The Operating LeverageNo. 001
Nobody pays for software any more. They pay for proof.
Nine weeks of watching what an unpriced AI question costs an asset, and this month the market finished setting the price.
The software repricing is essentially over. Forvis Mazars (opens in a new tab) puts median EV/EBITDA on private equity SaaS buyouts at 11.7x for the first half of 2026 against 20.4x in 2025, on PitchBook data as at 30 June, with non-SaaS buyouts at 11.0x. That is a fall of nearly nine turns in a year, and what is left of the software premium is seven tenths of one.
That price was set outside the firms that own the assets. Reuters (opens in a new tab) found 81% of software loans marked down across the filings of 44 US business development companies, against 40% outside software, and HarbourVest's benchmark (opens in a new tab) then marked Software and Services at -6.4% for the first quarter, with European Information Technology still positive at +11.1% over the trailing year.
So nothing you own is paid for being software, and a European book has not been asked yet. What is paid for is an answer to one question per asset: what does AI do to this revenue model.
Three practitioners in three countries built the same instrument
Sander van Woerden of Lexar Partners, on a Dutch panel reported by MenA.nl (opens in a new tab), screens for what resists AI pressure: deep vertical lock-in, systems of record, hardware coupled to the software. Three days later, Katja Schwarz of Verdane told pemacom in Munich, reported by Unternehmeredition (opens in a new tab), that the test is whether a model is strengthened by AI, substituted by it, or priced down by it. A week after that, BearingPoint (opens in a new tab)'s Andre Waßmann and Miriam Dusdal named where capability fails once a buyer looks: data that will not hold under production load, interfaces that pass testing and fail at scale, systems that do not fit the estate.
Three screens, three countries, eleven days, no coordination. The instrument is already public. Nobody has scored one.
The question has reached the people who give you money
Alexandre Armbruster of Caisse des Dépôts, quoted by PitchBook (opens in a new tab): "We are looking at new criteria and KPIs to better understand risk." He is the first LP on this beat to call AI resilience scoring criteria.
UK Private Capital (opens in a new tab) asked 76 senior decision makers what they do with AI: 92% use it in diligence and document review, 20% fund anything for AI at their portfolio companies. The deal desk is finished. The asset the LP is asking about has had nothing spent on it.
If you run a fund under $30bn
- Run one screen across your three largest assets this month, and write the answer down. Take van Woerden's, Schwarz's or BearingPoint's. They agree more than they differ, each one costs a morning, and the written version is the only one that survives a partner leaving.
- Reunderwrite against 11.0x, not against 2021. A thesis written when software-ness carried nine turns of premium is being exited into a market paying seven tenths of one. Your exit committee will get there before you do.
- Count your own 92 and your own 20. The diligence desk is almost certainly using AI already. Ask instead what you funded inside a portfolio company this year, and whether any of it came back with a number attached to it.
- Expect the question in your next fundraise, in writing. Armbruster is describing criteria and KPIs, not a conversation. Charterhouse (opens in a new tab) bought judgement for it on 29 September, appointing Professor Neil Lawrence and Mehdi Ghissassi as senior advisers. Two people, no platform. It is the cheapest credible answer on the table, and it includes nothing that re-runs next quarter.
What we'll be watching
Three things that would change this argument, and what each would mean.
- Whether a second named LP puts AI resilience into its scoring criteria. One allocator is a view. Two is a questionnaire, and the proof stops being optional for anyone raising in 2027.
- Whether the European marks turn. If the next quarterly benchmark takes European Information Technology negative, that +11.1% was a reporting queue, and the time we have just told you that you have does not exist.
- Whether a fourth screen arrives, and whether anybody scores one. Four practitioners reaching the same instrument independently would make a per-asset read the standard. If they stay unscored, a per-asset premium stays unprovable, and 11.0x is what the market pays for everything.