Investing in software? You bet your assets

The Wheel Is Come Full Circle

In August 2020 we published Investing in Software? You bet your assets. It argued that software is a distinct asset class, that revenue multiples are a rational lens on it, that a well-run software company can levelize at 45–50% cash-flow margins, that a 10x revenue multiple is defensible from first principles, and that software returns outrun most other asset classes. Six years is long enough for the market to have an opinion.

This addendum puts every load-bearing claim in that paper against the data and reports the result — including where the original was wrong, and where the arithmetic did not follow from its own inputs. We are publishing the corrections rather than quietly reprinting. An investor who acted on the 2020 paper deserves to know which parts survived contact with 2022, 2023 and 2026.

The Six Conclusions, Graded

The 2020 paper listed six conclusions on its first page. It is fair to grade those before explaining any of them. Five hold. One does not, and it is the one we stated most confidently.

Verdicts are against the sources listed in the appendix. Confirmation of a conclusion is not confirmation of the figures underneath it — Sections II to IV set out where those have moved.

The Machinery of Value: Now Is the Winter of Our Discount Rate

The 2020 paper built a discount rate from scratch and arrived at 6.87%. It is worth restating the build, because the method is still right and only the inputs are wrong. We took a 1% nominal risk-free rate (0% real, with a 1% inflation premium), added a 2% duration premium for a roughly six-year hold, added 2.5% for default risk using the spread between the 10-year Treasury and Moody's Baa corporate yield, and grossed the 5.5% subtotal up by 1.25 for a 20% private-company discount.

Every judgment in that build was defensible in 2020. The 0% real risk-free rate was, if anything, conservative: the 10-year TIPS yield actually closed 2020 at −0.98%. We leaned on Paul Schmelzing's Bank of England work arguing that real rates were in an eight-century secular decline and that negative real rates could persist for a long time.

That call is now decisively wrong, and the authority we cited has revised it himself. Rogoff, Rossi and Schmelzing published in the American Economic Review in August 2024 finding that real rates are trend stationary, not a random walk; that the long-run downward trend is roughly 1.6 basis points a year — about ten basis points over a six-year hold, which is to say nothing; and that mean reversion runs on half-lives of one to six years. On their revised reading, the post-crisis collapse in real rates was a cyclical deviation, not a secular shift. The 10-year real yield has now been positive every month since May 2022 and above 1.6% for thirty-nine consecutive months.

Rebuilding the Rate

Substituting today's observable inputs into the paper's own method: the three-month bill is 3.89%, the Baa–Treasury spread is 1.57%, and we keep the paper's 2.00% duration judgment even though the curve now pays about 0.70% at the six-year point. Credit is the one component that argues for a lower rate than we used — the spread has averaged under 2.0% every year since 2022, well below the 2–3.5% band we described. It is not nearly enough to offset the risk-free move.

What 10x Actually Requires

A revenue multiple is not a claim about cash flow today. It is a claim about the size and durability of cash flow later, and the arithmetic only works if growth is carried explicitly. The 2020 paper did carry it — it valued software with the present value of a growing annuity rather than capitalizing a flat cash flow — and that method is right. Capitalize 45% at 6.87% with no growth and you get 6.6x, which is why the critics of software multiples always arrive at a low number: they are pricing a business that has stopped compounding.

What the paper did not do was state the growth rate its 10x conclusion required, and that omission is worth repairing, because the growth rate is the whole argument. Run the method with its own inputs — revenue compounding at g for five years, then a 45% cash-flow margin capitalized at the discount rate less 2% terminal growth, discounted back — and the threshold falls out directly.

The rate shock did not invalidate 10x. It raised the price of admission. For a fund underwriting $1–8M ARR companies scaling to $15M — which is to say growth in the twenties and thirties — every one of those thresholds is inside the range. That is the useful form of the claim, and it is more demanding than the original. It says 10x is cheap for a company compounding at 30% and indefensible for one compounding at 10%. The 2021 vintage did not fail because 10x was wrong; it failed because people paid 10x and more for companies growing 15%, and then the growth slowed.

