What Survives When Building Gets Easy

A solo founder in 2025 sold a chat-based no-code app builder to its incumbent competitor for $80 million cash plus earn-outs running through 2029, six months from founding to exit and operating a team of one. A different solo founder runs 14 startups with $2.5M+ cumulative revenue and roughly $71K in any given month, by his own public reporting and a third-party tracker. Pieter Levels operates a portfolio at roughly $250K monthly revenue across Photo AI, Nomad List, Remote OK, and Interior AI, all built solo on PHP and SQLite with no team and no proprietary technology. None of these stories looked plausible as recently as 2023.

When the marginal cost of producing a product approaches zero, every traditional moat that depended on production scarcity disappears with it. Engineering is no longer the bottleneck; marketing has absorbed the budget that used to fund development; the operative question has shifted from whether a product can be built to whether it can be defended once it ships. Most early-stage companies as of April 2026 cannot answer the second question, and the gap is where the durable moats of the next decade hide. The previous chapter took up the rebuilt market structure where agents are the buyers and SaaS is callable. This chapter asks what survives that market: which competitive advantages compound once production costs collapse and the 2010-2024 differentiation playbook stops working.

The Skill File Test filters most pitch decks before the term-sheet stage

The Skill File Test is editorial framing for a structural question every AI-focused founder should run before walking into a term-sheet conversation: if the product can be reproduced as a markdown file in Claude or a custom GPT, the moat is missing. No canonical published statement of the rule exists; the closest public analog is the recurring analyst critique that infrastructure tools like Cursor have no defensible position against the underlying model platforms. The framing is the chapter's, but the test the framing encodes is real and easy to run.

The test filters on structure, not on technical novelty. An agent that can copy a product from public docs in a week kills any technical moat the founder thought existed; the test is passed only when the surface depends on something that cannot be encoded as instructions plus public documentation — proprietary data, regulatory access, physical infrastructure, or distribution loyalty. Most products that look defensible on a deck fail the test once the cost of copying drops below a single afternoon of engineer time.

A worked example. Take a generic AI sprint planner that ingests issues, scores by velocity history, and drafts a sprint. Three sentences of capability description. Public documentation needed: the issue tracker's API and a model provider's text-generation endpoint. An agent could rebuild that surface in a week. The moat for a real sprint planner has to come from the issue tracker's user graph and decision trace — not from the AI feature, which is the part the agent reproduces fastest. The same logic runs against AI customer-service products, AI marketing copy generators, AI code reviewers, and almost every product whose central value proposition is an LLM call wrapped in a UI.

The empirical pattern visible across solo-founder portfolios in 2025-2026 is that execution velocity alone produces revenue without durable defense. Pieter Levels operates a portfolio at roughly $250K monthly revenue across Photo AI ($132-138K/month), Interior AI ($38-45K/month), Nomad List ($38K/month), Remote OK ($35-41K/month), and a tail of smaller products, all operated solo on conventional infrastructure with no proprietary technology layer. Marc Lou ships a similar pattern at the consumer-tools tier: 14 active startups, $2.5M+ cumulative revenue, monthly revenue running between $65K and $95K through late 2025 and early 2026 across CodeFast, ShipFast, DataFast, TrustMRR, and a long tail of smaller products. Both founders survive competition by ruthless pricing, ruthless shipping cadence, and a personal audience that fills the marketing layer at zero cost rather than any product moat in the conventional sense.

The investor-side correlate: ARR growth alone no longer signals defensibility. Cluely raised $20.3 million in 2025 ($5.3M seed in April 2025 from Abstract Ventures and Susa Ventures, then $15M Series A from Andreessen Horowitz in June 2025) on viral marketing built on deliberately provocative messaging. The CEO publicly claimed $7M ARR in mid-2025; in March 2026 he retracted the claim on X, calling it the only blatantly dishonest thing he had said online. Distribution that produces fundraising momentum is a campaign rather than a moat; the gap shows up the moment the market discounts unverified ARR claims.

The diagnostic is the same on both sides. A founder writes the product's core capability in three sentences and asks whether an agent given public documentation could replicate it within a week. An executive runs the same test against the firm's three highest-margin surfaces. The surfaces that fail the test need a moat layer or a different product entirely.

Six structural moats survive when building gets easy

The moats that hold against agent-native competition share a common feature: each one depends on something that cannot be encoded as instructions plus public documentation. Six categories are visible as the surface of durable defense in 2026.

