Defensibility in the AI era
The moats that hold, the ones that break, and the ones that have not been built yet.
The thesis that networks, proprietary data, and communities are the only durable moats is directionally right but dangerously incomplete. The original framework correctly identifies that AI commoditizes execution and that cloneable products die fast. But pressure-testing it against 18 months of real-world carnage — Chegg's 99% stock collapse, Stack Overflow's question volume falling 97%, the early-2026 SaaSpocalypse erasing ~$2 trillion in B2B software market cap — reveals that the framework understates the fragility of data and community moats while entirely omitting several moat types that have proven more durable. A revised taxonomy must incorporate regulatory barriers, workflow embedding, process power, and AI-native switching costs alongside the original three, and must rank these by both durability and accessibility for founders who need to know what to build first and what to build toward.
I. Where the thesis holds: networks, data, and communities under pressure
The strongest pillar of the original thesis is network effects. NFX's landmark study of 336 companies that reached $1B+ valuations found that network effects account for roughly 70% of all value created in technology — a figure independently confirmed by Morgan Stanley. Hamilton Helmer's framework ranks Network Economies among the most durable of the 7 Powers, and Virta Ventures' AI-era adaptation argues they've actually strengthened because AI accelerates onboarding and liquidity accumulation for platforms that already have critical mass.
The evidence at the company level is concrete. Morningstar upgraded Airbnb's moat rating from narrow to wide in 2025, explicitly citing its network advantage and AI positioning. With 5M+ hosts, 8M active listings, and ~90% organic/direct traffic, Airbnb's two-sided marketplace has proven resilient to a decade of well-funded competitors. Veeva Systems maintained dominance in pharma CRM despite Salesforce's direct assault — 9 of the top 20 pharma companies committed to Veeva's Vault CRM because industry-specific network effects and data standards outweighed Salesforce's distribution advantage. Bloomberg's terminal network — 350,000 subscribers at $31,980/year, generating roughly $15 billion in annual revenue — combines a financial data network effect with a professional messaging community that functions as Wall Street's default communication layer.
Proprietary data holds as a moat, but only under specific conditions. The thesis overstates data's general defensibility while understating the conditions required for it to work. The strongest data moats share four properties: the data is exclusive (not scrapable), continuously refreshed (not static), embedded in workflows (not just stored), and compounds with usage (exhibits a genuine flywheel). Bloomberg exemplifies all four — proprietary real-time financial feeds that traders access through deeply integrated workflows, with AI features (BloombergGPT, trained on 363 billion financial tokens) that make the data more valuable and the switching costs higher. Palantir's ontology framework, covering 3,400+ patents and generating 128% net revenue retention, shows how data embedded in operational decision-making creates a self-reinforcing flywheel. Flatiron Health's oncology data — pulled from 265+ cancer clinics — was worth $1.9 billion to Roche precisely because it was regulatory-grade real-world evidence that couldn't be synthetically generated.
Community moats are the weakest of the original three, but the best examples are genuinely durable. Figma's design community — with its plugin ecosystem, template marketplace, and Designer Advocate program — creates ecosystem lock-in comparable to Salesforce's AppExchange. Notion runs 80+ self-organized communities and 30+ monthly events with only three full-time team members, generating user-created templates that make the product harder to leave. HubSpot's partner/agency ecosystem creates distribution defensibility: as SaaStr's Jason Lemkin observed, "Most of us in CRM will still just pick HubSpot or Salesforce today." A First Round survey found 28% of founders described community as their moat and critical to their success. In consumer AI, a16z documented how ElevenLabs' "Gibberlink" moment — two AIs spontaneously switching to a more efficient communication protocol — generated massive organic distribution through community evangelism, replacing traditional PR entirely.
