The Winner's Share: Who AI Actually Pays in European Software

Larix Research · Framework · This note extends the argument of The Loser's Game (11 July 2026). The businesses named below are large and well-covered; they are illustrations, not coverage. No rating, no target. Disclosures at the end.

The other half of the trade

The previous note made a defensive argument. In a loser's game, the disciplined player keeps the ball in play: adopt AI ruthlessly in how you build software, protect the proprietary substrate the interface layer will be forced to pay to reach, and do not bet the franchise on a moonshot AI product. The companies that follow that discipline survive. But survival is an odd destination for a business sitting on a scarce asset whose scarcity is increasing. If the framework is right — if AI commoditises interfaces and concentrates value onto what the interfaces must query — then the owners of the substrate are not merely surviving the transition. They are on the receiving end of it.

The distinction matters because the market has priced the two identically. Between January and July 2026, the European software complex traded as a single AI-impaired asset class: SAP had lost roughly $150 billion of market value from its 2025 peak by late January, and on 29 January alone the shares fell as much as 17% intraday — their steepest fall since October 2020 — erasing more than €40 billion in a single session; RELX fell 14% and Wolters Kluwer 13% in a single session on 3 February, when Anthropic's legal plug-ins triggered what one wire called a rout in "AI losers"; Dassault Systèmes suffered the worst day in its history on 11 February, down more than 21%. The software sector's twelve-month drawdown is, in J.P. Morgan's formulation, the largest non-recessionary drawdown in the category in over thirty years: roughly 34% peak to trough, about $2 trillion of market capitalisation erased. Barclays captured the mechanics in February: "sell first, think later." The selling was categorical. The assets are not.

This note is about the capture side of the framework, run on European names: what it means, mechanically, for AI to pay an incumbent, and where the evidence — in reported order books and adoption figures, not in management presentations — is already visible.

Three channels through which AI pays the incumbent

Strip the marketing language away and an incumbent software or data business benefits from AI through exactly three channels. Each is independently testable in reported numbers, which is what separates this framework from a vibes-based "AI winner" list.

The cost channel. Software businesses buy capability — their own developers' — and AI deflates the price of precisely that input. A company whose product is a proprietary asset and whose principal cost is building and maintaining the software around it is long the exact trade the market shorted: what it sells is untouched, what it buys gets cheaper every quarter. The tell is in disclosed internal figures, which we return to below — SAP claims up to 30% developer productivity uplift and is targeting a €2bn efficiency run-rate by 2028; RELX reports that more than three quarters of its AI coding interactions are now agentic.

The demand channel. Agents are new customers of the substrate. An assistant that replaces a dashboard still has to source what the dashboard displayed — and if that something is licensed, contractual, or liability-bearing, the agent is metered like any other client. The tell is whether the incumbent has built the toll booth: SAP sells premium AI through "AI Units," a consumption currency priced per user and per record; the exchange and index operators license the number itself regardless of which interface displays it; Wolters Kluwer signed an expanded enterprise collaboration with OpenAI — the assistant layer contracting with the corpus rather than around it.

The authority channel. Generative AI makes plausible content free, which raises — not lowers — the value of content that is authoritative, graded, current, and citable under liability. A hospital does not buy access to medical text; it buys accountability. A law firm does not buy case-law retrieval; it buys a corpus it can rely on in front of a court. As ambient synthetic content rises, the premium on the verified corpus rises with it. The tell is adoption and pricing of the AI products built on the corpus — Lexis+ with Protégé driving double-digit growth in RELX's legal division, UpToDate Expert AI signed by more than half of Wolters Kluwer's US enterprise hospital customers within two quarters of launch.

A business qualifies as an AI beneficiary when all three channels run in its favour simultaneously. Most software does not qualify — the capability sellers of our TECMA initiation are the counter-example, and Europe has plenty of them. But a specific class of European incumbent does qualify, and it is a larger share of the European listed software complex than of the American one. That composition fact is the payload of this note.

The canonical case: SAP

No name has been treated as a more canonical AI loser, and no name fits the three channels more completely.

