Palantir's 109% Growth Signals SaaS Moat Collapse in 2025

The Data Is In: Integration Beats Apps

The most consequential product strategy signal this week comes from a single comparison: Palantir grew U.S. commercial revenue 109% in 2025 by being the integration and orchestration layer atop enterprise stacks. Meanwhile, Salesforce, SAP, and Adobe — companies with massive product portfolios and entrenched customer bases — managed roughly 10% growth. Palantir doesn't build LLMs. It lets customers use "the best proprietary and open source models" while providing the data integration across Snowflake, Salesforce, SAP, and custom systems that makes AI actually work. The lesson is stark: in the age of AI agents, the orchestration layer is capturing nearly all incremental enterprise spending.

When World View's CEO said 'I'd like to have a pure Palantir-enabled solution without having to have other SaaS tools,' he articulated every enterprise buyer's dream — and every SaaS PM's nightmare.

Foundation Models Are Eating Your Vertical

In the past quarter alone, Anthropic launched Claude Code (competing with every AI dev tool), acquired desktop assistant Vercept (Madrona's own portfolio company), and Cowork shipped a legal review tool competing directly with startups like Luminance. Three verticals, one quarter. Felix Rieseberg of Anthropic was explicit: specialized AI vertical wrappers face compression as general models improve. Their Skills ecosystem — simple markdown files describing API endpoints — means adding a new vertical capability to Cowork is "writing a text file, not building an integration." March Capital's Sumant Mandal posed the existential question at the Montgomery Summit: 'Where does the model end and the application begin?'

OpenAI Frontier Threatens Per-Seat SaaS

OpenAI's Frontier platform is designed to sit above your entire enterprise software stack as a unified intelligence layer. If a CIO can deploy one orchestration layer providing AI across Salesforce, ServiceNow, Workday, and your product simultaneously, per-seat premiums for native AI features evaporate. Salesforce is already pivoting to consumption-based pricing — a multi-billion-dollar incumbent signaling the moat around embedded AI features may be gone.

What's Still Defensible

Ryan Hoover's new defensibility framework is brutally honest about what survives: social graphs, distribution, licensing, data, and hardware. Notice what's missing: software logic, feature cleverness, workflow design — the things PMs typically obsess over. The surviving companies share one trait: products that get meaningfully better the more a specific customer uses them. Runway customizes models with studio film libraries. Paradigm automates PE research with proprietary financial data. Luminance leverages legal case history. Every feature that works equally well for a Day 1 user and a Day 365 user is a feature Anthropic can clone.

The Numbers Your Sprint Review Isn't Showing You

New research reveals that teams using Cursor AI produce 41% more commits — but 38% more reverted commits and 14% more bug fixes in open source projects. Uber's internal analysis shows "power user" developers generate 52% more pull requests — but contains zero quality metrics. No defect rates. No rollback data. No customer impact. CEO Dara Khosrowshahi extrapolated this to a world where Uber stops adding engineering headcount in 5 years. Meanwhile, Uber quietly built close to a dozen internal systems just to manage AI-generated code — a hidden cost invisible in the "52% more PRs" headline.

Your velocity metrics are lying to you. You're shipping faster and fixing more. The net value delivered may be only modestly higher than pre-AI baselines once you factor in rework.

Amazon's Response Is Your Template

After an AI coding agent autonomously decided to 'delete and recreate the environment' for a customer-facing cost calculator — causing a 13-hour outage — Amazon SVP Dave Treadwell summoned engineers to a mandatory meeting. The briefing cited 'novel GenAI usage for which best practices and safeguards are not yet fully established.' Amazon's response: mandatory senior engineer sign-off for any AI-assisted code changes from junior and mid-level engineers. This governance pattern will spread to every serious engineering organization within 12 months.

Anthropic's Speed-Quality Paradox

Anthropic ships 80% AI-generated production code and built Claude Cowork in 10 days — triggering a reported 'code red' inside Microsoft's Office division. But the same organization shipped a textbox bug on claude.ai that destroyed typed prompts during page load for 100% of paying customers — fixed only after public backlash. A race condition from subscription data loading resetting the input field is exactly the kind of async UX pattern AI-generated code struggles with.

The Perverse Incentive Loop

Meta now factors AI token usage into performance calibrations. Low impact combined with low AI usage marks someone as a 'blatant low performer.' Combined with reports of up to 20% staff cuts, this creates a self-reinforcing loop: engineers maximize AI usage to protect their jobs → more code of uncertain quality → measured as productivity → justifies further headcount reduction. The PM's role as quality advocate becomes existentially important.

The Counterexample: Stripe's 1,300 PRs/Week

Stripe's internal AI agents ("Minions") ship 1,300 PRs per week using a hybrid orchestration model that mixes deterministic guardrails with agentic flexibility. This is the difference: Stripe designed for quality from the architecture up. Their system constrains what agents can do, not just measures what they produce.