managers-hedge-funds/
- Asana, Inc. Quarterly Report on Form 10-Q for the quarter ended 31 July 2026, filed 3 September 2026. https://www.sec.gov/Archives/edgar/data/1477720/000147772026000060/asan-20260731.htm
- DocuSign, Inc. Quarterly Report on Form 10-Q for the quarter ended 31 July 2026, filed 4 September 2026. https://www.sec.gov/Archives/edgar/data/1261333/000126133326000099/docu-20260731.htm
- GitLab Inc. “GitLab Reports Fourth Quarter and Fiscal Year 2026 Financial Results,” Form 8-K, Exhibit 99.1, 3 March 2026. https://www.sec.gov/Archives/edgar/data/1653482/000162828026013795/gitlab-ex99120260131fy26.htm
- VentureBeat (VB Staff), VB Pulse Research. “Agentic orchestration: Enterprise AI organizations know how to govern agents but still can’t meter what they cost,” 12 August 2026 (n=107, organisations with 100 or more employees, self-selected, July 2026 wave). https://venturebeat.com/resources/agentic-orchestration-enterprise-ai-organizations-know-how-to-govern-agents-but-still-cant-meter-what-they-cost