Start with what the market did. SAP briefly became Europe's largest company by market capitalisation in March 2025; from that peak the shares slid through the year, took their steepest one-day fall since 2020 in January 2026, and by late June were down 37% for 2026 alone. The bear case is the standard one: agents will execute business processes directly, the system of record becomes a commodity database the agents route around, and the seat-based ERP claim on corporate budgets evaporates. The January 29 print — in which current cloud backlog growth merely decelerated while two-thirds of cloud order entry contained SAP Business AI — was taken as confirmation of the bear case.

Now look at the asset. SAP's own longstanding formulation is that its customers generate 87% of total global commerce. However one discounts a marketing statistic, the structural claim underneath it is checkable and true: for most of the world's large companies, the general ledger, the order book, the inventory record, and the payroll are SAP tables. The AI bear case concedes the interface and misses the point. An agent that is going to execute a business process — place the order, release the payment, post the journal, move the inventory — needs more than read access to the system of record. It needs write access, and write access to a regulated, audited, permissioned system is granted by the layer that owns the authorisation model. SAP owns that layer. The disintermediation story assumes agents route around the ERP; the engineering reality is that agentic execution of enterprise processes runs through it, on terms the system of record sets, because the alternative is an agent writing unaudited entries into books that a CFO must certify and an auditor must sign.

That is the demand channel, and it is already in the order book rather than in a slide deck. More than two-thirds of SAP's cloud order entry in Q4 2025 contained Business AI, up more than twenty percentage points on the prior quarter. The monetisation mechanism is explicit: base AI is bundled into cloud subscriptions, and premium AI is sold through AI Units — a consumption currency, purchased annually, priced per user per month and per record processed. SAP has, in other words, already built the toll booth the previous note argued substrate owners should build. The Business Data Cloud announced in February 2025, with Databricks natively embedded, is the same move at the data layer: SAP's transactional corpus, packaged with its business semantics intact, as the substrate enterprise AI trains and grounds on. Databricks earmarked $250m to support deployments. The asset is being metered.

The cost channel is equally disclosed. SAP's Q2 2026 presentation — published this week — quantifies up to 30% developer productivity uplift from AI tooling and a €2bn efficiency run-rate targeted by the end of 2028. For a company with more than 110,000 employees whose largest controllable cost is engineering, AI is deflationary to the input and accretive to the output. That is the winning side of the trade, stated in the company's own numbers.

And the operating business did not, at any point during the derating, stop compounding. FY2025: revenue €36.8bn, up 8%; cloud revenue €21.0bn, up 23%; non-IFRS operating profit €10.4bn, up 28%, at a 28.3% margin; free cash flow €8.2bn, up 95%; a €10bn buyback announced alongside the results. The Q2 2026 print, reported this week: current cloud backlog €22.9bn, up 27%; cloud revenue up 22%; Cloud ERP Suite up 25%; full-year cloud guidance of €25.8–26.2bn reiterated. None of this is a business being disintermediated. It is a business being repriced for a scenario its own order entry contradicts — after a drawdown that took more than a third off the share price in six months, on 86% more-predictable revenue.

The honest bear case on SAP is not that AI bypasses it; it is the migration arithmetic — whether the remaining on-premise installed base converts to cloud at the pace the backlog requires, and whether the deceleration the company itself guided to is benign. That is a normal execution risk, priced as an existential one. The misclassification is the opportunity.

The corpus owners: RELX and Wolters Kluwer

The cleanest mispricing event of the whole episode happened in a single morning. On 3 February 2026 Anthropic released legal plug-ins for its assistant, and the market's response was instantaneous: Thomson Reuters fell 16% — intraday the loss approached 18%, its worst session on record — RELX fell 14%, Wolters Kluwer 13%. The logic, such as it was, held that a frontier lab shipping a plug-in had just competed with two companies whose products are graded, maintained, liability-bearing corpora of law and medicine. Deutsche Bank's response the next day remains the sharpest sentence written on the episode: RELX is "a mispriced AI beneficiary… proprietary deep data/intelligence names such as RELX should see their value increase as AI interface tools launch and look to better data." The plug-in is an interface. It has to query something.