To Be Forty-Five Percent, or Not to Be

The single most consequential number in the 2020 paper was this: once a software company shifts from growth generation to cash-flow generation, the sustainable net cash-flow potential should be 45 to 50% of revenue. Everything downstream depended on it — the valuation formula, the sales-efficiency ceiling, the comparison to a stabilized office building.

It is the claim we are challenged on most, and it is almost always tested the wrong way. Point at Microsoft's 20% free cash flow margin, or Salesforce's 34%, and the rule looks fanciful. But Microsoft is building data centers and Salesforce is spending roughly a third of revenue on sales and marketing. Both are making investment decisions; neither is reporting a cost of operation. Measuring companies that are still buying growth against a no-growth steady state answers a different question than the one asked.

The right test is software that has actually stopped growing.

Free cash flow margin against revenue growth, latest reported fiscal year, from company filings via stockanalysis.com. Excludes licensing and registry businesses (VeriSign 64.5%, InterDigital 63.4%) as unrepresentative of operating software. Figures are before stock-based compensation.

Qualys delivers 45.5% while growing 10%. Check Point delivers 43.0% growing 6%. Dropbox — the cleanest case in the set, because it has already cut sales and marketing to a maintenance level — delivers 36.9% at minus one percent growth. The rule is real. What is not true is the strong form of the objection, that essentially no software company reaches the band; it fails the moment you look at companies that have stopped buying growth.

The Cost Floor, Independently Measured

The structural argument holds up better than the original stated it. SaaS Capital's 2026 spending survey of more than a thousand private B2B SaaS companies puts customer support and success at 9% of revenue, G&A at 15%, and total R&D at 22%. Take support, G&A and half of R&D as the genuinely non-discretionary maintenance load and the floor is 35% of revenue below gross profit — the 2020 paper's number, arrived at independently. At public-company scale, where G&A runs 6–9% rather than 15%, the floor is nearer 28%.

And the evidence that the rest is discretionary is stronger than we had in 2020. Three findings:

1. Margin and growth are uncorrelated in the public cross-section. Across 69 public B2B software companies the correlation between net income margin and revenue growth is R = 0.18, p = 0.14. The standard defense of thin software margins — that the money is buying growth — has no statistical support.
2. Ownership model changes the spend, not the business. PE-backed private SaaS runs sales and marketing at 33% of revenue against 47% for venture-backed companies at comparable scale. Bootstrapped companies are within two points of breakeven or profitable 83% of the time; equity-backed, 52%.
3. Growth spend can be removed without killing growth. Instructure grew 23% in 2019, was taken private by Thoma Bravo in March 2020, and by its FY2021 re-IPO year reported revenue of $405.4m growing 34% with adjusted unlevered free cash flow of $168.7m — 41.6% of revenue. Growth accelerated while the business was being run for cash. Note this is one company, and roughly half the revenue gain came from price rather than customer count.

Three Qualifications That Do Bind

The rule holds. Its domain is narrower than the original implied, and three conditions decide whether a given company sits inside it.

1. Say whether stock compensation is in or out. Reported free cash flow adds back stock-based compensation, and for software that is not a rounding item: 7.5% of revenue at Check Point, 11.9% at Dropbox, 15.6% at Zoom. Deduct it and Check Point's 43.0% becomes 35.5%, Qualys' 45.5% becomes 34.0%, Dropbox's 36.9% becomes 25.0%. Nobody in the operating cohort clears 36% on that basis. The claim is defensible either way, but it is unfalsifiable until the definition is stated. Every margin figure here is pre-stock-compensation unless marked otherwise, and the 45–50% band should be read the same way.
2. The gross margin premise does more work than the cost floor. Every company above 30% cash flow margin in the table above has a gross margin at or above 80%. Every company below 15% either has gross margin below 76% — 8x8 at 64.6%, Blackbaud at 58.8% — or carries a services mix or acquisition debt. Median public B2B software gross margin is 76%, not the 85% the arithmetic assumes, and for product-led and SMB software it is 69%. Each point of gross margin below 80% comes straight off the ceiling.
3. Replacing churn is not discretionary. This is the one that changes the answer, and it is internal to the paper's own framework rather than imported from outside it. A company retaining 84% of its revenue gross must win back sixteen points every year simply to stand still. That is not growth spend. At a new-logo cost of $2.00 per dollar of ARR it consumes 32% of revenue; at a blended $1.30, 21%. It belongs below the line with support and G&A, not above it with growth.

