The Squeeze on Software Startups

Over the years I've worked with several accelerators, incubators, and early-stage investors. This fall, the conversations between pitches kept returning to one question. What happens to software startups when software is nearly free to write? What follows comes from those conversations. It describes what a set of early-stage investors are seeing and saying, not a survey of the venture market.

The earlier pieces in this series looked at the question from the market, from corporate buyers, and from PE-backed vertical SaaS. This piece looks at it from the seed stage, where new software companies usually start. The short answer is that many early-stage investors have stopped funding the kind of software company most founders know how to build. The longer answer is that software entrepreneurship and venture-backed software entrepreneurship may be starting to diverge.

Haves and have-nots at the seed stage

A view heard often in these rooms is that writing software has become about as hard as writing a good essay. At the seed stage, that view sorts companies into haves and have-nots. The haves hold something software alone does not reproduce, such as a chip, a robot, proprietary data, or a regulated position. The have-nots are horizontal software companies that, ten years ago, would have raised a seed round and grown into independent mid-sized businesses.

The film industry offers a parallel. Matt Damon has said he could not raise $23 million for Behind the Candelabra, a drama that would once have been easy to finance, and it ended up on HBO. Data from Stephen Follows suggest the change showed up more in shrinking budgets than in the number of dramas made. Financing for the middle got scarce while audiences kept watching. Something similar is happening to software startups, with seed investors in the financier's seat.

The seed stage feels this first because a new company has no lock-in yet. A PE-backed vertical SaaS company at least holds its customers' data while it works out what to do next. A seed-stage company has mainly its product, and the product is the part that got cheap.

Tools buyers can rebuild went first

The first software startups under pressure are the ones selling capabilities their customers can now reproduce with coding agents. Developer tools were an early example because their buyers are engineers. For a decade, developer tools were a reliable seed bet. Founders were their own customers, adoption ran bottom-up, and a good tool spread through engineering teams without a sales force. Several early-stage investors in these conversations say their earlier winners were dev tools companies, while their recent winners lean toward robotics and advanced manufacturing. Some say they have not funded a dev tools company in six months.

The usual explanation is simple. The buyer of a narrow developer utility is an engineer, and the engineer is now well equipped to replace it. With a coding agent, a passable version can take an afternoon. Product quality has also weakened as a signal. A well-designed product used to suggest the founder had real insight, and investors now see it as about as hard to make as a good homepage.

Other developer tools may gain. The more code agents write, the more there is to test, secure, monitor, deploy, and govern. AI changes which development problems stay hard. Categories outside developer tools are likely to feel the same pressure in roughly the order their buyers can build for themselves.

Copying is getting cheaper as well. A fast follower used to need months to study a competitor's product and rebuild it. Coding agents shortened the rebuild, and reverse-engineering tools such as REA (Reverse Engineer Anything) aim to shorten the study by letting an agent trace how a shipped application works. Its public demonstrations still cover single features rather than whole products, so it shows a direction more than a present threat. Whatever a startup ships to a user's device can be examined, and features on their own buy less time than before.

What investors in these conversations are saying

Several perspectives came up repeatedly this fall. None is new on its own. Heard together, from people who write first checks, they describe the squeeze at its starting point.

Interest has moved toward deep tech. Several investors describe deep tech companies, especially in semiconductors, drawing far more interest than SaaS and dev tools. At one recent investor matchmaking event, a deep tech company had more than 20 meeting requests, while a software company had fewer than five. That is a single event, and it is the only evidence here of a preference for physical deep tech.

Capital is concentrating. A separate trend shows up in the market data. AI companies took 86% of US venture dollars in the first half of 2026, and rounds of $100 million or more took 87.5%. Those figures do not separate foundation-model companies from robotics, chips, or manufacturing, and much of the money is likely going to model developers. What they show is capital flowing into a small number of very large AI bets.

Platforms absorb features. Coding agents from the frontier labs and GitHub, data platforms, and hosting providers keep turning point tools into features. Investors remember the generic agent startups of 2024 that faded once the labs released agents of their own. Large companies also build more and acquire less, which narrows a common exit.

Enterprises build. Investors describe selling new software into large enterprises as very hard. As the enterprise piece in this series describes, forward-deployed engineers from the labs, data platforms, and consulting firms are already inside large companies.

Theses go stale quickly. Several pitches this fall described products that would have been fundable in the spring. By the fall, the frontier models or the customers' own engineers had caught up.

The growth bar went up. Some investors in these conversations now describe expecting nine to ten times annual growth before a Series A, a bar that would have seemed extraordinary a few years ago. They also describe companies with a few million dollars in annual recurring revenue struggling to raise one because of foundation-model risk.

Where seed money still goes

Software still gets funded. The companies that get through have found a position where the squeeze is weaker. Five patterns stand out.

Buyers who cannot build. The software that seems to work now goes to non-technical, non-enterprise buyers the labs cannot easily reach. Investors call this diffusion, meaning bringing AI agents into traditional industries the labs are not yet in.

