Is AI Eating Software? The Moat Is Moving Down the Stack

Generative AI is compressing SaaS margins, redrawing moats, and pushing value toward platforms.

Marc Andreessen once argued that software would devour the economy. Today, investors fear the reverse: AI devouring software. The new question is whether the industry’s profits and power migrate away from application vendors and toward the layers below—cloud, models, and operating systems—much as the cloud era kneecapped hardware a decade ago. If that sounds familiar, revisit Andreessen’s 2011 WSJ essay on software.

What’s actually new

Generative AI can perform work that used to justify entire product lines. Drafting emails, designing mockups, summarizing documents, and answering support tickets now look like table stakes rather than premium features. That changes pricing power. Fast.

Value is sliding down the stack. Hyperscalers sell model access and GPU time; foundation-model providers compete on quality and latency; operating systems embed “good enough” assistants. Application software must either differentiate on proprietary data and workflow depth—or watch core tasks get commoditized. There’s no middle seat.

Evidence from the field

Incumbents are repricing to capture AI’s perceived lift. Microsoft sells Copilot as a premium add‑on. Salesforce bundles AI across clouds and is testing usage-based credits. Adobe is weaving Firefly into Creative Cloud to defend share while arguing its provenance tools add trust, not just tricks. Price is a signal here.

AI-native challengers move faster on experience. Canva and Figma collapse learning curves with prompt-based creation and collaborative defaults. In customer service, AI agents now resolve tickets end-to-end, pressuring per-seat models. In coding, assistants turn “number of developers” into a weaker proxy for output. Efficiency cuts both ways.

Economics under pressure

AI features carry ongoing inference costs. Traditional buttons were cheap to serve; large model calls are not. If vendors bundle unlimited AI, margins suffer; if they meter usage, customers scrutinize ROI and switch costs. The bill arrives either way.

Pricing is shifting from seats to outcomes and consumption. Per-resolution, per-generated-asset, and per‑token models align revenue with value but introduce volatility and procurement friction. Finance teams will ask how to budget “bursty” AI workloads without surprise overages. Predictability still wins enterprise deals.

Platform gravity intensifies

Operating systems and suites are swallowing routine tasks. If Windows, iOS, and the big productivity suites handle drafting, summarization, and simple edits out of the box, point solutions must prove sharper accuracy on real data, tighter workflow fit, or clear compliance advantages. “Good enough” at platform scale is a formidable baseline.

The hyperscalers capture more economics when AI usage rises. They monetize model hosting, vector databases, orchestration, and fine‑tuning pipelines that SaaS vendors need to deliver features. As with the cloud, the toll road gets paid first. That dynamic is hard to outgrow.

Strategy choices for incumbents

Three paths dominate: build, buy, or bundle. Build means shipping copilots that actually reduce time-to-value on messy, domain-specific tasks—think CRM forecasting off first-party signals or creative tools that respect licensing constraints. Buy means paying up for AI-native UX and distribution, with antitrust as a speed bump. Bundle means leaning on suites to defend price and consolidate vendors. Speed is the constraint.

Data moats matter again. The strongest AI advantages come from using a customer’s own, high‑grade data—with permission and trust already in place. That tilts the field toward vendors with big footprints and tight hooks into workflows—Salesforce in CRM, SAP in ERP, Adobe in creative. Without data advantage, UX advantages erode quickly.

Regulation, trust, and Europe’s turn

Enterprise buyers now treat AI risk like security risk: contractual promises, audit trails, and content provenance are must‑haves. Vendors that watermark outputs, expose usage logs, and wall off training from customer data reduce adoption friction. Trust is a feature.

Europe’s large software houses—SAP among them—face a double mandate: ship competitive copilots while meeting stricter privacy and AI governance expectations. That can be an advantage if compliance is productized rather than bolted on. It can also slow cycles. The margin for error is thin.

What to watch next

Investors should watch three numbers on earnings calls: attach rates for AI add‑ons, gross margin trends net of AI costs, and churn in seats where AI automation replaces human users. Claimed productivity gains are welcome; unit economics decide the winner. Results beat demos.

Startups should target narrow, painful workflows where incumbents’ generic copilots stumble—legal review on company‑specific templates, industry‑specific analytics, or vertical design systems. Proprietary data partnerships beat param tweaks. Small, fast, and focused still travels.

Limitations and second-order effects

AI will not erase deep software. Complex planning, compliance-heavy workflows, and multi-entity processes still need systems of record and controls. Copilots reduce clicks; they don’t replace audits. For many buyers, “explainability on my data” outranks “flashy generation.” That’s the quiet constraint.

Model churn is real. As vendors swap providers or mix open and closed weights, quality and latency can wobble. Buyers will demand SLAs and graceful fallback paths. Reliability is part of the product, not an ops detail.

Why this matters

• The software profit pool is being re-sliced: platforms and model infrastructure pull more value while application vendors fight for differentiated, data-rich workflows.
• Pricing and trust will sort the field; companies that meter AI sanely and prove safe, outcome-level impact will earn durable multiples.