The strategic frameworks broadly corroborate the thesis. Helmer's 7 Powers maps cleanly: Network Economies (networks), Cornered Resources (data), and community-driven Switching Costs (communities) represent three of the strongest powers. a16z identifies proprietary data as the "fruits of the walled garden." BVP's State of AI 2025 report argues that "memory and context are the new moats" — an evolution of the data thesis. Andrew Chen's "Revenge of the GPT Wrappers" essay argues that as models commoditize, the competitive axis shifts to distribution and network effects, exactly as it did when "database wrappers" (CRUD apps) competed in Web 2.0. The most defensible Web 2.0 companies weren't distinguished by their database technology but by their network effects.
II. Where the thesis breaks: five structural weaknesses
Data moats are far more fragile than the thesis implies
The most authoritative critique comes from within a16z itself. Martin Casado and Peter Lauten's "The Empty Promise of Data Moats" dismantled the standard VC data moat argument with three specific claims, all well-evidenced. First, most "data network effects" are actually just data scale effects with severe diminishing returns — their analysis of Eloquent Labs showed that 20% of data collection effort yields ~20% intent coverage, but the curve asymptotes at ~40%, meaning additional data adds zero marginal value. Second, data goes stale — "streets change, temperatures change, attitudes change." Third, the moat erodes as it grows because competitors race to catch up while the incumbent's marginal gains shrink.
Synthetic data has accelerated this erosion dramatically. Frontline Ventures documented 131 synthetic data companies by 2023, and Tomasz Tunguz (Theory Ventures) reported that Google's synthetic data research achieved 500-1,000x cost reductions versus human data labeling, with student models trained on synthetic data actually outperforming teacher models trained on real data. LBZ Advisory's Liat Ben-Zur pointed to a clear precedent: "Companies like SDL and Lionbridge built massive competitive advantages around proprietary linguistic datasets. Then Google Translate became good enough for most use cases using publicly available training data." The pattern — proprietary data moat → foundation model subsumes the value → data moat collapses — has now played out in computer vision, NLP, and predictive analytics.
Chegg is the canonical casualty. Its database of 79 million solved problems — once a formidable data asset — became worthless overnight when ChatGPT could generate solutions to any problem instantly. Stock fell from $113.51 to under $1 (99% decline), erasing $14.5 billion in market cap. The company slashed 45% of its workforce in October 2025. Employees had asked for resources to build AI tools in 2022 and were denied; by the time the pivot came, it was too late.
The critical distinction the original thesis misses: static data moats (libraries, databases, content archives) are dead. Dynamic data moats — where data is continuously generated through operations, deeply embedded in workflows, and impossible to replicate without doing the actual work — remain viable. V7 Labs' CEO Alberto Rizzoli articulated this well: "The moat isn't the data itself, it's the ability to put that data to work." Or as feedback from GenAI company operators told unique.ai more bluntly: "Data moat is bullshit" — unless the data is exclusive, continuously refreshed, and operationalized.
AI purchasing agents are eroding marketplace network effects
ARK Invest published a direct examination of this threat. In the scenario they describe, "autonomous AI purchasing agents act on behalf of shoppers, continuously sweeping marketplace listings, brand sites, social-commerce feeds, and loyalty inventories. In real-time, agents scan catalog data, compare specs across sellers, and negotiate prices." This fundamentally changes marketplace dynamics: when an AI agent makes purchasing decisions algorithmically, brand loyalty and platform habits — the behavioral glue of network effects — get bypassed. ARK estimates AI agents could influence the shift of $9 trillion in spending power.
NFX acknowledged the threat directly, noting that "incumbent marketplaces aren't likely to build a new system around the AI agent model" because it would replace the providers who constitute their supply side. The Network Law Review published academic research warning that AI agents create self-reinforcing feedback loops that can consolidate bargaining power into "dominant agent hubs" — potentially overriding existing marketplace network effects entirely.