Read those together and the argument between "45–50% is a rule" and "the median delivers 30%" resolves completely. Both are right, about different companies. At 96% gross retention the cost of standing still is five to eight points of revenue and the steady state lands at 47–50%. At the current private median of 84% it is twenty-one to thirty-two points and the steady state lands at 23–34%. The cost structure was never the binding constraint. Retention is.

That is a more useful statement than the original, because it makes the 45–50% claim testable in diligence rather than aspirational. It also unifies two sections of the 2020 paper that were written as separate arguments: cash-flow stability and cash-flow magnitude are the same variable. Churn is not merely how stability gets expressed, as we wrote in 2020 — it sets the achievable margin directly, at roughly 1.3 to 2.0 points of steady-state margin per point of retention.

Fair Is Foul, and Foul Is Fair: What the Market Paid

The 2020 paper made two claims about market pricing. The first was that exit multiples are driven by revenue growth, with a relationship "linear, bordering on exponential." The second was that the SEG SaaS index had outperformed the Nasdaq 100, which in turn outperformed the S&P 500. The first has held and sharpened. The second has inverted.

Median EV/NTM revenue for the Clouded Judgement public software basket (~100 companies, consistent methodology throughout). The June 2026 trough of 3.1–3.2x was below the December 2022 low, and in the first quarter of 2026 software traded at or below the S&P 500 forward multiple for the first time in the cloud era.

Two full cycles in six years. The first drawdown, 2021–22, was the ZIRP unwind: median multiples fell from 17.7x to 5.2x, roughly $1.4 trillion of cloud market capitalization evaporated, and the punishment was quality-sorted — companies below Rule of 40 compressed 78% against 58% for those above it. The second, in the first half of 2026, had no recession behind it at all. It was an AI-disruption repricing: the software category fell 34% over twelve months, the steepest non-recessionary decline in more than thirty years, with the sharpest single-day damage following a model launch rather than an earnings miss.

Median EV/NTM revenue by growth cohort, Clouded Judgement. Cohort cut-offs moved with the market — "high growth" meant above 30% in 2020 and above 22% in 2026 — so this is a directional comparison, not a like-for-like one.

The convexity we described is intact and steeper. The ratio between the top and bottom cohorts widened from roughly 3.8x in the SEG 2020 data to 4.6x today, and on SEG's own trailing series the spread is 5.3x: 2.4x for companies growing under 10%, 12.7x for those growing 20–30%. What has changed is that every level fell, and slow growth is now barely priced at all. A software company growing under 15% is worth 3.9x revenue, half what the same company was worth in 2020.

It is worth being precise about what the market is actually saying, because it is not that the model is wrong. Solve the two-stage model backwards: at the public median growth rate of about 12%, today's 4.4x implies a stabilized cash-flow margin of roughly 29%. That is not a rejection of the 45–50% rule — it is a price consistent with a business retaining 88–90% of its revenue rather than 95%, which is exactly where the public median now sits. The market is pricing the retention deterioration, and pricing a little more risk on top for AI. Public multiples are a statement about the median company's durability, not a test of what a durable company is worth.

One caveat against our own thesis: the statistical relationship between growth and multiple is weakening even as the cohort medians move apart. First Analysis measures the correlation for 2026 estimated revenue at R = 0.35 in the second quarter, down from 0.48 in the first. Dispersion within cohorts has exploded. The right description of today's market is not a curve but a bimodal distribution: five companies above 10x forward revenue, 80% of the universe below 5x, and very little in between. The market is not paying for growth so much as it is paying for a small set of names it believes are on the right side of AI.

The Relative-Returns Claim — the Plainest Failure

One basis for all three, since a total-return series is not published for the cloud index (FRED NASDAQEMCLOUD; QQQ and SPY price levels). On a total-return basis the ordering is unchanged. The 2020 paper asserted the ordering SaaS > Nasdaq 100 > S&P 500.

This is the plainest failure in the paper. We wrote in the autumn of 2020, at the top of a software melt-up, and extrapolated a two-year index history into an asset-class property. Cloud software lost to the Nasdaq 100 in 2021, 2022 and 2025, and its five-year annualized return through June 2026 was −49.7%. The SEG SaaS index fell 22.2% in 2025 against an S&P up 17.9%.