RELX's own numbers are the authority channel rendered in sterling. FY2025 revenue of £9,590m grew 7% underlying, adjusted operating profit of £3,342m grew 9%, and margin reached 34.8%. The Legal division — the one the plug-in was supposed to kill — grew revenue 9% and profit 12%, and the April 2026 AGM statement attributes continuing double-digit growth in law-firm and corporate legal to "the adoption and expansion of our core AI-enabled legal research and analytics platform with its integrated agentic legal assistant (Lexis+ with Protégé)." Management has described AI as "a key driver of our business for well over a decade," which is not a slogan: RELX's entire business mix shift toward analytics and decision tools is a decade-long repositioning onto the substrate. Internally, more than three quarters of its AI coding interactions are agentic — the cost channel, again, disclosed. The market's response to all this was to derate the shares from roughly 31 times forward earnings at the 2025 peak to about 17 times 2027 earnings by April 2026, at which point Quilter Cheviot published a note titled, in effect, that the derating is overdone. We agree, and would only add the framework: what was repriced was a classification error, and the classification error persists.

Wolters Kluwer we covered in the previous note and will not re-litigate; the numbers since have done the arguing. FY2025: revenue €6,125m, up 6% organic, 83% recurring, margin up 40 basis points to 27.5%. Nearly 70% of digital revenues now come from AI-powered solutions, per the CEO's own formulation in the annual report. UpToDate Expert AI launched commercially in October 2025; by the Q1 2026 trading update, more than half of US enterprise customers — the hospital systems whose clinicians carry the liability — had signed to adopt it. In June the company expanded its enterprise collaboration with OpenAI: the assistant layer contracting with the corpus, on the corpus owner's terms. The February panic priced Wolters Kluwer as a victim of the assistant layer. The June contract shows it as a supplier to that layer.

The smaller systems of record: Temenos, Sage, Dassault

The SAP logic generalises downward in size, and Europe happens to own the category.

Temenos is the system of record for banking: over 950 core-banking clients in more than 150 countries run their ledgers on it. Be precise about what that does and does not mean, because the obvious objection — the ledger's contents are the bank's, not the vendor's — is correct. Temenos owns no data moat, and the argument here does not require one. What it owns, as at SAP, is the write path: a core banking system is the most audited, most permissioned, most operationally conservative software in any enterprise, and the agent that wants to post to it will do so through it, because the alternative is an agent writing unaudited entries into books a regulator supervises. FY2025 ARR grew 12% to $860m at a 34.7% EBIT margin; the company raised its FY2028 targets in February. The Q2 2026 print this week showed deal slippage — explicitly, per the release, slippage rather than losses, with the slipped deals including one of the largest banks in Central and Eastern Europe signed in the first three weeks of Q3 — and the market's habit of reading any SaaS softness as AI impairment remains the standing error. Temenos is shipping AI the disciplined way: copilots and agents inside the core (the FCM sanctions-screening agent, the Intelligent Core launched at its 2026 conference), and AI inside its own software factory.

Sage is the marginal case in this note, and it is worth being precise about which threat applies, because the obvious defence answers the wrong attack. The wrong attack is the in-house-build story — no 20-person business is going to vibe-code its own ledger. The real one is competition from below: AI collapses the vendor's cost of producing accounting software, and Sage gets outcompeted by cheaper AI-native entrants rather than replaced by its customers' own efforts. That threat has to be taken seriously, because SMB accounting is the least defensible product named anywhere in this note: the workflow is standardised, the switching event arrives at natural breaks — business formation, year-end, a change of accountant — and the customer base churns mechanically as small businesses die. What Sage owns, if the framework applies to it at all, is not the software but three adjacent assets: the accountant channel, the recommendation layer through which European small businesses actually choose their software, which an entrant must replicate firm by firm; the compliance engine, payroll and tax rules maintained and liability-bearing across jurisdictions, which sits closer to Wolters Kluwer's corpus than to application code; and the ledger itself as system of record, migrated only at real cost inside a live relationship. The evidence so far is genuinely two-sided: at £2,513m of FY2025 revenue, 97% of it recurring, renewal rate by value of 101% says the channel is holding; ARR up 11% with Sage Copilot scaled across the core products says the demand channel is open. But the marker that settles it is churn at the entry tier, not the average — if AI-native ledgers start taking share at business formation, the franchise unwinds from the bottom and the consolidated figures will be the last place it shows. The sell-side lists Sage among the potential AI winners; our own read is narrower. Sage is the live test of where the framework's boundary lies — the one name in this note where the risk that matters is being outcompeted, not replaced.

