The pace of change surrounding artificial intelligence will undoubtedly affect software companies. The future of software is less clear today than it was a year ago, in our view, which informs our moat rating downgrades. We no longer think software companies will earn excess returns over the next decade with near certainty. We instead view it as probable, and we have even less confidence beyond the next decade.

We are constantly reassessing our moat ratings and assumptions, and we think the current opacity can be clarified as the AI era unfolds. This leaves open the possibility of upgrading the moat ratings on some of the companies we’ve recently downgraded.

Our lower fair value estimates are consistent with lowering the moat duration time horizon to 10 years from 20. Shorter stage 2 durations within our discounted cash flow models (typically now five years, down from 10 years previously), with lower growth now expected in stage 2, drive our fair value estimates lower. Beyond the long-term adjustments, our models are largely unchanged.

Which software companies will fare best as AI unfolds?

Our investment thesis on software is that system-of-record companies like Salesforce CRM and ServiceNow NOW will remain strategic, even as their business models evolve to contain more AI usage on a consumption-based model. Further, we believe that these major software platform companies will be advantaged in distribution since they are already widely adopted, and that over the next few years, customers will likely adopt “agentic” solutions within functional software platforms.

For example, with its strong position in both the customer relationship management and customer service markets, we think Salesforce is well-positioned to capture agentic AI wins. In this context, as AI usage increases, consumption revenue will increase in the mix, and the business model will evolve organically.

Investors looking for metrics supporting our position can focus on retention rates, which have not changed and suggest that moats are largely intact (at least for now), even if the probability and duration of excess returns are less clear. Annual recurring revenue and remaining performance obligation measures continue to outpace revenue growth. We also note that software margins are still increasing, despite fears that margins will be pressured by the need to serve up AI. These data points continue to support moatiness within software.

For all of the hype, AI products are not generating substantial revenue for software vendors, as management teams fear hallucinations and rogue agents. This is supported by disclosures from publicly traded software companies, which generally indicate that AI solutions account for approximately 1%-5% of revenue or ARR, and further by OpenAI’s revenue, which skews heavily toward consumer subscriptions. The flip side of this is that revenue from both OpenAI and Anthropic is rapidly expanding.

The bear case is overblown

We also consider that even if the bear case were true, there is simply not enough inference capacity to replace humans with AI in any meaningful sense at present, nor do we expect there to be in the next five years. We remind investors that the software market generates approximately $1.4 trillion dollars annually.

Further, from a practical standpoint, it is not really possible for a global enterprise to transform its entire software stack and all of its business processes in a few years. Again, if the AI fears manifest, we would expect the software seat count to slow faster than expected, then flatten out and start declining, which we do not think can happen in the medium term.

We also hear fears around “vibe coding” encroaching on proprietary software. While this is possible in some cases, it is highly unlikely to have a substantial impact on the software sector in our view. Enterprises operate within their fields and are great at manufacturing pumps, selling shampoo, or operating restaurants. They are not great at creating and maintaining software. These customers pay for leading-edge solutions that underpin their workflows, are proven reliable and scalable, are well-documented and supported, and are constantly updated based on feedback from hundreds of thousands of customers.

We also find it easy to believe that a developer can create a good application quickly in their garage, but we are highly skeptical that it will be better than the competitive product from a leading vendor, or that they can drive paid adoption by enterprise customers in just a few years. We also note that any increased velocity in innovation will be available to incumbents.

Open-source software is a relevant precedent for the vibe-coding concern. During the internet bubble and for as many years that followed, we saw the rise of open-source software. Open source was supposed to be the death knell for proprietary software because it is free and its advances can be rapidly incorporated into the codebase. Today, there are dozens of proven open-source alternatives to major software applications. Despite this apparent headwind, we still saw the rapid ascent of many proprietary software firms, including Salesforce and ServiceNow. It is not just about code; holistic platform design, support, problem solving, documentation, and other factors matter a great deal.

Demand for software seats likely to decline over the long term

The most reasonable element of the bear case for software, in our opinion, is that over time, as human workers become more efficient based on agentic augmentation, seat count growth will slow and can eventually compress at some point. Assuming the industry evolves along that narrative, AI consumption would approximately offset the lost seat revenue. For context, we certainly did not see sales reps evaporate over the last 25 years due to Salesforce’s automation approach. In the meantime, we can also look at headcount, which continues to increase across functional areas despite a few large-scale headline corporate restructurings that seem to have more to do with poor management than AI. Even for software developers, which is the immediate epicenter of AI usage, hiring bottomed out and has been picking up in recent months.

Of course, these data points are being gathered in real time, whereas we are less confident about what these trends look like in five to 10 years, in accordance with our moat framework. It is certainly possible that these fears may manifest more directly within the next five years and worsen over time.