The distinction that matters — and the paper should have drawn it at the time — is between the asset and the entry price. The 2020 paper's other conclusion, that software returns should approximate revenue growth, is close to an identity: buy and sell at the same multiple and the return is the revenue CAGR, less dilution. Everything that went wrong for the index went wrong at the entry point.

A company compounding revenue at 25% for five years (3.05x cumulative), sold at 4.5x revenue. The only variable is what the investor paid going in. Gross multiple on invested capital, before ownership, fees and dilution.

This is why an index that bought at 17.7x in December 2020 lost to the S&P while the underlying companies kept compounding, and it is why our own funds did not. Fund I is marked at 2.0x and Fund II at 2.01x TVPI through the steepest software repricing in two decades. Neither number is an argument that software outperformed as an asset class over the period. Both are arguments that entry discipline is the variable, and that the 2020 paper was right about the mechanism and wrong to attach it to an index that had no discipline at all.

The Operating Reality: Something Is Rotten in the State of Retention

The intellectual core of the 2020 paper was not the multiple. It was the claim that software cash flows are stable, and that the stability comes from a specific mechanism: the customer's problem, the value proposition, the accumulating data asset, switching costs and contract strength. We wrote the stability as a function of those five variables and said churn is how it gets expressed. That was the argument for treating software like a bond, and it is the argument the last two years have damaged most.

Benchmarkit annual B2B SaaS benchmarks (n ≈ 340–580). The 2025 fall in gross retention appeared at every quartile — median 88% to 84%, upper quartile 95% to 91%, lower quartile 78% to 74% — which makes it a distributional shift rather than a tail event.

Two things are wrong here relative to what we wrote. The first is that we treated 105% net dollar retention — −5% net churn — as an unremarkable good case. It was never the private median. SaaS Capital measured 100% in 2019, the paper's own vintage; 105% was the 2021 peak, and it has fallen every year since to 101%. We anchored on the best number available at the top of a cycle.

The second is more serious. Net retention is a blended figure, and the blend has been flattering. Gross retention — which is the direct measurement of the stability we theorized — fell from 88% to 84% in a single year, at every quartile. A company losing sixteen cents of every revenue dollar each year and replacing it through expansion is not the annuity we described. Expansion now supplies 40% of net new ARR at the median and 67% above $100M ARR, which means the reported top line increasingly rests on selling more to a shrinking installed base.

The Variable We Did Not Have

In 2020 essentially all B2B software was priced per seat and the question did not arise. Usage-priced companies now retain at 108% net; seat-priced companies retain at 98%. That is a ten-point structural gap in a variable we never modeled.

This belongs in our cash-flow-stability function as a sixth term. The mechanism is straightforward: a seat-priced contract is a claim on the customer's headcount, and headcount is now the thing customers are most actively trying to reduce. A usage-priced or outcome-priced contract is a claim on the customer's work volume, which is not falling. ServiceNow reported that 50% of net new business in its second quarter of fiscal 2026 was non-seat-based, and its renewal rate is 98% — the 2020 paper's stability thesis, intact, in the most AI-exposed workflow category there is, at a company that repriced away from seats.

The Course of True CAC Never Did Run Smooth

The 2020 paper contained a genuinely provocative claim, and we made it on purpose: that with a median cost of $1.34 to acquire $1 of new ARR, and a rational willingness to pay up to $4.50 (or $6–7 with a near exit), software companies as an industry were not spending enough on sales. Not burning enough cash. It is the argument we were most often challenged on, and it has not survived.

Three corrections. First, on the baseline: we could not re-source the $1.34 figure to a primary KeyBanc document. The defensible baseline for that vintage is KeyBanc's CY2020 blended figure of $1.20, and that is the number to compare against. Second, on the level: the correct comparator for "cost to acquire $1 of new ARR" is the new-logo ratio, and that now sits at $2.00 at the median and $2.82 at the fourth quartile. Acquisition is roughly 50% more expensive than the number we built the argument on, and it rose 14% in a single year while blended CAC was falling. Blended CAC fell over the same period only because expansion ARR — which costs about $1.00 per dollar — rose from 25% to 40% of net new ARR. That is a mix shift, not an efficiency gain.