Multi-homing dynamics compound the problem. HBS research documents that many drivers and riders use both Uber and Lyft, weakening lock-in despite both platforms' efforts to reward loyalty. Homejoy — a marketplace connecting house cleaners to customers — shut down because once users found a reliable cleaner, they bypassed the platform entirely. Groupon's market cap peaked in 2013; its revenue has declined since 2015 despite claiming network effects.
However, the strongest network effects survive this pressure. Morningstar's systematic reevaluation of 132 companies concluded that "simple customer embedding is likely to face pressure, but genuine network effects — where scale improves liquidity, relevance, content depth, telemetry, or ecosystem utility — are expected to increase in value." The distinction is between behavioral network effects (habits, familiarity) that AI agents erode and structural network effects (liquidity, data density, ecosystem integration) that agents cannot replicate. Airbnb's host reviews, booking history, and trust scores are structurally embedded; an AI agent can't synthesize a host's 500 five-star reviews.
Community moats are audience moats in disguise — and audiences disperse
Stack Overflow's collapse is the sharpest illustration. Monthly questions fell from a peak of ~200,000 to just 6,866 in early 2026 — roughly the volume when the site launched in 2008. The site's 84% of developers now use AI coding tools. The community moat was hollowed out by the very AI systems it helped train. The company survived by pivoting to data licensing (revenue doubled to ~$115M), but the community itself is functionally dead for simple questions.
Clubhouse reached a $4 billion valuation in early 2021 backed by Andreessen Horowitz, then collapsed within months. Competitors — Twitter Spaces, Discord Stages, Spotify Greenroom — replicated the core mechanic instantly. Without monetization, top creators left. The community dissolved because it was an audience dependent on creators, not a self-sustaining network.
Twitter/X demonstrated that even the most entrenched community moat in media history — journalists, politicians, and thought leaders all structurally dependent on the platform — could erode rapidly when trust was broken. Advertiser revenue dropped 60%; major brands (IBM, Disney, Apple) halted spending. User migration to Bluesky, Threads, and Mastodon accelerated. Reddit's API changes provoked 8,000+ subreddits going dark and destroyed the third-party app ecosystem that had been a key community feature.
The pattern is clear: community moats depend on governance, trust, and continued investment. They are not self-sustaining assets like network liquidity or proprietary data flows. Stop investing in the community — or betray its trust — and it dissolves faster than any other moat type.
Execution speed is currently winning, but it's a wasting asset
Cursor's trajectory is the single most compelling challenge to the moat thesis. Four MIT graduates took Cursor from zero to $2 billion ARR in roughly three years — doubling revenue in just 90 days in early 2026 — while competing directly against Microsoft's GitHub Copilot, which had VS Code's distribution, GitHub's 100M+ developer user base, and a deep OpenAI partnership. As Notorious PLG observed: "The fact that a startup can compete this effectively against Microsoft's resources speaks to both Cursor's execution and the openness of the developer tools market."
But Cursor's own vulnerability proves the thesis's deeper point. The LLMs powering Cursor are commoditizing rapidly; a competitor can fork the same open-source VS Code base, plug in the same model, and produce a very similar product. Cursor's moat is currently speed — Elad Gil documented roughly 60-100 internal releases daily with ~5 PRs per engineer per day. Speed alone, as Oxx VC cautioned, "simply takes advantage of the current arbitrage between what's technically possible at the vanguard and the typical customer's view of what's possible. Over time, this gap will atrophy."
NFX frames this elegantly as the motte-and-bailey model: speed is the bailey (outer defense, easy to breach), while network effects and deep embedding are the motte (inner fortress). Elad Gil concurs: "One of the few advantages a startup has against an incumbent is speed... but other startups may be able to copy said advantages." Speed earns the right to build moats; it is not itself a moat.
Foundation models are compressing the value of many "proprietary" advantages
Each major LLM upgrade delivers "free" capability improvements to every application built on top of it. When OpenAI released GPT-4, many AI products saw quality leaps simply by switching their API to the new model — no additional data collection, engineering, or proprietary advantage required. Insignia VC's analysis concluded that "a clever algorithm fine-tuned on small domain data can rival one trained on a giant private dataset thanks to knowledge already baked into modern LLMs."
