AI Impacts on Software Power

February 2026

Persistent differential returns are driven by power. Power comes in seven forms:

For anyone familiar with Hamilton Helmer, this will sound familiar. His "seven powers" framework has become a popular starting point for analyzing profitability. In this piece, I aim to analyze the powers of software businesses and how I think they will be impacted by AI. I'll briefly summarize the definition of each power, but for anyone not familiar, I recommend Acquired's interview with Hamilton Helmer.

Scale Economies

Scale economies occur when unit costs decrease with volume. This is a core power of all software businesses. The high fixed cost of building the software can be amortized across many customers at basically zero marginal cost. It is inefficient for each company to build their own version of software systems that are similar across companies, in the same way that it is inefficient for companies to generate their own electricity, design their furniture, or manufacture their own merchandise.

AI has obvious impacts on the scale economies of SaaS. The cost of creating code has gone down tremendously. How does this impact the potential for persistent, differential returns in software businesses? Scale economies protect profits in proportion to the upfront cost. Other players looking to compete with scale would need to be unprofitable in order to catch up. To be unprofitable, they need access to external capital. Access to external capital requires promise of a risk-adjusted return. If there is an outsized risk-adjusted return to be had, capital and competition will enter the market. This will push down prices until the outsized return has been competed away. Therefore, if only scale economies are taken into account, an existing piece of software should generate returns in proportion to the upfront cost of recreating it today.

This conclusion is devastating for software companies relying primarily on scale economies for profits. The reduction in the cost of creating software depends on the specifics, but it's easy to argue a 10x decrease. It might be 100x within a year or two. Earnings should decrease proportionally. To assess the impact on a specific business, imagine that all their existing code became open source. The impact on Meta or Netflix would be zero or close to zero. They are all about content, not code. For a website builder, open source code would erode almost all profits. Clones would emerge and compete away the margin. We won't reach the zero-cost-of-code limit, but if software profits survive that hypothetical, they will survive in a future with a low cost of code as well.

Network Economies

Network economies exist when the value of the product increases as other people use it. Social media is the classic example. People join the same network as people they know. In software, there are also examples of network economies. Slack becomes more useful as others join, both internally and externally. Salesforce integrates with everything, incentivizing both adaption and further integrations. Snowflake's instant sharing feature reduces the cost of data collaboration between customers.

AI coding tools won't impact networks of people or companies. For software application networks, coding tools will change the dynamics. Cheaper code means lower cost for an application developer to support additional networks. This equates to reducing the cost of multi-homing in Hamilton's nomenclature. Lower cost of multi-homing limits the profits that can be extracted. Compare Uber to Windows. The cost of checking Lyft is much lower than the cost of porting a software application to a new operating system. Consequently, Microsoft was able to extract far greater profits than Uber. With AI coding tools, software application networks (e.g. Salesforce's AppExchange or Windows itself) are becoming more like Uber.

There is a likely world in which agents are the primary producers and consumers of digital information. In this world, multi-homing costs radically drop across the board. Agents are patient, rational, and operate at digital speed. They can search 10 sites simultaneously to serve the video you want with the shortest possible advertisement. They will look at 10 ride sharing companies to find the cheapest way to your destination.

If a few companies control the agent layer and dictate default traffic, agents could switch network en-masse. Agent providers could even move traffic into networks they controll. However, anti-trust would likely follow, just like it did for Microsoft in 1998. One company controlling the operating system, the base interface with digital information at the time, didn't automatically imply the same dominance in internet browsing. Courts will certainly be challenged in a similar way with AI agents. Let's hope they reach similar conclusions.

Aggregators (e.g. OpenTable or Expedia) are also powerful because of network economies. They have consumer mindshare and producers therefore participate. If agents are the new way consumers interface with information, an aggregator will need agent mindshare instead. What it takes to gain agent mindshare is largely a function of how many agent providers are out there. If a single company dominates, independent aggregators will have very little leverage. Today, Booking makes almost no profit when somebody visits their site through Google. Dollars flow into their pockets only when people browse the site directly. Will anyone browse sites directly when browsers are replaced by agents? Will there be any power left for independent aggregators?

The same dynamics that led a single player to dominate search exist for chat. Advertising is valuable in proportion to users. Users are valuable in proportion to advertising. Mutual reinforcement and aggressive traffic acquisition spend leads to one player running away with the market. This is bad news for independent aggregators that depend on consumer mindshare and direct browsing of their applications.