Third, on the willingness-to-pay ceiling. The $4.50 figure was derived from the 45–50% cash-flow margin and a 9x exit. Both of those inputs are now materially lower. Rebuilt on a 22% cash-flow margin and a 4.5x exit, the rational ceiling is closer to $2.20–$3.00 per dollar of new ARR — which happens to be roughly where the fourth quartile is already spending. The industry is no longer under-spending on sales. At the margin it is spending about the right amount, and the top quartile may be over-spending.

There is a related finding that changes how we would run the diligence. The variance in growth outcomes is no longer well explained by sales spend at all. High Alpha's 2025 cohort analysis splits companies by retention and payback rather than spend: companies with net retention above 106% and CAC payback under ten months grew 71% at the median with a Rule of 40 score of 47; companies below 98% retention with payback over fifteen months grew 10% with a score of 5. A seven-fold growth spread, and sales-and-marketing intensity is nearly flat across every revenue band at 25–30% of revenue. The throttle we described is real, but it is downstream of retention, not independent of it.

If Recessions Be the Bane of Returns — a Reckoning

The 2020 paper made five falsifiable claims about how software behaves in a downturn, drawn from 2008–09 and from our own earlier piece, The Golden Age of B2B SaaS? The 2022–23 software recession was a clean test. We got two right, one half right, and two wrong.

The profitability call is the one we should be most pleased with, and it is worth being precise about why we were right for slightly the wrong reason. We argued margins would expand because buyer behavior improves in a downturn. In fact margins expanded because vendors cut — R&D fell eight points as a share of revenue, 703,000 people lost jobs, and prices went up. It was supply-side austerity, not demand-side rescue. Margins also got worse before they got better, bottoming twelve to eighteen months after the shock rather than during it.

The CAC call failed because we modeled the wrong buyer psychology. We predicted that managers pushed into loss territory would become more willing to buy change. What actually happened was procurement scrutiny: average contract values in Vendr's transaction data fell 51% from the second quarter of 2022 to the second quarter of 2023, renewal ACV fell 11% across 2023, one-year contracts rose from 79% to 85% of the market, and vendors held price and lost volume rather than discounting. Buyers behaved procyclically and defensively.

And Yet the Demand Thesis Held

Gartner worldwide IT spending growth by segment. Gartner restates prior years at every release, so no single press release contains this set. Later vintages put 2022 devices between −6.3% and −10.6%; the direction is stable, the decimal is not. Software grew 15.5% in 2026 to $1.47tn.

This is the chart that rescues the section. In the two worst years for software as an investment, software was the fastest-growing category of enterprise spending on earth. Total IT spend was roughly flat in 2022; devices fell 8.4% and again 8.6%; data center systems fell. Software grew 10.7% and then 12.4%. Gartner's own language at the time was that "enterprise IT spending is recession-proof," with 69% of finance leaders planning to increase digital spend into the teeth of the downturn.

So the correct formulation is narrower than the one we published. Software demand is countercyclical at the aggregate spend level, because enterprises substitute software for labor and hardware when they need efficiency. Software economics are not countercyclical at the unit level: it gets harder and more expensive to win a new logo precisely when budgets tighten. Both statements are true, and the 2020 paper conflated them.

What Has Appeared Since: All the World's an Agent

AI is not the only thing that changed since 2020, and this addendum deliberately does not treat it as the whole story — the rate move did more damage to the valuation math than AI has. But AI is the only development that attacks the paper's assumptions rather than its inputs, so it deserves its own section. We will take the bear case and the bull case at full strength, because both are strong.

The Case Against the 2020 Assumptions

ICONIQ State of AI, roughly 300 AI product leaders per wave. AI products were 32% of software revenue in 2025 and 42% in 2026, projected at 53% for 2027. Inference alone is about 23% of total spend at the scaling stage, up from 20% pre-launch — a cost line the 2020 paper's 10–15% cost-of-sales assumption has no room for.