This creates a specific threat pattern: any advantage that exists primarily in the model layer — custom training, fine-tuning, prompt engineering — gets compressed every 6-12 months as foundation models improve. The state-of-the-art proprietary model stays only ~6 months ahead of open-source alternatives (per Andrew Chen). DeepSeek's emergence demonstrated that near-frontier performance can be achieved with dramatically less capital.
III. The expanded moat taxonomy: nine additional defensible assets
The original thesis lists four moat types. The evidence supports at least nine additional categories, each with distinct durability and accessibility profiles. Ranked below by a composite of durability and evidence strength.
Regulatory and compliance barriers are the most durable moat AI cannot compress
A Mighty Capital analysis of 578 AI companies that raised $50M+ found that regulatory moats ranked as the most durable and hardest to replicate — above data, networks, and all other categories. The mechanism is structural: certifications like HIPAA, SOC 2 Type II, FedRAMP, FDA validation, and financial licensing (banking charters, broker-dealer registration) take 2-5 years to establish and require sustained operational compliance. The EU AI Act, enforceable by 2026, imposes fines up to €35 million or 7% of global revenue. No amount of AI compresses the bureaucratic timeline.
Stripe illustrates the accumulation strategy: state money-transmitter licenses across jurisdictions, a Georgia MALPB charter (April 2025) allowing direct card network access, and a conditional OCC national trust bank charter for its Bridge subsidiary (February 2026). Each layer took years. Epic Systems' HIPAA certifications and Joint Commission audit trails make it unreplaceable in healthcare — not because its software is superior, but because "the compliance infrastructure is the product." Durability: very high (5-15+ years). Accessibility for small teams: very difficult. Evidence strength: strong.
Workflow embedding creates switching costs that survive AI disruption
When a product becomes the system of record — the spine that other tools connect to — every new integration increases the cost of removal. Salesforce has 3,000+ applications in its AppExchange; unplugging it means renegotiating every integration simultaneously. Insight Partners identified workflow depth as one of two primary moat categories for vertical AI companies, documenting how FDE-model companies like Filevine (legal OS) and Basis (accounting) create implementations that take 6-12 months to build and would cost a full year of productivity to switch away from.
A critical emerging concept is "representation switching costs" — in the AI era, lock-in comes not from data but from dependence on the machine-readable representation of reality that workflows, agents, and institutions depend on. A spreadsheet of transactions is portable; a machine-operational understanding of a customer's risk profile, consent boundaries, and exception patterns is not. BVP's observation that "systems of action are replacing systems of record" captures the evolution: AI-native products don't just store data, they act on it, and the accumulated actions and decisions create deeper embedding than data storage ever could. Durability: high (5-10+ years). Accessibility: moderate — vertical startups can achieve this in 18-24 months by designing as the system of record for a narrow domain. Evidence strength: strong.
Process power and the demo-to-production gap
The gap between showing an AI capability and making it production-ready is enormous and well-documented. Harvard and Stanford research finds that 90-95% of AI initiatives fail to reach sustained production value. Among the 5-10% that ship, only 6% qualify as high performers. Task completion rates in real business settings average 50-55% versus 95% in demos. The "99% Rule" states that reaching 99% reliability for a complex system takes 10-100x the effort of a basic MVP.
This gap is the moat. Greenlite AI (YC, KYC/AML compliance) is the YC Lightcone podcast's illustrative example: "rivals could write a toy KYC demo in days, but Greenlite's production system took years of data and iteration — the final 10% of performance took 10-100x the work of a hackathon prototype." Harvey (legal AI) grew from $3B to $8B valuation in 2025 through deep understanding of how law firms structure work — something no new entrant can replicate overnight even if code is free. Scale AI secured a 5-year, $100M DoD agreement requiring SCIF infrastructure, security clearances, and relationships that take years to build. Durability: strong (3-7+ years). Accessibility: low — requires forward-deployed engineering teams and years of refinement. Evidence strength: strong.
