The end of software is nonsense

“The end of software is nonsense. This is going to be a layer on top of software.” – Jensen Huang, 16 September 2026

Jensen Huang’s comment at Salesforce’s Dreamforce conference overnight gets to the heart of what I think the market may have misunderstood about the death of software narrative, aka the SaaSpocalypse. The original argument was that if AI could write software cheaply and quickly, enterprise applications would be commoditised and incumbent SaaS vendors would see their economics deteriorate. But what we are increasingly seeing is something quite different. AI is changing how users interact with software, while potentially increasing the value of the underlying enterprise context.

Marc Benioff described the SaaSpocalypse not as the end of software, but potentially “the end of software that makes humans do all the work”. More importantly, he demonstrated an AI architecture that looks very different from the idea that frontier models simply replace existing applications. The emerging model looks more like the image below. 

The important point is that AI models are not replacing the enterprise software underneath it. The model provides the intelligence and may increasingly shape the interface, but the enterprise system still provides the underlying business context. This has ramifications for software investors.

1. Systems of record may be more defensible than feared

Large language models are probabilistic. In simple terms, they generate the answer they judge most likely to be correct rather than retrieving the correct answer from a database. That flexibility makes them powerful, but it also means they can be wrong. Enterprises, however, operate on deterministic data.

A sales manager cannot have an agent inventing revenue numbers. A manufacturer cannot allow an AI system to make up the steps required to configure a production line. Salesforce repeatedly made this distinction at Dreamforce. The model brings intelligence, but the enterprise application supplies the trusted context required to turn that intelligence into useful action. That potentially makes systems of record more important, not less.

2. The fixed UI may matter less, but the interface itself may become more valuable

One of the most interesting demonstrations at Dreamforce was the command centre shown by Salesforce executive Patrick Stokes.

What made it unusual was that it was not a standard Salesforce interface. Stokes said he had personally asked Claude to create it for him, built on top of Salesforce’s underlying data. That points to a very different future for enterprise software. Rather than every user working through essentially the same vendor designed application, each one could ultimately have a personalised interface generated around their own role and objectives.

In that world, the traditional fixed UI becomes less important as a source of differentiation. But that does not mean the software underneath becomes a set of “dumb pipes”. Quite the opposite. The personalised interface is only useful because it sits on top of a trusted system that knows how the business actually works. The architecture could therefore shift from one application UI serving thousands of users to one system of record supporting thousands of personalised AI generated interfaces. That is an important distinction. The interface is not disappearing. It is becoming dynamic and personal. 

3. AI may create a second monetisation layer

The original SaaSpocalypse thesis assumed AI would cannibalise subscription software. Instead, it may add another revenue layer. Historically, enterprise software has largely been monetised through seats based subscriptions. Agents introduce potentially enormous volumes of queries, actions and automated workflows. Salesforce said Agentforce now has more than 30,000 customers and is already processing billions of agent work units.

If that scales, AI could expand the addressable market for software rather than simply reduce the cost of providing the existing product. And if the large scale job losses predicted by some do not materialise, seat based revenue can remain largely intact while agent consumption is layered on top. 

4. Model commoditisation may strengthen the application layer

Another important theme was model choice. Salesforce is happy to work with Anthropic, Nvidia, open models and its own models. It does not necessarily need to predict which model ultimately wins. That matters because, as models become cheaper and increasingly interchangeable, the scarce assets migrate elsewhere: to the proprietary context embedded inside enterprise software. A general purpose model knows an extraordinary amount about the world. It knows very little about the specific customer relationship, inventory position, approval hierarchy or regulatory obligations inside an individual company. That context already lives inside enterprise software.

5. Security and governance could become a substantial new software category

Agents also create a problem humans never created at the same scale. An enterprise may have tens of thousands of employees, but potentially millions of autonomous agent actions. This creates challenges and opportunities for software vendors. 

Salesforce introduced products specifically to manage agent identity, permissions, data security, governance, and cost. I suspect this becomes increasingly important across enterprise software. It is also central to our investment thesis on Microsoft. We have long viewed Microsoft’s advantage as broader than simply productivity software or its cloud business. Its position across enterprise identity, permissions, security, governance, productivity and workflow gives it an unusually strong position as agents increasingly act on behalf of employees. The more capable agents become, the more valuable the control layer around them may become.

Not every software company wins

None of this means enterprise software is immune from disruption. The fixed, vendor-designed UI may become less important as AI lets users generate interfaces around their own role, objectives and workflows. That could put pressure on software businesses whose differentiation rests mainly in the presentation layer, or on narrow point solutions that own little of the underlying enterprise context. But for the stronger enterprise platforms, the opposite may be true.

If personalised AI interfaces sit on top of trusted systems of record, then those underlying systems could become more valuable rather than less. So the shift may not be from software to “dumb pipes”. It may be from fixed applications to intelligent, personalised interfaces built on top of increasingly important enterprise systems. That is a very different conclusion from the SaaSpocalypse narrative earlier this year. Perhaps what we are witnessing is not the death of SaaS at all, but the beginning of a RenaiSaaSnce.