Regulatory licenses and government-enforced bottlenecks. Bank charters, broker-dealer registrations, insurance carrier permits, healthcare provider licenses, and legal admissions remain hard-gated by approval processes that run at human speed. The OCC received 14 de novo applications for limited-purpose national trust bank charters in 2025, primarily from fintech and digital-asset firms seeking to vertically integrate payments, custody, lending, and stablecoin issuance and reduce reliance on third-party banking intermediaries. The mechanism is straightforward: license supply is administratively fixed, demand rises as applicants begin filing with AI assistance, and the regulator's approval cadence does not scale with applicant volume. Failure mode: regulatory capture itself becomes attackable when applications surge. The licensing bottleneck shifts from "does the applicant qualify" to "can the regulator process volume," and the eventual response is automated review, which removes the moat the applicant was relying on.

Proprietary data loops and decision traces. The accumulation of structured records of why decisions were made under which constraints with which outcomes compounds across operating time and cannot be reproduced by competitors operating on public data alone. The platform-tier examples (Glean Enterprise Graph, Palantir Foundry Ontology, Databricks Unity Catalog, Salesforce Agentforce on Data Cloud) developed in the prior part are the production reference shape. Two further 2025-2026 anchors landed in the same category: Celonis markets its Process Intelligence Graph as "a living digital twin of business operations" feeding AI agents with business context they cannot get from public data; Snowflake's Horizon Catalog provides "a shared view of your metadata, lineage, and rules" across engines with policy enforcement at the retrieval layer. The defensibility threshold is editorial rather than measured: as a working heuristic, a decision-trace ontology starts to defend a moat once it captures around 50 validated exception cases the model would otherwise mishandle, and becomes hard to reproduce around 500 such cases drawn from real customer transactions rather than synthetic data. The numbers are a practitioner sketch rather than benchmark data, but the structural distinction holds: below the lower band the trace is internal documentation, and the moat opens up only when the trace covers enough real-customer cases that a competitor cannot reconstruct the equivalent from public data alone. Failure mode: data ingestion without retrieval-time policy enforcement is a data lake, not a moat, and competitors close the gap from the public-data side as foundation models improve at extracting structure from logs, transcripts, and regulatory filings.

Trust-based distribution and audience concentration. Andrej Karpathy held 2,343,628 followers on X as of April 29 2026. His April 2 2026 post on LLM knowledge bases drew 104K likes, 57K bookmarks, 8.9K reposts, and 2.8K replies on a topic any reasonably calibrated agent could write 1,000 words about. The mechanism is that when AI-generated content saturates every channel, attention concentrates on creators with established trust, and that trust is itself the scarce resource. Failure mode: trust survives only when it translates to repeat usage and revenue rather than first-touch impressions; viral distribution that produces fundraising momentum without retention is a campaign that runs out before it becomes a moat.

Physical infrastructure and capital-intensive supply chains. TSMC took 38% of the global foundry market in 2025 with $56 billion in planned 2026 capex, with advanced process technologies at 7nm and below accounting for 74% of TSMC's wafer revenue in Q4 2025. Samsung's 3nm yields sat in the 30-40% range for much of 2025, insufficient to attract large external orders, and its first 2nm product entered mass production by December 2025 with yields below 50%. ASML produces "the world's only commercial extreme ultraviolet lithography tools" used in advanced chip fabrication, and the next-generation High-NA EUV machines cost roughly $400 million each, twice the cost of original EUV machines. The mechanism: the equipment, the operational knowledge, and the customer relationships were assembled across decades of capital and engineering investment, and no startup can replicate any of the three at speed. Failure mode: geographic concentration makes the moat geopolitically vulnerable, and the physical-infrastructure moat is real on the decade horizon but exposed to disruption by export controls and conflict on shorter horizons.

Taste, judgment, and aesthetic discipline. When intelligence becomes abundant, the scarce resource is the judgment about which output to keep. The mechanism is well-established outside the AI context (Steve Jobs at Apple; the small handful of ML researchers whose intuitions reliably guided which model architectures to fund and whose 2025 individual signing packages reportedly ran into nine figures because the market is paying for taste rather than headcount). Editorial framing: agents generate options at near-zero cost, and humans pick from them. The moat is durable per-operator and fragile per-firm, because taste does not transfer across domains and rarely transfers cleanly across founder generations. A worked non-software case: a custom-furniture maker's competitive advantage is the proprietor's taste in joinery, materials, and proportion. None of those reproduces by instruction alone, and the firm's value migrates with the operator who holds the taste rather than residing in the firm itself.