Counterpositioning exploits incumbent business model paralysis
Helmer defines counterpositioning as a newcomer adopting a superior business model that the incumbent cannot mimic without anticipated damage to existing business. In AI, this manifests primarily through pricing: legacy SaaS charges per-seat, but AI reduces the number of seats needed (the 2026 SaaSpocalypse saw CIOs reducing seats at a ~1:5 ratio per AI agent deployed). AI-native companies pricing per-task or per-outcome offer superior economics that per-seat incumbents cannot adopt without cannibalizing revenue.
Meta's Llama strategy is the textbook case — open-sourcing frontier-class models to commoditize competitors' proprietary model revenue, with Llama surpassing 1 billion downloads by late 2025. Avoca.ai chose outcome-based pricing for home services call centers, deeply optimizing for trades workflows in ways that make it structurally unattractive for broad call-center SaaS companies to follow. Durability: moderate (2-5 years) — persists until incumbents decide to self-cannibalize. Accessibility: high — requires strategic insight more than capital. Evidence strength: strong.
Brand operates as a trust proxy when AI outputs are opaque
When capabilities commoditize, brand becomes the heuristic users rely on to evaluate quality they cannot directly inspect. ChatGPT commanded 64.5% of tracked AI chatbot usage in January 2026 versus Gemini's 21.5% — despite Google achieving comparable benchmark performance. TLV Partners called ChatGPT "the Kleenex of AI." Perplexity built its brand through "no ads, no hidden agenda, and transparent sourcing," reaching $75M+ ARR by earning trust iteratively.
The paradox: AI-generated noise makes trust more scarce (strengthening brand moats) while simultaneously enabling trust fraud (weakening them). In regulated and mission-critical domains — healthcare, finance, legal — brand durability is higher because predictability and auditability matter. In consumer AI, the Nokia/Netscape precedent looms: first-mover brand advantages can collapse rapidly when a superior product arrives. Durability: moderate (3-10 years, domain-dependent). Accessibility: moderate — achievable by small teams through PLG and community evangelism. Evidence strength: moderate.
Context and memory accumulation create AI-native switching costs
BVP's central defensibility thesis for 2025 is that "persistent memory and semantic understanding create emotional and functional lock-in." When an AI system accumulates a user's project history, preferences, communication patterns, and workflow context, switching means teaching the new system everything from scratch — a process that can take weeks. NFX identifies this as a "personal utility network effect": the more memory your AI copilot builds alongside you, the more personal utility you derive.
OpenAI's developer documentation frames this explicitly as "a strategic moat — a way to continuously capture, refine, and apply high-quality behavioral data, creating denser, higher-signal information about your users than typical clicks or history data." However, this moat faces emerging threats: universal memory solutions like AI Context Flow aim to make context portable across platforms, and data portability regulations could force export of personalization data. Durability: moderate (1-3 years). Accessibility: very high — any AI app can implement memory features using widely available vector databases and memory frameworks. Evidence strength: moderate.
Agentic orchestration reliability is the next process power moat
If each step in a 10-step AI agent workflow achieves 95% accuracy, end-to-end success is only ~60%. Building multi-agent systems that work reliably in production — with orchestration, observability, governed data access, retries, and audit trails — is an engineering challenge that creates meaningful defensibility. Bain identifies three required layers: orchestration (workflow management), observability (tracing and monitoring), and governed data access (security and permissions). Companies reaching 80-90% autonomous task completion do it through extensive prompt engineering, RAG grounded in company knowledge bases, tightly bounded use-case design, and iterative tuning loops — not by swapping models.
