Former Microsoft Windows president Steven Sinofsky is pushing back against Wall Street's assumption that large language models will commoditize standalone SaaS companies into irrelevance.
His thesis: AI creates demand for more software, not less. "Automating with software is HARD," he told Fortune, pointing to how AI is moving products up the stack through agents and personalization rather than replacing them.
The numbers back a software expansion story. AI coding tools now generate 25% of Google's new code. Amazon saved $260 million and 4,500 man-years on legacy updates using AI. Developer adoption is nearly universal: 76% are using or planning AI tools, with 92% of US developers already using AI coding assistants.
What this means in practice: The cloud analogy holds. When cloud platforms emerged, fears that Amazon and Microsoft would build everything proved wrong. Apps became more essential, not less. The AI market is tracking the same trajectory, projected to hit $1.8 trillion by 2030 compared to cloud's $2.3-2.6 trillion.
The enterprise procurement question is which layer wins. AI-native SaaS companies argue they're building fundamentally different architectures. Legacy vendors counter that enterprises won't rip-and-replace working systems for unproven AI products. Both are partly right.
Stratechery offers a competing view: AI's real impact is top-down replacement of workers (41% of enterprise time spent on low-value tasks), shifting buying decisions to C-suite cost-cutting initiatives rather than organic app growth. That's a different growth model than Sinofsky's enhancement thesis.
The trade-offs matter for technology leaders. AI-native products may offer better architecture but bring integration risk and vendor uncertainty. Legacy vendors with AI bolted on bring familiar procurement but potentially awkward implementations. History suggests both will coexist, with the winners determined by execution, not positioning.
Worth noting: Sinofsky himself acknowledges historical skepticism around developer tools (syntax coloring, Fortran compilers) that proved essential. Some platforms do try to build everything. The pattern is that apps endure, but which apps is still being determined.