Vertical integration of compute, model, and distribution. xAI's Colossus reached 555,000 GPUs across two Memphis facilities by early January 2026, with capex tracking toward 2 gigawatts of total power capacity and a third building (publicly named MACROHARDRR by Elon Musk) breaking ground for 2026 deployment. Owning the chip layer, the data center, the model, and the distribution channel stacks the markup at each layer of the stack, producing a meaningful unit-cost advantage relative to competitors who rent at any layer. Failure mode: vertical integration at this scale requires capital only the largest technology firms can deploy, and the long-run market structure concentrates among that small set of firms rather than diffusing across the broader market.

The energy framing makes the moat's scale legible. One gigawatt is roughly the output of one nuclear reactor. xAI's 2 GW Memphis cluster is therefore the energy footprint of two nuclear reactors deployed for one firm's training compute, and Musk has publicly stated the goal of pulling ahead of every other compute holder combined within five years. The EU's average electricity load across 2024 was approximately 115 GW (about 1,005 TWh distributed across the year). The EU's average electricity load across 2024 was approximately 115 GW (about 1,005 TWh distributed across the year)↗. A vertically-integrated compute roadmap that builds gigawatt-scale data centers on a quarterly cadence is operating in the same order of magnitude as a mid-sized European country's average power demand. The token-supply moat at that scale is what produces the cost advantage further down the stack; the firms that match the integration play are competing on energy infrastructure, not on software cleverness. The companion 2025 announcement of Macrohard — Musk's public framing of an end-to-end AI-built software firm — reads as the natural product surface for that compute base2.

Six moats; six distinct holding mechanisms; one prescriptive claim. The firms that survive the next decade pick at least two of the six and execute against both:

• Regulatory bottlenecks hold because supply is administratively fixed.
• Decision traces compound over operating time and cannot be backfilled by competitors.
• Trust survives because attention does not multiply with content volume.
• Physical infrastructure resists software cloning by definition.
• Taste stays scarce when intelligence does not.
• Vertical integration prices almost every potential competitor out of the stack.

What disappears in the same window is the inverse list:

• Code complexity collapses once agents implement specs.
• Team size shrinks inside factories.
• Foundation models extract structure from public filings, transcripts, and logs, so information hoarding stops protecting margin.
• Once table-stakes velocity becomes the floor, execution speed alone no longer separates one firm from another.

The moat layer is the part of the firm that is not in the disappearing list, and the firms that pick at least two from the six survive.

The vendor-side margin re-rate landed in public markets in early 2026. The February-2026 SaaSpocalypse selloff erased roughly $285 billion in software market cap inside a 48-hour window, and by late March the iShares Expanded Tech-Software ETF (BATS:IGV) was down more than 21 percent year-to-date2. Per-name declines clustered around the per-seat pricing model: Salesforce trading roughly 30-33 percent below its 2025 peak, Workday roughly 33-40 percent off, Oracle roughly 19 percent off, SAP roughly 15 percent off. The mechanism named across coverage is direct: when ten agents do the work of a hundred reps, the customer needs ten Salesforce seats rather than a hundred, and the per-seat revenue model that built the SaaS category re-rates as the seat count compresses.

The counter-examples define the structural distinction. Datadog and Cloudflare priced on usage rather than per-seat compressed less because their revenue rises with agent-driven traffic; AWS, Azure, and GCP carry the floor because agents drive compute consumption upward at every layer; Snowflake and Databricks hold because they store the data agents read. The sector pattern that surfaces from the price action is the one this chapter has been making in slower form: code-as-moat erodes (Salesforce, Oracle, SAP, Workday) while infrastructure-as-moat holds (Cloudflare, Datadog, the cloud providers, the data warehouses). The selloff is the public-market expression of the chapter's underlying claim that the bubble was code being a moat, not software being a moat.

The decision matrix runs roughly by company stage. A solo founder under $1M ARR has access to trust-based distribution and taste; regulatory and physical-infrastructure moats are unavailable at that capital and timeline. A Series A SaaS at $5-20M ARR has access to decision-trace data and trust; vertical integration is unavailable until much later. A regulated incumbent defends the regulatory license and layers decision-trace data on top. A vertical integrator at the chip-and-data-center scale is one of fewer than ten entities globally and runs the integration play as the highest-tier strategy. The reader's pick depends on stage and accessible capital, not on which moat sounds most defensible on a deck.