Emerging standards like MCP (Model Context Protocol) and A2A (Agent-to-Agent) protocols may reduce orchestration complexity over time, creating an erosion risk. Durability: moderate-to-strong (2-5 years). Accessibility: low — requires cross-functional teams and months of iterative tuning. Evidence strength: moderate (mostly forward-looking).
Vertical integration creates the highest-durability moats but requires massive capital
NVIDIA's CUDA ecosystem — built over 20 years with 4+ million developers, 3,000+ optimized applications, and deep integration into PyTorch, JAX, and TensorFlow — represents perhaps the deepest moat in technology. Hugging Face hosts 500,000+ models, the overwhelming majority trained on NVIDIA hardware with CUDA kernels. Apple invested $20B+ growing its silicon team and now controls the full stack from Neural Engine to Core ML to operating system. Tesla designs custom AI chips, builds Gigafactories, and uses data from millions of vehicles to train FSD.
This category is essentially inaccessible to startups and early-stage teams. Even AMD and Intel, with billions in resources, have struggled to replicate NVIDIA's software ecosystem. Durability: very high (10-20+ years). Accessibility: extremely difficult — requires billions in R&D. Evidence strength: strong.
Talent density is real but volatile
Anthropic leads AI industry retention at 80% for employees with 2+ years tenure versus OpenAI's 67% and Meta's 64%. Engineers at OpenAI are 8x more likely to leave for Anthropic than the reverse. Palantir employs 50% of all forward-deployed engineers in the US. But talent is inherently mobile: Mira Murati left OpenAI and pulled 20 staffers before even announcing her startup. Meta offers $100M signing bonuses; OpenAI counters with $20M+ equity packages. The FDE-based moat (encoding institutional knowledge into customer relationships) is more durable than pure talent concentration, but neither is accessible to small teams competing on compensation. Durability: low-to-moderate (1-5 years). Accessibility: very difficult. Evidence strength: moderate.
IV. The updated framework: a defensibility matrix and build sequence
Defensibility matrix
The table below ranks all 13 moat types across three dimensions: durability (how long the moat lasts under competitive pressure), accessibility (can a small team of 5-15 people build this in 12-24 months?), and evidence strength (how confident we should be that this moat actually works, based on available data).
| Moat type | Durability | Accessibility | Evidence | Confidence |
|---|---|---|---|---|
| Structural network effects | Very high | Moderate | Strong | High |
| Regulatory/compliance | Very high | Very low | Strong | High |
| Vertical integration (HW+SW) | Very high | Very low | Strong | High |
| Workflow embedding / switching costs | High | Moderate | Strong | High |
| Process power (demo→production) | High | Low-moderate | Strong | High |
| Dynamic proprietary data (flywheel) | High | Moderate | Strong | High |
| Counterpositioning | Moderate-high | High | Strong | High |
| Brand as trust proxy | Moderate | Moderate | Moderate | Medium |
| Community (genuine, not audience) | Moderate | Moderate | Moderate | Medium |
| Context/memory accumulation | Moderate | Very high | Moderate | Medium |
| Agentic orchestration reliability | Moderate | Low | Moderate | Medium-low |
| Talent density / FDE | Low-moderate | Very low | Moderate | Medium |
| Iteration velocity (speed) | Transitional (6-18mo) | Very high | Strong | High |
What a small team should build, and in what order
The evidence across sources converges on a consistent build sequence. Elad Gil states it most directly: "Most SaaS software starts off default non-defensible and tends to build a moat over time." NFX's motte-and-bailey model, YC Lightcone's episode on the 7 Powers, and Insight Partners' vertical AI framework all describe essentially the same progression.
Phase 1 (Months 0-6): Speed + Counterpositioning. Ship fast in a narrow vertical. Choose a business model that incumbents can't adopt (outcome-based pricing, open-source wedge, or AI-native workflow replacement). This is the bailey — easy to breach but buys time. Cursor exemplifies this: 60-100 internal releases daily, rapid iteration, no moat except velocity. Elad Gil's "GPT Ladder" concept applies: identify workflows that become viable at the current model capability level and capture the market before models improve further.