Dassault Systèmes is the most interesting of the three because its moat is neither contractual nor regulatory but physical. The virtual twin is a simulation of a product, a factory, or a human heart, built on physics engines and decades of accumulated industrial data; its 3D UNIV+RSES strategy and the NVIDIA partnership announced this year position it as the environment in which industrial AI models are trained on customers' protected IP. An assistant can write you an essay about an aircraft wing; it cannot certify one. FY2025 revenue grew 4% to €6.24bn with subscription up 11% and 3DEXPERIENCE software up 10%, at a 32% operating margin — and the stock had the worst day in its history in February. The classification error again: a company whose product is the verified physics of the physical world, priced as if a chatbot were a substitute.

The rails: exchanges, marketplaces, distribution

The exchange test from the previous note needs no new argument, only new numbers. Deutsche Börse reported record FY2025 results in February — net revenue up 9% to €5.2bn, EBITDA up 14% — with the ISS STOXX index-and-data business sitting inside it at 93% recurring revenue, a 44% adjusted EBITDA margin, roughly 90% share in European structured products referenced to STOXX and DAX indices, and north of $150 trillion of assets benchmarked to its indices. Whoever builds the agent that answers "where did the DAX close," the number is licensed from one place. An exchange does not fear a new client interface. It meters it.

Amadeus is the distribution-rail version. Its FY2025 — revenue €6,517m, up 6%, at a 29% adjusted EBIT margin, with €1.4bn of R&D spend equal to 22% of revenue — arrived with a CEO statement that is the framework's claim in management language: "as the neutral and embedded execution layer at the heart of the industry, we are uniquely positioned to orchestrate the AI-enabled travel ecosystem… AI reinforces and augments the Amadeus platform." The travel agent that is an actual agent still has to file the booking through the rail that airlines and hotels are wired into. Amadeus spent February buying SkyLink, an AI corporate-travel startup — the incumbent buying the interface to sit on top of its own rail, which is the correct direction of acquisition, the reverse of the disruption story.

And the two marketplaces named in the previous note — Scout24, with group EBITDA margin at 60.1% and rising on AI-integrated product discipline, and SMG, whose liquidity pool no assistant can conjure — are the small-cap end of the same trade. The framework is what transfers; the point of the large names in this note is that they make the transfer legible.

The European inversion

Step back from the names and the composition fact becomes visible. America's listed software complex is dominated by capability sellers — application SaaS built in the last fifteen years on seat pricing, exactly the asset class AI deflates, and its derating, however overshot, has a rational core. Europe's complex is different. Europe missed the application-SaaS wave almost entirely; what it has instead, by accident of industrial history, is a collection of substrate businesses that predate it: the German ERP system of record (SAP), the Anglo-Dutch information corpora (RELX, Wolters Kluwer), the exchanges and index houses (Deutsche Börse, Euronext, the LSEG data franchise), the travel rail (Amadeus), the banking core (Temenos), the physics of the industrial world (Dassault), and a long tail of national-champion marketplaces (Scout24, Auto Trader, Vend — all three on Barclays' February list of derated names to overweight, alongside RELX, Wolters Kluwer, Experian and Adecco — plus our own SMG).

The European tech sector was, in short, positioned for precisely the transition the market feared would kill it. It owns almost no frontier-model complex — a fact lamented in every industrial-policy speech of the decade — and it turns out not to need one. The frontier models are the commoditising layer; the value migrates to what they must query, transact with, and ground on, and that is what Europe happens to own. BofA's formulation after the February rout — "highly regulated, complex, and low-churn industries remain difficult to disrupt… innovative software incumbents are best positioned to build high-value AI agents by leveraging their proprietary datasets," naming RELX, Amadeus, SAP and Sage among the potential winners — is the sell-side arriving, five months and several hundred basis points of derating later, at the conclusion the framework implies.

The inversion has a second-order consequence worth stating plainly. For two decades the standard complaint about European technology was the absence of the winner's-game winners — no hyperscalers, no platform giants, no frontier labs. In the game software is now actually playing, that absence is the exposure you want: Europe is short the asset being deflated and long the assets being concentrated upon. The 2026 derating priced the continent's software champions as if they were American application SaaS. They are not, and the spread between what they are and what they are priced as is the analytical opportunity.