The gross margin problem is the concrete one. We wrote that software runs at 70–90% gross margin with cost of sales at 10–15% of revenue, mostly hosting. Traditional SaaS still delivers on that — the median is 76–80%. But AI-delivered revenue does not, and AI-delivered revenue is now 42% of the total. Blend a 42% share at roughly 53% margin with a 58% share at 78% and you get about 68% — already below our floor. The trajectory is genuinely improving as per-token inference costs collapse by an order of magnitude a year, but it lands roughly twenty points below what we assumed, and cost of goods is now variable with usage in a way it never was for seat-based software. Compression is worst in exactly the band we invest in: High Alpha measured gross margin down seven points year over year below $1M ARR and four points at $1–5M, against flat above $5M.

The second problem is subtler and, we think, the more important one. The 2020 paper said selling software is selling change, and people want change like they want a hole in the head. We presented that friction as the reason acquisition is hard. What we did not notice is that the friction and the moat were the same thing. AI has reduced change-resistance measurably — sales cycles fell from 25 weeks in 2025 to 19 in 2026, and 20% of new startups charge a first customer within thirty days against 8% in 2020 — and that reduction runs in both directions. It is easier to sell into an account and easier to be displaced from one. Gross retention falling four points at every quartile is the same phenomenon viewed from the other side.

On durability, the AI-native cohort is a warning rather than a template. ChartMogul's 2025 panel of 200 AI-native companies shows median gross retention of 40% and net retention of 48%, against 82% net for B2B SaaS in the same panel. Speed is real — median time to $1M ARR is 11.5 months against roughly 15.5 for the fastest traditional SaaS, and Bessemer's fastest cohort reaches $100M in eighteen months against seven years — but a dollar of that revenue is worth considerably less than a dollar of ours. In discounted cash flow terms the two effects substantially offset, which is a point neither the bulls nor the bears usually make.

The Case That the 2020 Assumptions Hold

The bull case is not sentiment; it has numbers behind it, and they are large.

1. Aggregate demand accelerated. Gartner has software spending growing 15.5% in 2026 to $1.47tn, faster than 2025's 13.9%. IT services — the category actually being substituted by AI — grew 5.3%. Software is the beneficiary of that substitution, not the victim.
2. Incumbents are the largest AI monetizers in absolute dollars. Salesforce reported Agentforce and Data 360 ARR of about $3.9bn growing 210% in its second quarter of fiscal 2027. ServiceNow crossed $1bn of AI ACV growing 40% quarter on quarter. Microsoft reported over 30 million paid Microsoft 365 Copilot seats for the quarter ended 30 June 2026.
3. Systems of record are holding. ServiceNow renews at 98%. Expansion still costs $0.80 per dollar of ARR against $1.63 for a new logo — the installed base remains the cheapest growth available, which is precisely the switching-cost argument the 2020 paper made.
4. The market has repriced ahead of the evidence. We could not find a single large-cap enterprise software company that attributed a seat decline to AI on a 2025 or 2026 earnings call. Chegg blamed search interfaces; HubSpot blamed its own agent price cuts. Deutsche Bank found no software company guiding to negative AI revenue impact in 2026. Roughly $2tn of software market value has been marked down for a disclosed deterioration that has not yet appeared.
5. The TAM argument is quantified, not hand-waved. Enterprises spend roughly $6 on services for every $1 on software. A vendor that captures part of the services line through outcome pricing expands its own market by a multiple. Gartner expects 40% of enterprise applications to embed agents by the end of 2026, against 5% in 2025.

Where We Come Out

The build barrier has genuinely fallen and the maintain-secure-integrate-and-be-trusted barrier has not. Google reports 75% of its new code is AI-generated and 90% of developers use AI tools; against that, METR's randomized trial found experienced developers were 19% slower with AI available while believing they were 20% faster, and GitClear's analysis of 623 million code changes finds refactoring down 70%, duplicated blocks up 81% and two-week churn up 15%. Google's own DORA research finds AI positively related to throughput and negatively related to delivery stability.

For a vertical B2B software investor that resolves cleanly. Getting to a demoable v1 is dramatically cheaper, so the number of competitors at the bottom of the market goes up. Getting to a system a regulated mid-market business will run its operations on is not cheaper, and may be getting harder. The moat was never the code.