Phase 2 (Months 6-18): Workflow embedding + Context accumulation. Become the system of record for your vertical. Every customer implementation should deepen integration with existing tools and accumulate context that creates switching costs. HappyRobot (logistics AI, YC '21) illustrates this: after 6-12 months of forward-deployed work, "switching would risk losing a full year of productivity." Simultaneously, build persistent memory that personalizes the product with each interaction.
Phase 3 (Months 18-36): Dynamic data flywheel + Process power. If your workflow embedding is working, you're now generating proprietary data through operations that nobody else has. This data must be exclusive (not scrapable), continuously refreshed (not static), embedded in decisions (not just stored), and compounding (each data point improves the product). The demo-to-production gap becomes your friend: accumulate edge-case knowledge, regulatory trust, and operational reliability that competitors cannot shortcut.
Phase 4 (Months 36+): Network effects + Community. Once you have enough users and data density, architect for network effects — marketplace liquidity, data network effects, platform effects where others build on your product. Community follows naturally from a product people depend on, but must be genuinely self-sustaining (not just an audience that dissolves when you stop investing).
Regulatory moats can be pursued at any stage if you're in a regulated vertical (healthcare, finance, legal) — but they're a parallel track, not a sequence step. Start compliance work early because it takes 2-5 years and cannot be accelerated.
Where evidence is thin versus strong
The thesis has high-confidence support for: network effects as the single most valuable digital moat (NFX, Morgan Stanley, Morningstar all converge); static data moats being dead (Chegg, SDL/Lionbridge, Casado's analysis); workflow embedding creating durable switching costs (Salesforce retention despite inferior product, Insight Partners portfolio data); and the demo-to-production gap as a genuine defensibility source (Harvard/Stanford failure statistics, YC Lightcone case studies).
The thesis has moderate confidence for: community moats being durable when genuinely self-sustaining (strong examples exist but so do rapid collapses — Clubhouse, Stack Overflow); brand as a trust proxy in AI (ChatGPT market share data is strong but Nokia/Netscape precedent suggests fragility); and context/memory as a switching cost (mechanism is clear but portability solutions are emerging).
The thesis has low confidence on: whether AI purchasing agents will actually erode marketplace network effects at scale (theoretically compelling but limited real-world evidence so far); whether agentic orchestration reliability will remain a moat as standards like MCP mature; and the exact durability of counterpositioning moats (historically they last until incumbents decide to self-cannibalize, but timing varies enormously).
What actually defends a business when AI makes building free
The consensus across every authoritative source — a16z, BVP, NFX, Helmer, Insight Partners, Elad Gil, Morningstar, ARK — can be distilled into a single formula: models commoditize, workflows differentiate, data flywheels compound, networks defend. The original thesis captures three of these four elements but underweights workflow embedding, entirely omits regulatory barriers and process power, and fails to distinguish between static data (dead) and dynamic data flywheels (alive).
The most important update is structural: moats in the AI era are not binary (you have one or you don't) but layered and sequential. Speed earns the right to embed. Embedding generates proprietary data. Data enables network effects. Each layer reinforces the others — what NFX calls "the network effect of network effects." The companies best positioned in 2026 are not those with a single dominant moat but those stacking multiple reinforcing advantages: Bloomberg (data + network + workflow + brand + regulatory), Palantir (process power + FDE + switching costs + data), and Veeva (network + regulatory + workflow).
For a small team starting today, the implication is clear: don't try to build the moat first. Build the product fast, embed it in workflows, generate data through operations, and let the moat emerge from the work. The thesis's core principle — "anything worth building must be defensible" — is correct, but the corollary is equally important: defensibility is built through use, not through architecture. The moat is not a feature you ship. It is an emergent property of doing the work, in the right sequence, better and faster than everyone else.