A structural reason the giants will not copy the product. Data that cannot leave a regulated site, a licensed-practitioner rule, or deep legacy integration hold up under questioning. A claim that an incumbent is too slow does not.

Owning the outcome instead of selling the tool. Executives increasingly want outcomes rather than capabilities, much as companies outsource advertising to an agency. Founders with a real technical edge are pushed to capture its value directly, for example by trading energy markets with a weather model instead of selling forecasts to utilities.

A model of its own, when it earns its cost. For tools that sit close to the labs, some investors push founders to stop wrapping someone else's model. That advice holds under specific conditions. The company has training data others cannot get, a specialized model clearly outperforms general ones on the task, the advantage can survive the next frontier release, and owning the model improves the product's economics or control. Without those, a proprietary model adds cost and risk without adding defensibility.

Mission-critical checks. Few buyers want a vibe-coded control on a system where a failure makes the news. A buyer who needs an independent second check has a reason to pay a vendor.

Each pattern is a way to capture part of the surplus that cheaper software creates. A company that sells outcomes keeps the value of the work itself, which a buyer cannot rebuild in an afternoon. Startups earned an estimated 63% of enterprise spending on AI applications in 2025, up from 36% the year before.

Physical deep tech is less directly exposed to software commoditization, because much of its value sits in physical assets that cannot be produced by generating code. AI is lowering its design and firmware costs, while manufacturing, certification, and capital requirements still stand in the way of anyone copying it. Reverse-engineering tools can read firmware, but they do not copy a chip or a robot. Investors with semiconductor backgrounds are actively looking for more of these companies, and many describe applications built on top of AI as a hard pitch right now.

Two kinds of software company

For decades, building software was expensive, so a product only paid off across a large market with a broad customer profile and a standardized feature set. Customers adapted their workflows to the software because customizing it cost too much. When building gets cheap, that math changes. A product can serve a small, specific group of customers and still cover its costs, and a founder can start with one narrow problem, solve it well, and expand wherever the data leads. Whether it works is an open question. If the buyers cannot build, in a market no lab or platform will bother with, a founder keeps the savings that cheap code creates instead of passing them to a platform.

Starting narrow also raises a question many founders skip. A narrow market rarely produces the returns a venture fund needs, especially against the growth expectations some investors now describe. A company that serves a few thousand customers well can be profitable without that growth. For many of these businesses, bootstrapping, revenue-based financing, or a small angel round fits better than venture capital. Venture money commits a company to growing from a narrow product into a large software business, and that path runs through the middle that investors are now avoiding. These companies also have a different kind of exit available, since acquirers that specialize in small vertical software businesses, the model Constellation Software built, buy profitable niche companies that venture funds would decline.

The strongest objection to this piece is that cheaper development may expand startup formation instead of shrinking the opportunity. That objection may be right, and it does not conflict with the squeeze. Both can happen at once. The software industry may produce more viable businesses and fewer conventional venture-scale software companies. Founder data point in that direction without settling it. Carta reports that about 36% of startups founded on its platform in 2025 were led by solo founders, up from 31% in 2024, and that the share has roughly doubled over ten years. Those figures describe who is starting companies. They say nothing yet about how those companies fare. A rising share of solo founders fits the idea that building got cheaper, but no data yet show more small software businesses becoming profitable or lasting, so this part of the argument is preliminary. If it holds, a shrinking middle in venture-backed software would not mean fewer software companies. It could mean many more small ones, serving narrower markets, owned by their founders.

Staying small carries its own risk. A narrow product that was cheap to build is also cheap for the next founder to copy. What protects it is what accumulates over time, such as the customers' data and workflows, standing in a small community, and detailed knowledge of how that niche operates.

What it means for the middle

Gartner sees a longer path for the few venture-backed companies that get through. In an October 2026 webinar, VP Analyst Patrick Connaughton noted that two or three vendors hold most of the market in each major enterprise application category, which makes entry hard. He compared today's AI-native startups to Salesforce in 1999 and expects some to become the next large vendors after 2030. His test for an AI-native product is simple. Remove the AI, and the product stops working.

For the middle, the evidence so far suggests fewer independent software companies will reach mid-size on venture funding. Some seed-stage software companies may grow into services firms that use their own tools. Others may sell to larger companies early. Many investors in these conversations expect this year's horizontal tool startups to end as acqui-hires or close.

These are perspectives from one season of conversations, and the claim is narrower than it may sound. At the seed stage, a horizontal tool whose buyers can rebuild it is now one of the hardest pitches in software, and the categories next to it are moving the same way.

Questions founders need to answer

Four questions come up early, three in the pitch and one before it.

Why won't the buyer build this, and why won't a fast follower copy it? The answer needs to come in the first two minutes, and it has to be structural.

Which part of the company stays expensive to copy once code is cheap? Data, domain knowledge, accountability, and physical product qualify, and a polished interface usually does not.

What happens after signup? As software becomes a weaker moat, habit formation after signup becomes the evidence that distribution worked.

Does the company need venture money at all? A narrow market served well may be a better business than a large one the founder has to chase.