One finding does cut against us specifically, and it should be said plainly. ICONIQ reports that 43% of AI products being built are vertical applications and roughly 70% of AI builders are focused on vertical AI. The AI-native attack is aimed directly at vertical software, because vertical workflows are where domain-specific agents have the clearest return. What protects a vertical company is proprietary workflow data and regulatory embedding — not the vertical label. And on multiples the market currently agrees: SEG's public data puts vertically focused software at 3.7x, near the bottom of its category table, while AI-native applications trade at 4.1x against 2.2x for horizontal.

Neither a Borrower nor a Lender Be — Until 2023

The 2020 paper's most optimistic section argued that the banking system was rearchitecting itself around software, that venture debt had grown from under $5bn in 2006 to more than $25bn in 2019, that losses were under 2%, and that revenue-based lending had been competed down from 25–35% APRs to as low as 6–7%. We named Pipe as the emerging market for ARR as a tradeable asset. The direction of that call was right and larger than we imagined. The pricing was wrong by several hundred basis points, and one of our two named examples no longer exists.

Two structural developments we did not anticipate matter more than any of the above. The first is the failure of Silicon Valley Bank in March 2023. SVB held 60–70% of the venture debt market. Nobody replaced it at scale — Hercules, Trinity and Horizon between them added a few hundred million against a $6.7bn book — and the durable result was a repricing of 300–400 basis points, terms shortened from three or four years to two or three, and the long tail of small borrowers simply losing access. Dollar volume is at a record; availability for a $5M ARR company is worse than it was in 2020.

The second is the arrival of private credit at a scale that dwarfs venture debt. Loans outstanding to software companies went from roughly $8bn in 2015 to more than $500bn at the end of 2025 — 19% of all US direct lending. One third of private credit funds have made a SaaS loan; BDCs put over 15% of 2025 lending into the sector. The capital-stack rearchitecture we predicted happened, at sixty times the scale, and it happened in private credit rather than in banks.

That is worth watching rather than celebrating. Roughly $130bn of the $657bn broadly-syndicated software acquisition loan universe now trades below ninety cents on the dollar. A sector that took on a great deal of ARR-backed leverage at 2021 valuations is being repriced for AI risk, and the loans were underwritten against multiples that no longer clear.

What's in a Multiple? The Worked Example, Repriced

The 2020 paper's worked example is the piece most often quoted back to us, so it deserves the most direct correction. We described a founder who raised $2.25m in total, reached $19.2m of run-rate revenue, and sold at 9x revenue for $175m — a 77x return on invested capital. Every step of that is arithmetically fine. The 9x is not available.

Median EV/Revenue by transaction size across 1,325 disclosed software transactions, 2015–2025 (Aventis Advisors). Berkery Noyes finds the same split on separate 2023–25 data: 2.6x for $10–160m deals against 5.4x above $160m. A $19M-revenue company with a $70–110m enterprise value lands in the $50–100M band.

The binding constraint is not the cycle. It is deal size, and the discount has been stable across a decade of transactions. Our 9x sits above the 75th percentile of all software M&A including mega-deals, and well above the 90th percentile for deals of the size we actually produce.

There is a structural reason it is unlikely to re-rate. Private equity software platform buyouts fell to a decade low in 2026 — 41% of deal value against 71% in 2025 — while add-ons roughly doubled as a share. A $10–20m ARR vertical software company is no longer a platform; it is an add-on. Add-on buyers price off accretion math and their own platform's trading multiple. They do not pay scarcity premiums.

The paper's own case repriced: $2.25m total invested, $19.2m run-rate revenue at exit. Returns are gross multiples on invested capital before ownership, fees and dilution, exactly as in the original. The 5.8x case is the 2025 vertical SaaS M&A median; the 7.5x case requires category leadership plus Rule of 40 compliance plus growth above 30%.

This is the section where the honest revision is also the reassuring one. Substituting a realistic multiple takes the modeled return from 77x to somewhere between 34x and 49x. That is a very large reduction and still an excellent outcome, and the thesis is more credible stated that way than it was at 77x. What has to go is the pretence that 9x is a base case.

The Variable the Paper Never Modeled: Time

The original has no treatment of holding period, and that omission is now the largest single risk in the model. Hold periods have stretched from roughly four and a half years to over seven. Median DPI for the 2018 vintage sits at 0.15x against a historical norm of about 0.7x by year eight. Fewer than half of the venture funds Carta tracks have returned any capital at all. The secondary market prices venture NAV at 78 cents on the dollar — which is the market's own opinion of the marks.