What would prove this wrong

Frameworks that cannot fail are marketing. Three developments would break this one, and all three are observable rather than speculative.

Write-access bypass. The entire SAP-and-Temenos argument rests on agents transacting through the system of record rather than around it. If agentic frameworks emerge that execute business processes in a parallel layer — keeping their own state, syncing to the ERP only as a dumb store — the authorisation-layer moat thins from both directions. We do not see this in enterprise architectures today; the audit and liability incentives run against it. But it is the single load-bearing assumption of the note, and it deserves the monitoring rather than the rhetoric.

Monetisation slippage. The demand channel assumes the toll booths price at scarcity value. SAP's AI Units, RELX's Protégé tiers, Wolters Kluwer's Expert AI contracts are all young; if competitive pressure from genuinely free ambient AI compresses their pricing faster than volumes ramp, the channel exists but pays less than the framework implies. The observable markers: the AI share of SAP's cloud order entry (two-thirds and rising, as of Q4 2025), RELX Legal growth (double digits, as of the April 2026 AGM), UpToDate Expert AI enterprise adoption (over half, as of May 2026). If any of these stalls while the narratives continue, the story is ahead of the revenue.

The TECMA problem. None of this rescues the capability sellers. Europe's small-cap software tail is full of businesses that sell exactly what AI deflates — configurators, marketing tooling, feature software rented by the seat to customers who can now build it — and our TECMA initiation is the standing reminder that the framework is a classification instrument, not a blanket endorsement of the column labelled software. The boundary case inside this note is Sage, where the threat is competition from below rather than replacement from within; the marker there is entry-tier churn at business formation, and if it turns, the classification of Sage changes with it. The inversion argument applies to the substrate owners. Applied to anything else, it is a value trap with better branding.

The discipline, then, is the same one the previous note ended on, stated now from the capturing side. Ask of each name whether all three channels run in its favour — cheaper inputs, metered demand, rising authority premium — and whether the order book, not the presentation, confirms it. Where they do, the 2026 derating has handed you a business on the winning side of the trade at a price set by people who filed it with the losers. In 2026, in European software, that is most of the opportunity set this firm exists to work through.

Sources and method

This is a framework note, not an initiation of coverage, and contains no recommendation, rating, or price target. It extends the argument of "The Loser's Game: Why Software Survives AI by Not Losing" (Larix Research, 11 July 2026). Figures are drawn from company reporting and public wires; the companies named (SAP, RELX, Wolters Kluwer, Dassault Systèmes, Temenos, Sage, Amadeus, Deutsche Börse, Scout24, SMG) are illustrations of a structural argument; with the exception of SMG, they sit outside the undercovered universe this firm exists to examine, and several are covered by sell-side research.

SAP: Q2/H1 2026 results release and presentation, 23–24 July 2026 (current cloud backlog €22.9bn, +27%; cloud revenue €6,281m, +22%; Cloud ERP Suite €5,525m, +25%; FY2026 cloud guidance €25.8–26.2bn; developer productivity uplift up to 30%; €2bn efficiency run-rate by 2028). Q4/FY2025 quarterly statement, 29 January 2026 (revenue €36.80bn; cloud €21.02bn, +23%; non-IFRS operating profit €10.42bn, +28%, margin 28.3%; FCF €8.24bn; €10bn buyback). Integrated Report 2025 (more than two-thirds of Q4 2025 cloud order entry containing SAP Business AI, +20pp vs Q3; more than 110,000 employees). "87% of global commerce" per SAP's own corporate FAQ. Business Data Cloud and Databricks announcements, 13 February 2025. AI Units pricing per SAP's published pricing pages. Share-price context (~$150bn wiped from the 2025 peak by late January 2026; intraday fall of as much as 17% on 29 January 2026, more than €40bn erased in the session, steepest since October 2020; shares down 37% year-to-date as of late June 2026) — Reuters wire, 29 January 2026, via syndicated copy; Morningstar, 24 June 2026.

