The Ice Cube Test

Sponsored by Crawford McMillan

Every investment committee evaluating software right now is pricing a risk none of them has ever measured.

Q2 made that expensive. Software deal value fell 65.7% year over year, per PitchBook's Q2 US PE Breakdown, not because buyers think every legacy software business is doomed, but because they cannot tell which ones are. Interest rates can be modeled. AI exposure, apparently, cannot. So price discovery stopped.

Two weeks ago this newsletter put it precisely. Some legacy software businesses are coiled springs, others are melting ice cubes, and at underwriting they show remarkably similar ARR, retention, and EBITDA margins.

Here is what should bother you about that framing. The difference is not unknowable. It is unmeasured. The two types of companies look identical because diligence reads the standardized scoreboard, and the scoreboard was never designed to show you how AI-vulnerable a business is.

The difference lives one layer down, in data almost every target already has and almost no investment committee ever asks for.

______________________

Here are four unasked questions, all answerable from the target’s own systems inside a diligence window.

1. What is the quality of the renewals, not the rate?

A 92 percent gross retention number is an average, and averages are where truth goes to hide. Pull the renewal history apart by cohort and by circumstance.

• Which customers renewed after genuinely evaluating alternatives?
• Which renewed because the renewal was automatic and nobody looked?
• How much retention is carried by price escalators on a shrinking base of logos?

A coiled spring gets renewals because customers keep choosing it. An ice cube gets renewals because switching is a project nobody has prioritized yet. The renewal rate alone cannot tell you which one you are buying.

2. What does the telemetry say people actually do in the product?

Most software companies sit on years of usage data they have never been asked to produce in a deal.

• How many seats log in weekly versus seats that exist on an invoice?
• How many modules that carry daily workflow versus modules that were bought and abandoned?
• How much usage is deepening into the workflows that are hard to leave or pooling in the features any competent team could rebuild.

The login data does not care what the management presentation says. If four engineers with modern tools could quickly rebuild what customers actually use, you are not buying a product. You are buying the customers’ inertia, on a timer.

3. Where is the switching cost?

Switching cost is claimed in every CIM and quantified in almost none.

• How much of the customer’s own operational history lives inside the product and nowhere else?
• How many integrations would break on exit, and who owns fixing them?
• How many of the customer’s daily processes exist as configuration built up over years?

These are queryable facts. A useful early warning sits in the support queue, because customers preparing to leave start asking for their data. Export requests are the sound of ice starting to melt, and they show up quarters before churn does.

4. Does the business own data that a rebuild cannot replicate?

This is the question AI has made decisive. The old moat was the cost of recreating the product. If that moat is now forty engineers, the durable question becomes what those engineers could not recreate at any headcount.

Ten years of proprietary ground truth, benchmark data that only accumulates from being the system of record, a network of records competitors cannot assemble. Some legacy software businesses hold data assets that make them more valuable in an AI world, because the models everyone can buy are only as good as the data nobody else has.

That is the coiled spring. A feature set with no accumulated state underneath it is the ice cube, however elegant the AI-enabled interface design is.

________________________

None of this is exotic. It is renewal records, usage logs, integration inventories, and schema reviews. So why does almost nobody look?

Partly because the data room does not volunteer it. Data rooms are curated by sellers, and sellers of ice cubes have no incentive to surface the melt rate.

Partly because it falls between workstreams.

• Financial diligence stops at whether the numbers are right.
• Commercial diligence asks customers what they think, and customers are polite right up until they leave.
• Technical diligence reviews the architecture, not the evidence of demand.

The layer where coiled springs and ice cubes diverge, the data a software business generates about its own durability, has no owner in the standard diligence stack.

For twenty years nobody needed to look - software got the benefit of the doubt. The doubt is now the whole discount.

This is the uncomfortable arithmetic for anyone holding legacy software at 2021 marks. The repricing has already happened in rejected bids and pulled processes. It has not happened in the marks, because marks are patient when nobody forces them to meet a transaction.

When that patience runs out, the funds that can show diligence-grade evidence under their retention and usage numbers will defend their valuations. The funds waving averages will take the ice cube discount whether they are holding ice or not.

That’s the part that is genuinely avoidable. The market is not precisely punishing melting businesses, it is pricing every unmeasured business as if it were melting.

In the same quarter the category’s deal value collapsed, the largest software-focused fundraise on record closed at $21 billion. The specialists are not braver than everyone else. They are underwriting one layer down, company by company, where the difference is visible.

And software is only where this shows up first, because software is where the AI question is loudest. The same logic is coming for every mid-market business whose numbers look fine from altitude. The M&A market has already bifurcated.

Assets trade fast at premium valuations while everything else waits, and what separates the two is rarely the business. It is whether the evidence underneath the numbers survives a hard look.

Mid-market deals get the thinnest diligence budgets and carry the most held-together-by-three-spreadsheets data, which means the unmeasured discount lands there hardest of all.

The spring and the ice cube only look identical from the altitude most deals are done at. Fly lower. The data has been keeping score the whole time.

Graeme Crawford is the CEO of Crawford McMillan, a data advisory firm that works the layer underneath financial diligence for PE-backed mid-market companies, and writes Inside The Data Room, a weekly memo for operators who know the data is the deal.

Sources

PitchBook, Q2 2026 US PE Breakdown - pitchbook.com/news/reports/q2-2026-us-pe-breakdown

Francisco Partners, $21 billion close across FP VIII and Agility IV, July 23 2026 - franciscopartners.com/media/francisco-partners-closes-21-billion-across-flagship-and-agility-funds