Information-asymmetry business models face structural margin pressure

Industries whose core competency is calculating something better than their customers face structural compression as the calculation becomes available to every customer at near-zero marginal cost. The pattern runs across four classic information-asymmetry business models:

• Insurance prices actuarial risk on data the customer cannot reproduce.
• Telecom runs pricing optimization across millions of usage profiles the customer never sees.
• Financial services captures the interest-rate spread on a calculation the customer never runs.
• Consulting sells the implementation knowledge accumulated across prior engagements that the customer's organization could not internally reconstruct.

Each model rents on the customer's information disadvantage, and each becomes attackable the moment the customer's agent can run the same math at near-zero marginal cost. The agent does not have to be smarter than the incumbent's analyst; it has to be approximately as accurate at a small fraction of the cost, and it has to be available to every consumer rather than the incumbent's enterprise customers alone. When that condition holds, the economic equilibrium shifts away from the incumbent's information rent toward a thinner margin on the underlying service.

Horizon: The empirical compression wave is forward-looking. The IAIS Global Insurance Market Report 2025 shows the non-life reinsurance combined ratio holding stable at 95% in 2024, which means there is no macro-level evidence of margin compression attributable to consumers contesting charges with AI assistance as of late 2025. The compression argument is structural rather than measured. The supply-side investment signal (GenAI-in-insurance market projected to grow from $1.11B in 2025 to $14.35B by 2035 at 29.1% CAGR). The compression argument is structural rather than measured. The supply-side investment signal (GenAI-in-insurance market projected to grow from $1.11B in 2025 to $14.35B by 2035 at 29.1% CAGR)↗ reflects incumbent investment in defending against the compression that they see coming, not compression that has already arrived. The reader should treat the structural argument as a 2026-2030 horizon claim, not a 2025 measured outcome.

The personal consumer-advocate category is the demand-side bridge. DoNotPay was the first AI-branded entrant and was fined $193,000 by the FTC in February 2025 for deceptive claims about its AI legal capabilities. The category itself survived the regulatory action. Flightright, operating within the EU's passenger-rights framework, reports €700 million in cumulative compensation paid out to its users with a stated 99% success rate on accepted cases, charging a 20-30% commission on recovered amounts. AirHelp reports 28 million customers helped under a similar model with a standard 35% fee. Both companies operate inside well-defined regulatory frameworks where the underlying entitlement (passenger compensation under EU Regulation 261/2004) is fixed by law and the AI value is in claims processing rather than legal interpretation. The category opportunity that opens further from this point is the broader consumer-side bridge product — not formal legal advice, which the DoNotPay action defined as the regulatory tripwire, but structured contestation of charges, refunds, and entitlements where the agent ships a verifiable outcome rather than a recommendation.

The CFO diagnostic for an information-asymmetry incumbent runs as a three-step procedure. First, take 100 historical pricing or underwriting decisions from the past 12 months, run the same input through a current frontier model with public reference data only, and record the agreement rate plus median delta. Second, divide the firm's fully-loaded analyst cost per decision by the agent's per-call cost on the same task. Third, score whether the firm's regulatory wrapper (the licensing requirement, the audit obligation, the carrier permit) captures value independent of the calculation itself. If the agent matches the underlying calculation closely on a small fraction of the analyst cost, and the wrapper does not capture meaningful independent value, the firm faces compression on the next 4-8 quarter earnings horizon. The exact window is structural rather than measured — the empirical wave is forward-looking, not present-state — but the diagnostic is what tells the CFO whether to defend the calculation or build the wrapper.

Vertical integration is the moat that compounds while the others stabilize

The six moats above defend market position; vertical integration compounds it. xAI bought a third Memphis building in late December 2025 specifically to keep its compute build ahead of OpenAI and Anthropic, and the cost advantage from owning chips, data centers, and models stacks at every layer of the inference pipeline. The deployment cadence is the visible signal: 555K GPUs across two facilities in roughly the time it takes a competitor running on rented infrastructure to negotiate a single multi-gigawatt power purchase agreement.

The integration thesis runs through the whole frontier-model tier in roughly parallel form, with each firm owning a different stack of layers:

• NVIDIA — chip (Hopper, Blackwell), software stack (CUDA), cloud service (DGX Cloud), and a foundation-model offering through partner deployments.
• Google — chip (TPU), foundation model (Gemini), search distribution (Google Search and AI Overviews), and productivity surface (Workspace).
• Microsoft — cloud (Azure), partnership-based model layer (Azure OpenAI Service), productivity surface (Copilot inside Office), and increasingly its own custom silicon (Maia, Cobalt).
• Meta — training infrastructure, Llama model family, and three of the largest distribution surfaces in the world (WhatsApp, Instagram, Facebook).