RELX: 2025 results presentation, 12 February 2026 (revenue £9,590m, +7%; adjusted operating profit £3,342m, +9%; margin 34.8%; Legal +9% revenue, +12% profit). AGM trading update, 23 April 2026 (Lexis+ with Protégé driving double-digit legal growth). Company overview deck, June 2026 (AI as key driver; more than three quarters of AI coding interactions agentic).

Wolters Kluwer: 2025 Full-Year Report, 25 February 2026 (revenues €6,125m, +6% organic; 83% recurring; margin 27.5%; ~70% of digital revenues from AI-powered solutions; UpToDate Expert AI launch October 2025). Q1 2026 Trading Update, 6 May 2026 (more than half of US enterprise customers signed to adopt UpToDate Expert AI). OpenAI expanded enterprise AI collaboration announced 3 June 2026.

Dassault Systèmes: Q4/FY2025 results, 11 February 2026 (revenue €6.24bn, +4% constant currency; subscription +11%; 3DEXPERIENCE +10%; margin 32.0%; NVIDIA Industry World Models partnership); Q1 2026 results, 23 April 2026. 3D UNIV+RSES strategy announcement, 4 February 2025. "Worst day ever," 11 February 2026 — Morningstar, 25 February 2026.

Temenos: 2025 Annual Report and Q4 2025 release, 24 February 2026 (950+ core banking clients; ARR $860m, +12%; EBIT margin 34.7%; raised FY2028 targets); Q2 2026 release, 22 July 2026 (ARR $881m, +11% constant currency; deal slippage, none lost; Q3 subscription and SaaS guidance of at least +30%).

Sage: FY2025 results, 19 November 2025 (revenue £2,513m, +10%; ARR £2,574m, +11%; 97% recurring; renewal rate by value 101%; Sage Copilot scaled across core products).

Amadeus: FY2025 results, 27 February 2026 (revenue €6,517m, +6.1%; adjusted EBIT €1,894m, margin 29.1%; R&D €1,434m, 22% of revenue); CEO statement and SkyLink acquisition per contemporaneous trade-press coverage, 27 February 2026.

Deutsche Börse: FY2025 results, 11 February 2026 (net revenue excluding treasury €5.2bn, +9%; EBITDA €2.7bn, +14%); ISS STOXX investor presentation, March 2025 (93% recurring revenue; 44% adjusted EBITDA margin; ~90% share in European structured products on STOXX/DAX; $150tn+ of assets benchmarked).

Market context: 3 February 2026 sell-off (Thomson Reuters −16% at the close, RELX −14%, Wolters Kluwer −13%) — Morningstar Equity Research, 4 February 2026; the intraday ~18% Thomson Reuters decline and the "worst session on record" characterisation are from Reuters syndicated copy. Sector drawdown framing (−34%, ~$2tn, largest non-recessionary twelve-month drawdown in more than thirty years) — J.P. Morgan note by Dubravko Lakos-Bujas and colleagues, quoted in Fortune, 10 February 2026. Barclays "sell first, think later" report of 11 February 2026 (team led by Emmanuel Cau; overweight software with specific calls on Experian, Adecco, RELX, Wolters Kluwer, Scout24, Vend and Auto Trader) — Dow Jones Newswires via Morningstar, 11 February 2026. Deutsche Bank on RELX ("a mispriced AI beneficiary… proprietary deep data/intelligence names such as RELX should see their value increase as AI interface tools launch and look to better data") — via interactive investor, 4 February 2026. BofA "proprietary datasets" winners framing — Morningstar, 25 February 2026. Quilter Cheviot on the RELX derating (~17x 2027 earnings from a ~31x peak) — 23 April 2026. Dassault Systèmes' more-than-21% single-day decline per contemporaneous secondary wire coverage, 11 February 2026.

Scout24 and SMG references are as discussed in "The Loser's Game" (11 July 2026); TECMA references are to the TECMA Solutions initiation (3 July 2026). We hold no position in the securities mentioned and received no compensation from any party in connection with this note. All market-move figures were verified against primary or high-authority sources on 24 July 2026; two rest on contemporaneous secondary wires where the original wire could not be retrieved and are noted as such above (the "worst session on record" characterisation for Thomson Reuters; the Dassault Systèmes single-day decline).