Over the past year I have advised deep tech and AI companies on raising capital, worked with software companies on how to respond to AI agents, and sat in on startup selection sessions. Each group described the same change from a different seat. Writing software has become cheap, and the businesses built on selling it are being repriced.
This series asks one question about that change. When software becomes cheap to produce, who captures the economic surplus? The short version of the answer is that the barriers to creating software are falling faster than the barriers to building a durable software business. This first piece covers what changed in the market and why. The next three look at enterprise buyers, mid-market software companies, and startups, and a closing piece draws the answers together.
A premium that turned into a discount
Private equity firms buying middle-market companies used to pay extra for technology. In 2023, technology companies in PE-sponsored deals between $10 million and $500 million in enterprise value sold for 10.2 times EBITDA, about 40% more than the 7.2 times paid across all industries. In the first half of 2026, technology sold for 5.7 times against 7.1 times for everything else, in transaction data from GF Data. The premium had become a discount of about 20%, and technology now carries the lowest multiple of any sector in that data set.
Randy Woods charted that reversal under the title "The tech valuation advantage has flipped." These are EBITDA multiples on middle-market technology transactions of every kind, not a measure of SaaS valuations alone. They show that buyers of technology companies stopped paying a premium for them.
Why profitable software companies are being repriced
Software valuations had been falling for years before AI agents arrived, and several forces contributed. The extraordinary valuations of the pandemic era were bound to normalize. The Federal Reserve raised its benchmark rate from near zero in early 2022 to a peak of 5.25% to 5.50% in 2023, which hit long-duration growth assets such as software hardest. Revenue growth slowed and customers renewed for less, and competition, including from AI, intensified. Listed software, measured on revenue rather than EBITDA, tells a similar story. The median public cloud software company traded at just under 18 times revenue in 2021 and at 4.1 times forward revenue on January 30, 2026, about the lowest level in a decade.
The strongest objection to this series is that slower growth and higher rates explain most of that decline, with AI as a footnote. That objection carries real weight. Rates do not hit every industry equally, since long-duration growth businesses are more sensitive to them than mature ones, and the mix of technology companies changing hands can move a sector's average multiple. The reversal relative to other sectors is consistent with an additional reassessment of technology businesses, although the transaction data alone cannot isolate how much of it reflects AI rather than growth expectations, deal composition, or financing conditions. Software multiples also kept falling into 2026 after the Fed began cutting in September 2024. The stronger claim is narrower. AI adds a new risk to future pricing power and revenue durability, on top of the earlier financial correction.
The operating numbers explain why investors grew nervous. Median revenue growth fell from 27% in 2021 to 10% in 2025, and net revenue retention slipped from 117% to about 108%. Margins moved the other way. Median EBITDA margin rose from −6.8% to 9.1%, and 93% of software companies in the S&P 500 beat profit estimates in the fourth quarter of 2025. Investors sold profitable companies because they doubted how long the revenue would last.
Then the doubt acquired a date. Anthropic released Claude Cowork on January 12, 2026, and added plugins on January 30. On February 3, software stocks lost about $285 billion in market value in a single day, and a Jefferies trader named the event the SaaSpocalypse. The SEG SaaS Index fell 25.7% in the first quarter while the S&P 500 fell 4.8%. The IGV software fund kept falling until April 10, about 37% below its September 2025 high, before recovering part of the loss over the summer.
The companies themselves are still operating. Recent acquisitions show how far the price of their future revenue has fallen since the peak. The declines are measured from valuations set in 2021 and 2022, so they include the end of the pandemic boom, and they show the size of the repricing more than its cause. HashiCorp agreed to its sale in April 2024, before coding agents were a factor.
Company | Peak value | Deal value | Decline | Buyer | Status |
|---|---|---|---|---|---|
Miro | $17.5B equity (2022 round) | ~$1.79B equity | ~90% | Bending Spoons | Announced Sep 2026, pending |
Airtable | $11B equity (2021 round) | ~$2.25B equity | ~80% | Bending Spoons | Announced Aug 2026, pending |
Confluent | $93.60 per share (Nov 2021) | $31 per share | ~67% | IBM | Announced Dec 2025, completed Mar 2026 |
HashiCorp | $97.60 per share (Dec 2021) | $35 per share | ~64% | IBM | Announced Apr 2024, completed Feb 2025 |
Each product on that list still has paying customers. What fell was the price a buyer would pay for the revenue still to come. Deal terms come from the Miro announcement, Fortune, Financier Worldwide, and TipRanks.
AI agents threaten that future revenue in two ways. Per-seat pricing assumes revenue grows with a customer's headcount, so when one agent does the work of five people, the renewal can shrink with the seat count. Gartner put it directly in July, saying agents break "the link between user growth and revenue growth for many enterprise software vendors." Users also increasingly ask an agent for an answer instead of opening the application. When an agent becomes the main user of a product, the product either becomes the infrastructure the agent depends on or something the agent can route around.
What became cheap and what stayed expensive
Three costs are falling at different speeds. Separating them is the framework for the rest of this analysis.