Each operates on the integrated stack with a meaningful cost advantage relative to competitors renting at any single layer. The exact multiplier is not publicly verifiable; the qualitative direction shows up in deployment speed, model release cadence, and the gap between retail token pricing and the bundled-stack pricing inside each integrated firm's first-party products. The cost advantage does not appear automatically, either — Meta's pre-Llama hardware investments did not produce a model advantage on their own, and the integration premium materializes only when each layer is operationally competitive in its own right.

The compounding is geometric rather than additive. Each integration tier contributes a multiplicative cost factor on top of the prior tiers — own chip times own data center times own model times own distribution — and a competitor that buys at any layer pays the markup at that layer while a competitor that tries to build at any layer needs years to match the integrated incumbent's operational depth. The compounding does not depend on better execution at the integrated firm; it depends on the geometry of the cost stack, which is why no amount of execution discipline at a non-integrated competitor closes the gap once the structural advantage is in place.

The deployed pattern in 2026 is that the highest-tier strategy in agent commerce is to operate the factory of factories: ship product variants against the integrated stack at high volume, kill the variants that fail to draw demand, and let the surviving variants compound the cost advantage across each iteration. Vertical integration for everyone else is necessarily smaller in scope. The Series A or growth-stage SaaS founder who cannot deploy multi-billion-dollar capex on chips and data centers can still own the data ingestion pipeline rather than rent it from Segment or Fivetran, fine-tune on the firm's decision traces rather than calling base models off the shelf, and ship agent SDKs for customers' agents in the shape Stripe ships its Agent Toolkit at the payments layer. Smaller-scale vertical integration is real and accessible at a five-figure fine-tuning budget on the customer-decision-trace tier, even when the chip-and-data-center version of the moat is unavailable.

Software factories produce speed and the moat layer is what makes the speed defensible

A common 2026 founder mistake is treating factory speed as the moat. The chapter on software factories developed the engineering shape of the high-cadence build pipeline (spec-and-tests as contract, agents as the implementation layer, branch-per-agent isolation, plan-approval gates). The factory delivers speed and only speed. Without a moat layer, the firm stays copyable inside a quarter.

Factories accumulate moat only when one of the six categories above catches the factory's output. The pattern across solo-founder portfolios already cited in the chapter — Marc Lou and Pieter Levels running multi-product operations at solid revenue — produces income but is structurally vulnerable to the next operator who runs the same factory pattern at the same cadence. The factor that decides which Business Factory operator wins as factory speed becomes table stakes is which moat the operator pairs with the factory. Execution velocity alone runs into a race against every other factory operator. Execution velocity plus regulatory license, or plus decision-trace data, or plus trust-based distribution, is a different game.

The gate between factory speed and moat is the discipline that turns a Business Factory from a churn machine into a compounding firm. The exact thresholds are operator-set rather than industry-standard, but as a working framing: before scaling product N+1, the previous product should clear at least one moat-clearance metric — 90-day retention above 40 percent, top-of-funnel from non-paid sources above 30 percent, or trust-source attribution (newsletter, founder audience, repeat customers) above 50 percent of new customers. Products that do not clear get killed, and the marketing spend they consumed gets reinvested in the next experiment.

The non-software case runs the same logic at a different scale. A custom-furniture shop running a portfolio of 50 product designs per quarter with no defensive layer is indistinguishable from 100 other shops doing the same thing within a year. The same shop with a portfolio plus a regional reputation moat — taste in joinery and proportion, an audience that buys repeat, an apprenticeship that compounds across years — keeps the customers that competitors copy the designs to. Speed produces the products; the moat layer keeps the customers from migrating to the next shop that copies the designs the following quarter.

Verifiable rewards explain why some industries fall first

The moat ranking above lines up roughly with the speed at which each category erodes, and the explanation runs through the verifiable-reward structure that drove the 2025 model-training jump. Reinforcement Learning with Verifiable Rewards (RLVR) emerged through 2025 as the de facto new training stage for frontier models, with OpenAI's o1 (late 2024) the first demonstration and o3 (early 2025) the release where the practitioner-visible difference landed. The mechanism: programming has structurally fast feedback at a cost of milliseconds per signal, since code compiles or it does not, unit tests pass or fail, type checks succeed or surface specific errors. Math (verifiable against ground truth), scientific computation (verifiable against simulations), and any task with a deterministic checker share the same property and trained the same way.