Against that, and genuinely in our favor: the sub-$100m software M&A market is liquid. SaaS transaction counts hit an all-time record of 2,698 in 2025, up 28%, and 2,784 on a trailing twelve-month basis through the second quarter of 2026. Private equity and venture-involved buyers are 59% of it. Search funds provide a floor bid at a $16m median purchase price. Vertically focused targets rose to 54% of deals from 46% a year earlier. The exit market for $5–15m ARR software is not frozen. It is busy, and it clears at roughly half the price we assumed.

Revisions: There Are More Things in Heaven and Earth

The 2020 paper opened with a hundred pages' worth of macro argument compressed into five: the industrial revolution, total factor productivity, the shift from manufacturing to professional services, and a prediction about inflation. Those deserve a scorecard too, and one of them is the worst call in the document.

The inflation call is worth dwelling on for one sentence, because the reasoning was not stupid and the outcome was still wrong. We argued a medium-run supply-side proposition and the world delivered a demand-and-supply-chain shock that no productivity trend was going to absorb. That is a real distinction, but we did not draw it at the time, and a forecast that requires a later caveat to survive did not survive. What is quietly interesting is that the mechanism may finally be showing up four years late: core inflation has come down to 2.5% while output grew 2.5% and hours grew 0.4%.

The one macro claim in the paper that has strengthened without qualification is the simplest. Business investment in software rose from $478bn in 2020 to $755bn in 2025 and an annualized $817bn in the second quarter of 2026 — from 2.24% to 2.45% of GDP, compounding at roughly 1.5 times the pace of nominal GDP straight through the multiple compression, the rate shock and the exit freeze. Whatever happened to software as a trade, software's share of national capital formation never stopped rising.

The Framework, Revised

The seven-step framework in the 2020 paper is the part we would change least. Value proposition, market depth, normative metrics, sales efficiency, stabilized cash flow, competitive analysis, team. Each step still asks the right question. What changes is the answers we would accept, and one addition.

GSV 8.0 — Pricing Architecture

We are adding an eighth step, between sales efficiency and stabilized cash flow. The question is simple: what is this contract a claim on? If it is a claim on the customer's headcount, it is exposed to the one line item every customer is actively cutting, and the evidence says it retains thirteen points worse. If it is a claim on work volume, transactions or outcomes, it grows with the customer's activity rather than their payroll.

What We Would Keep Unchanged

Three things in the 2020 paper we would write again, word for word. The first is the primacy of the value proposition, tested by trying to prove the null rather than by admiring the pitch — in a market where building a demo is nearly free, the discipline of asking whether the customer's problem is worth cash is the whole job. The second is the insistence that market size means early-adopter market size, which is more true now that there are more competitors chasing the same innovators. The third is the SaaS return on capital metric — new ARR booked in the period over total capital raised less current cash — which has aged better than anything else in the document, because it is indifferent to multiples.

Sources and Notes on Method

All figures were re-sourced to primary or named-analyst publications between 1 and 5 September 2026. Where two credible sources disagreed we have said so in the figure notes rather than picking one. Six cautions on method are worth stating explicitly.

• The indices are not interchangeable. SEG measures EV to trailing revenue across ~106 companies; Clouded Judgement measures EV to next-twelve-months revenue across ~100; SaaS Capital measures market cap to run-rate ARR across ~63. Time series here stay within one source.
• Two figures from the 2020 paper could not be re-sourced and are withdrawn: the $1.34 median CAC ratio, and the $875k average monthly coastal burn. The 2020 OpenView report contains no geographic burn breakdown and its highest band-level median was $375k.
• Rule of 40 attainment is basis-dependent. The same public cohort scores 15% on EBITDA and 46% on free cash flow. Any statement about attainment has to name the basis.
• Gartner restates prior years at every release, and the segment figures above are each taken from the release closest to the year in question. Directions are stable across vintages; decimals are not.
• Index returns are price returns for all three series, because no total-return series is published for the cloud index. On a total-return basis the ordering is unchanged.
• Free cash flow is stated before stock-based compensation unless marked otherwise.

Principal Sources