Production cost is falling first and fastest, although how fast depends on what is being built. On a narrow, well-defined task, developers using GitHub Copilot built a small web server 55.8% faster than a control group in a 2023 experiment. On mature codebases the picture is less clear. In a 2025 randomized trial by METR, 16 experienced open-source developers took 19% longer on real tasks in their own repositories when they could use AI tools, while believing the tools had sped them up. METR's 2026 follow-up could not produce a reliable estimate, partly because 30% to 50% of developers held back tasks they did not want to do without AI, and METR judged that developers are likely faster now. Across whole organizations, Google's 2025 DORA survey of nearly 5,000 technology professionals found 90% using AI at work, with AI adoption now associated with higher software delivery throughput and still associated with lower delivery stability. Read together, these studies suggest that AI makes writing code cheaper faster than it makes running software cheaper. A prototype or a narrow tool can take days or weeks instead of months. A production system still has to be tested, secured, and maintained.
Replication cost may be falling next. Copying a competitor's product used to require months of study, and a newer class of tools aims to shorten it. REA, short for Reverse Engineer Anything, is an open-source project that connects coding agents to reverse-engineering tools so an agent can trace how an existing program works and write a replacement. Its public demonstrations cover single features in small programs, so for now tools like it show a direction of travel, and nothing in the valuation data can be traced to them. Exposure also depends on where the code runs. Software installed on a customer's device hands over a copy that such tools can take apart, while software sold as a service keeps its business logic and data on the vendor's servers. Cheap production hurt SaaS more than most software, because its features are easy to rebuild. Cheap replication hurts it less.
Switching cost is falling slowest. A company might reproduce the features of Salesforce cheaply and still find it expensive to move its sales operation off Salesforce, because the data, the integrations, the trained staff, and the history of every customer interaction all have to move too. Reverse engineering can make software easier to understand without making it safe or legal to replace.
Around those three costs, several things have stayed expensive. Data and the history behind it are hard to move. Knowledge of a domain, such as how a regulated process works or what a contract allows, takes years to build. Accountability costs money, because someone has to answer for what software does when it acts on a customer's behalf. Customer acquisition and distribution have not become cheaper in proportion to code. Physical products such as chips and robots depend on manufacturing, certification, and capital that cheaper code does not replace.
Who keeps the surplus
Software has been declared dead before, by open source, by no-code, and by the financial crisis, and spending kept growing each time. This time spending is still growing too. Gartner raised its forecast for 2026 software spending to $1.44 trillion in April, up 15.1% from 2025, and its analyst George Brocklehurst described the shift as "less an apocalypse and more of a metamorphosis." What changed is how much of that spending investors believe vendors will keep.
Cheaper software creates a surplus. A feature that once took a team months can now take weeks, and someone ends up keeping the difference. Whether the repricing is a temporary correction or a lasting shift depends on who that is. Five groups are competing for it.
Software vendors keep some of it through lower engineering costs. The rise in median EBITDA margin from −6.8% to 9.1% shows vendors holding part of the surplus today. The fall in multiples shows investors doubt they can hold it once renewals come due.
Customers take it through lower prices, custom tools, and bargaining power. In a survey by the pricing consultancy Cruxy, 82% of SaaS CEOs said customers had already asked for AI-related price cuts. Software budgets are still growing, so customers have taken relatively little of the surplus so far.
AI infrastructure and model providers collect it through inference, model access, and compute. Enterprises spent an estimated $37 billion on generative AI in 2025, more than three times the year before, including $12.5 billion on model APIs. Anthropic, OpenAI, and Google together handled 88% of enterprise model usage.
AI-native companies capture it by selling outcomes instead of tools. Startups earned 63% of enterprise spending on AI applications in 2025, by the same estimate, nearly two dollars for every dollar earned by incumbents.
Systems of record and data owners can keep some of it by becoming the infrastructure agents have to consult, provided they charge for more than raw access. In an October 2026 Gartner webinar, VP Analyst Patrick Connaughton argued that putting AI interfaces on top of a system of record can deepen a customer's dependence on it, and that the model labs show little interest in replacing those systems, which he called "hard, complicated" and "low margin."
The evidence for each group sits at a different stage of the same change.
Stage | What happens | Evidence available today |
|---|---|---|
Cost reduction | Software becomes cheaper to create | Coding-agent capabilities and productivity studies |
Bargaining and competition | Buyers, vendors, and providers try to capture the savings | Renewal negotiations, pricing changes, internal builds |
Realized value capture | Savings show up in profits, lower spending, or new revenue | Margins, realized prices, revenue shifts, ownership economics |
The available data show where the pressure is building more clearly than where the savings have landed. Higher vendor margins, requests for discounts, and growing AI infrastructure spending each measure part of the transition, but none of them establishes how much of the surplus any group has captured. Most of the evidence in this series covers the first two stages, with suggestive evidence of the third. The distribution will become clearer as enterprise contracts renew and AI operating costs settle, and the closing piece of this series sets out the signals that would show which way the surplus is moving.
Where investors pay a premium now
Investors still pay a premium. They pay it to fewer companies and for different things.