Switching Costs

SaaS has a wide range of switching costs. Switching costs capture the productivity lost when switching from one tool to another. Familiarity is almost always there as a switching cost. Even if it takes 15 clicks and two emails to file an expense in the current system, it is initially easier than a two step process that is completely unfamiliar. A second kind of switching cost is engineering. Moving from one cloud provider to another takes a lot of engineering effort. Switching database providers, observability platforms, or experimentation frameworks also come with engineering costs. A third form of switching cost is risk. Some services are so mission-critical that even a small chance of something going wrong is unacceptable. Many banks are famously still running Fortran. IBM reports that 71% of Fortune 500 companies still use mainframes in their operations. This is partly due to engineering costs, but also due to how much revenue and reputation depend on the flawless functioning of these systems. A fourth form of switching costs are actual dollars. Moving large volumes of data can be either inherently or contractually expensive.

The impact of AI on switching costs is highly heterogeneous. AI enables new products to have more intuitive interfaces and more powerful features. This lowers the learning curve and increases the reward once familiar. For engineering switching costs, AI will help significantly. Coding agents excel at the highly structured migration tasks that often constitute the bulk of engineering switching costs. For risk, AI will also have an impact, although smaller. AI can reduce risks and increase rewards but my bet would be that the impact on the decision making of risk-averse legacy businesses will be negligible. When the switching costs are actual dollars, like the cost of moving data, I don't see AI having a significant impact.

Counter Positioning

Counter positioning occurs when new companies adapt business models that incumbents won't copy because it would hurt their existing business. It must make sense to invest in something as a startup, but at the same time not make sense when you have existing cash flows. ChatGPT is a recent example. Despite having the critical innovations, it didn't make sense for Google to cannibalize their sponsored search links with free-form chat.

With a lower cost of software, new, lower-margin business models can start making sense. Marketplaces that take a cut of each transaction will probably be challenged by ones with a fixed cost to participation; per-seat is already being challenged by consumption models.

Custom software will start to make sense at smaller scale. Google already builds all their systems in-house. It makes sense both because they can amortize across their scale and because of their bespoke challenges. It will always be more efficient to share a common system, but all companies have slightly different needs. With AI, the value of filling one's bespoke needs stays constant, but the cost of doing so drops. Companies will transition from shared platforms to in-house systems in proportion to the gains that are to be had from customization. This is important; it will still be cheaper to share a system, even if the provider makes a profit, than rebuilding in-house. The saved dollars won't be enough to motivate in-housing; only when gains from increased control are taken into account will the decision to in-house start making sense.

Cornered Resources

A cornered resource is defined by Hamilton as "access on attractive terms to an asset that can independently increase value". I don't know if there is much to say about cornered resources and software. Cornered resources are things like mining rights or patents. They are a reliable path to profits, but seldom star players software.

Branding

Branding power is defined as a "durable attribution of value to an objectively identical offering due to historical information about the seller". Some software companies have managed to create durable brands, but it's not a core power of the software industry.

AI is already becoming part of branding. While it might increase valuations marginally in the short term, simply branding oneself as AI is unlikely to produce differential returns over the long term. AI will become ubiquitous.

It's unclear what the impact of AI on software brands will be. Companies that fail to integrate AI might come to be seen as outdated. Or an avalanche of slop might place a premium on old school deterministic systems. Overall, I don't think AI will have a large impact on branding.

Process Power

Process power is a set of practices and activities that enable lower costs and/or superior products. What makes particular processes powerful is the time and effort required to copy them. Software companies have a lot of processes that bring them power. How do large teams iterate quickly on a complex system while keeping several nines of reliability? How is feedback from users surfaced, triaged, and integrated into the product?

How does the rise of AI impact process power? AI has the potential to improve or eliminate many processes. At an established company, processes can be hard to alter. Many minds have to align even for minute changes. Companies with a culture of constantly questioning and trying new things will be faster at making their processes AI native.

Part of establishing process power is identifying problems in need of processes. This favors incumbents. While a challenger might have higher speed, an incumbent can make up for it by picking better problems.

AI enables the creation of new, more powerful processes, but also benefits from organizational knowledge about important problems. On net, the value of existing process power decreases. AI gives challengers an ability to leapfrog incumbent processes that didn't exist before.

Conclusion

To think clearly about the impact of AI, it helps to start from power. Power breaks down the concept of a "moat" into a more nuanced set of attributes. The impact of AI on software power is highly heterogenous. Scale economies are likely to take a large hit. Application networks will see a lower cost of multi-homing from AI coding tools. Networks of people or companies might see a significant hit when agents become primary interface for information. But AI is not reducing the cost of multi-homing petabytes of data. Lower engineering costs and more powerful, intuitive products will entice companies to replace their current tools. We are already seeing AI challengers in surveys (Listen), compliance (Delve), and recruiting (JuiceBox), just to mention a few. The new generation of software companies will make their processes AI native, decreasing the process catch-up to incumbents.

While I've said things about specific companies, products, and predictions that you might agree or disagree with, the most important thing is to think structurally about this new situation. Instead of indiscriminately selling software stocks because AI can generate code, one should think about the impacts of AI on the powers that enable long-term profitability for a particular software business.

Edvin T. Berhane