Tasks without fast verifiable feedback do not improve at the same rate. Management decisions take months to reveal their quality, with the signal arriving conflated with market conditions, team execution, and a dozen other variables. Design choices reveal themselves through customer behavior over weeks, never cleanly attributable to any single cut. Film edits depend on storytelling and casting calls made years earlier; the audience response that grades the cut is the noisiest signal of the three. None of these can be trained against the automatic reward signal that RLVR exploits, and the rate of model improvement on judgment tasks lags the rate on programming-and-math tasks by a substantial margin.

Industries whose core competency is a verifiable-reward task compress before industries whose core competency is judgment, which is why the moat ranking above lines up roughly with the speed at which each category erodes. Information-asymmetry incumbents compress first because the underlying calculation is verifiable. Regulatory-license-protected services compress slower because the license is itself a non-verifiable judgment about applicant qualification. Taste-based moats compress slowest because taste is the textbook non-verifiable judgment task. Decision-trace and physical-infrastructure moats erode on a different axis altogether — neither is a task with a verifiable reward gradient, so neither shrinks from RLVR-style improvement; decision traces erode from public-data extraction and physical infrastructure erodes from geopolitical risk on different time horizons.

The moats that look durable today are still time-limited

Each of the six moats currently holds against agent-native competition for a specific reason; all six are also under attack from the same compute-and-capability curve that produced the cost compression in the first place. The closing posture is structural honesty: the moats name which surfaces are defensible right now, not which surfaces are defensible permanently.

Take regulatory bottlenecks first. The bottleneck holds while the regulator's approval cadence runs at human speed, but applicants filing with AI assistance generate orders of magnitude more applications than the regulator can process. The first response is human triage at the regulator; the eventual response is automated review of standardized applications, which removes the moat the applicant was relying on. The window between application surge and regulatory automation is what defines the holding period for a regulatory-license moat in any given category.

Decision traces erode from a different angle. Foundation models continue to improve at extracting structure from unstructured public data — logs, transcripts, news, public records, regulatory filings — and the moat narrows as the gap between what private decision traces show and what public data can reconstruct closes. The part of the trace that holds longest is the genuinely private layer: internal exception lists, customer-specific configurations, hiring decisions, the why-this-was-rejected reasoning that no public source records.

Trust-based distribution depends on a stricter test. Repeat usage and revenue have to translate from the trust signal, or the moat collapses inward. The Cluely case is the public illustration: $20.3M raised on viral marketing built around deliberately provocative messaging, then a March 2026 retraction of the $7M ARR claim that had carried the funding round. Distribution-without-retention reads as a moat from the outside and a campaign from the inside; the metric that distinguishes them is repeat usage of the underlying product rather than first-touch impressions of the founder's posts.

Physical infrastructure runs on a decade-scale moat that is real on its own time horizon and exposed to model-generation-scale geopolitical risk. The firms holding lithography, foundries, and energy infrastructure are concentrated in jurisdictions whose long-run stability is not guaranteed across the playbook's relevant horizon. The moat is durable on the decade and uncertain on the quarter where the geopolitical axis is concerned.

Taste runs the most operator-dependent risk profile of the six. The moat is durable per-operator and fragile per-firm — taste rarely transfers cleanly to the firm at exit, does not always survive the founder generation, and can be replicated by a successor only when an operator with similar taste is hired or developed before the founder leaves. The exit price for taste-based businesses reflects the discount the market applies to non-transferable advantage.

Vertical integration compounds the longest of the six because the capital threshold is prohibitive, but the same threshold concentrates the moat among fewer than ten entities globally. The integration play is therefore a path to oligopoly rather than to broad market participation; the structural risk it carries is regulatory consolidation pressure rather than competitive erosion.

For startups (5-50): pick at least two of the six moats before raising the next round. The Skill File Test is the diagnostic. Founders raising on factory-speed alone face a 90-day window before the next factory operator clones the surface, and the moat layer is what extends that window into a defensible position.

For enterprise transformations (500+): audit the firm's top three margins for information-asymmetry exposure. The compression window starts the day customers can run the firm's underlying calculation themselves, and the CFO diagnostic above is the formal procedure for sizing the window.

The moat layer is defensible right now in each of the six categories, and the time-bound character of that defensibility is itself why the chapter that closes the playbook takes up the political and regulatory consequences. The cyber-economy fork between hyper-efficient self-regulating markets and centralized surveillance commerce runs through exactly the territory the moats above define. The next chapter develops what happens to the political and regulatory environment when the moats above hold for a small set of incumbents and the moats below collapse for everyone else.