The partial recovery in software stocks since April has been uneven. The ten most valued public software companies rose from 11.7 times to 20.8 times revenue between April and July 2026, in Meritech data, while the typical company moved from 3.2 times to 3.8 times, still well below its 2019 level of 8.7 times. Listed business software points the same way, with generative AI companies at a median of 4.7 times revenue in August against 2.5 times for vertical SaaS.
McKinsey studied 471 companies that had received PE equity or debt financing and sorted them by how deeply they use AI. Companies using AI opportunistically had a median multiple of 13 times revenue, those using it to run operations 14 times, those with AI built into the product 20 times, and those that had built a new AI business 31 times. In software the top two groups sat at 24 and 33 times. The sample covers relatively high-valued, privately financed companies with revenue multiples between 5 and 300 times, so its figures are not typical buyout valuations, and the top group has only 21 companies. What the study shows is an association within that sample. It does not show that adding AI to a product raises a company's multiple, because stronger companies may simply be more likely to adopt AI.
Read together, the evidence suggests a pattern. Investors are paying more for companies whose customers buy AI from them and for the few companies they expect to capture the surplus, and less for the typical vendor whose value sits in features that became cheap to build.
A plausible shape for the industry
In September, Mike Vernal of the venture firm Conviction described where this could lead in a post titled "Moats & the Barbell-ification of Software." Drawing on a comparison Ben Thompson made, he points to newspapers after the internet. National brands grew, solo writers built audiences around a niche, and many regional papers in the middle closed.
Applied to software, frontier labs and large platforms sit at one end and add features with fleets of agents. At the other end, small teams and in-house engineers with coding agents build what a startup used to sell. The independent, mid-sized software company sits in the middle, where multiples compressed hardest.
The barbell is a model of where software could go, not a law it has to follow. Platforms could stumble, a new generation of mid-sized companies could form around agents, and the small end could fill with far more businesses than the old middle ever held. The rest of this series looks at the corporate buyers, the mid-market vendors, and the startups living through that change, and a closing piece draws their answers together.
Questions boards and investors need to answer
Three questions separate a repriced software business from one that holds its value.
How much of the company's revenue depends on work that coding agents now do cheaply?
What does the company hold that stays expensive to copy or to move, whether data, domain rules, accountability for agent actions, or a physical product?
Do customers buy AI from the company, or does AI mainly lower the company's own costs?
Read next in When Software Gets Cheap
What the Corporate Buyer Decides (Enterprise). The next piece moves to the customers on the other side of the contract, whose choices about what to build, buy, and renegotiate decide which vendors keep their value.
Also in the series:
What Still Stands in the Middle (Mid-Market): How vertical SaaS vendors can keep their customers and still lose the value of the workflow.
The Squeeze on Software Startups (Startups): Why software entrepreneurship and venture-backed software entrepreneurship may be diverging.
Who Keeps the Value When Software Gets Cheap (Closing): The answers drawn together, and the signals worth watching.
Selling the Finished Job (Services): When software gets cheap, does the money move to the work the software does?
Sources
The tech valuation advantage has flipped, Randy Woods, data from GF Data via Forvis Mazars and Mercer Capital
Middle Market Transaction Update, Fall 2026, Mercer Capital
The value of software revenue may have finally stopped rising, TechCrunch, September 17, 2021
Clouded Judgement 1.30.26, Jamin Ball
Software vs. AI Q1 2026: SaaSpocalypse Examined, Long Angle
The SaaSpocalypse Is Over, SaasRise, citing Meritech and other market data, July 2026
SaaSpocalypse analysis, Dealroom, updated September 11, 2026
Bending Spoons agreement to acquire Miro, SEC filing, September 10, 2026
Bending Spoons buys Airtable at a $9 billion discount, Fortune, August 5, 2026
IBM agrees $11bn Confluent deal, Financier Worldwide
IBM finally closes $6.4B HashiCorp deal, TipRanks
Confluent and HashiCorp stock price history, Macrotrends
Gartner Forecasts Worldwide IT Spending to Grow 13.5% in 2026, Gartner, April 22, 2026
Gartner Says $234 Billion in Enterprise Application Software Spend Is at Risk from Agentic AI, Gartner, July 1, 2026
The Future of Enterprise Applications: Navigating AI, SaaS Disruption and Vendor Lock-In, Gartner webinar with Patrick Connaughton, October 9, 2026
The Impact of AI on Developer Productivity: Evidence from GitHub Copilot, Peng, Kalliamvakou, Cihon, and Demirer, February 2023
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, July 10, 2025
We are Changing our Developer Productivity Experiment Design, METR, February 24, 2026
Announcing the 2025 DORA report, Google Cloud, September 23, 2025
REA: Reverse Engineer Anything, GitHub
Reverse Engineer Anything: AI REA guide, Kingy AI
SaaS chiefs expect pricing shift as AI threatens sales, CFOtech UK, April 16, 2026 (Cruxy survey)
2025: The State of Generative AI in the Enterprise, Menlo Ventures, December 2025
Beyond productivity: How AI creates value in private equity, McKinsey, June 2026
What happens as software development costs move toward zero, DigitalToday, on Mike Vernal's "Moats & the Barbell-ification of Software," September 2026