Software investors have spent years paying for a fairly simple promise: a company owns a useful slice of work, wraps that work in a product, and can keep charging because replacing the workflow is painful.
Frontier AI complicates that promise because the unit of disruption is no longer only the application. It can be the boundary around the application.
Reuters reported that U.S. software shares weakened on Sep. 8 as investors reacted to another jump in AI capability. Salesforce, Intuit and ServiceNow were among the names under pressure. The useful interpretation is not that one model launch magically erased three businesses. Markets are much better at panic than prophecy. The useful interpretation is that investors are trying to price a new form of product compression.
The model is moving up the stack.
OpenAI describes GPT-6 Astra as capable across software engineering, browser use, computer use and professional work. Those categories matter because they overlap with tasks that used to require a person moving among several specialized applications. If the model can operate the browser, write and inspect code, reason over documents, navigate interfaces and complete longer sequences, then some software products stop being destinations and become tools the model calls on the way to an outcome.
That does not make the tools worthless. It changes where their value has to live.
Features are easier to compress than systems of record.
A thin layer that reformats text, summarizes a document or automates a predictable click path is exposed when a general model can perform the same operation directly. A product with proprietary data, embedded permissions, regulated workflows, switching costs, audit history, deep integrations or a trusted system of record has a different defense.
This distinction matters more than the fashionable argument over whether “AI replaces SaaS.” SaaS is not one thing. A CRM database with years of customer history is not the same economic object as a clever text box attached to an API. A payroll system with compliance obligations is not the same object as a standalone drafting assistant. The market is beginning to charge different risk premiums to products that previously lived under the same software umbrella.
Incumbents get to use the new models too.
Disruption stories often freeze the incumbent in place. That is convenient for storytelling and terrible for forecasting. Salesforce, ServiceNow, Intuit and other large vendors can buy or integrate frontier models, expose their own data through agents, redesign interfaces and automate work inside the systems they already control.
The threat is therefore not simply “model versus software company.” It is a race over who owns the context, permissions, data and customer relationship once the interface becomes more conversational and the execution layer becomes more autonomous.
The valuation question is about retained friction.
Historically, software could monetize friction. Users needed a specialized interface, specialized training and a sequence of manual operations. AI removes some of that friction. That is good for users and uncomfortable for any valuation that assumed the friction would remain billable forever.
The durable company will have to prove that what customers pay for is more than the effort required to operate the interface. It may be the data model, compliance layer, transaction network, proprietary corpus, integration graph, distribution, liability structure or simply a product trusted enough to sit in the path of expensive decisions.
That is a healthier test than asking whether every software company is about to die because a model benchmark went up.
The Sep. 8 selloff is evidence of repricing, not proof of extinction. Frontier models are forcing investors to separate software that owns a durable system from software that merely owns a temporary interface. The companies that survive this transition will not necessarily be the ones with the flashiest AI feature. They will be the ones whose value remains after the model absorbs the easy part.