myainote

Foundation models are *partly* commoditizing — but "fully commoditized" isn't settled (frontier capability, cost, reliability still differentiate); the model alone is insufficient as a moat only where a task's capability is widely available and substitutable. Durable advantage lives in four things beyond the model: proprietary data rights and feedback loops, distribution, embedded workflow, and trust. The era's fight is AI-native attacker (unburdened, but moats not yet proven) vs. incumbent.

AI agents can become a single entry point that reaches into the systems of record/engagement/action behind the scenes, putting the interface (engagement) layer at risk. This opens the entry-point battle, with two extremes — the orchestrator (holds user intent and the customer relationship) and the backend pipe (called invisibly, losing brand and pricing power) — but they're ends of a spectrum, not a binary; many middle positions keep bargaining power via data, compliance, or execution rights.

Software pricing's third revolution — after perpetual licenses and per-seat subscriptions — is the shift to usage/outcome billing, triggered because AI agents dissolve the "person" that seat pricing was anchored to. Seat incumbents (Salesforce's Agentforce, Microsoft's Copilot) race to open usage-based curves; a too-slow transition drops them into the gap. Usage-native firms (Snowflake ~$4.7B, Datadog ~$1B/q, Twilio ~$1.4B/q) benefit only if AI lifts metered workloads.

Software's premium rests on near-zero marginal cost → high gross margin, which makes each revenue dollar high-quality. AI threatens it: every feature carries an inference cost that lands in COGS ("COGS creep"), and if it isn't matched by price, efficiency, or retention gains, margin is pressured. Snowflake's ~67% vs. Veeva's ~75% is a *mechanism reference* for compute entering COGS — NOT a causal claim (you can't attribute the ~8 points to compute alone).

Part Four opens with the frame that decides every SaaS company's AI fate: Christensen's sustaining (strengthens leaders) vs. disruptive (topples them) — answered company by company via one test: does AI thicken the moat or thin it? The optimistic scenario: AI is sustaining for quality incumbents, because it makes data and distribution (which they own) more valuable, while models commoditize on many tasks so the model alone isn't a durable moat where capability is substitutable.

SMB SaaS pairs the biggest customer base (tens of millions of small businesses) with the most fragile (high closure rates, price-sensitive). The economics are brutal: low ACV, high churn, and CAC that dwarfs deal size — an enterprise-sales motion is usually uneconomic here. Winners use three weapons: low-touch self-serve acquisition, cross-sell (raises ACV, lowers churn), and embedded finance (the super-lever from software fees to the whole money flow).

Developer tools run on product-led growth (Atlassian's legend): try → love → adopt, viral and cheap. But a gulf splits love from payment — a free version that's "good enough," and enterprise procurement running on different logic than the individual fan. Atlassian (~$1.79B/quarter) still posts a GAAP loss (~$98.39M), driven by SBC (~$408M, ~4× the loss); GitLab (~$264M) is near break-even. Cloud vendors can also absorb open source into hosted services.

Vertical SaaS (pharma, restaurants, construction, government) hides deep compounding: small TAM the giants ignore (gentle competition), industry-specific switching-cost moats, and pricing power that grows with the industry. Veeva, the life-sciences OS, earns ~75% gross / ~30% net margin (~$883M/quarter) via the path tool → system of record → industry OS.

Security is the category where "good enough" is a catastrophe, so buyers resist bundling and tolerate premium pricing — some of the best pricing power in software, driven by fear. Two hidden costs: abnormally high SBC (CrowdStrike ~23% of revenue, flattering the ~$1.31B FCF that stress-tests toward ~$210M) and self-inflicted-incident tail risk that attacks trust itself.

Data infrastructure (Snowflake, Datadog, Confluent) charges by usage, not seats, reflecting a shift where data itself becomes the workflow. The sharp side: expansion is automatic, so NRR often runs high and pricing aligns with consumption. The other edge: contraction is automatic too — customers just use less, and "cloud cost optimization" makes usage procyclical and volatile (high but not automatically guaranteed).