Business Moats in AI

Overview

Michael Bloch's framework for defensibility in the AI era. The core filter: AI compresses the time it takes to DO things. It does not compress the time it takes for things to HAPPEN.

Platform Shifts and Value Capture

Benedict Evans frames AI as the latest in a recurring series of platform shifts (mainframes → PCs → web → smartphones → AI), each of which reshuffles dominance: old gatekeepers fall, new ones emerge, and the first movers are very rarely the companies that capture lasting value. PCs had many early entrants before IBM and Microsoft; web browsers, search, social, and smartphones all followed the same pattern. The implication for AI: we should presume that half of what's being built now won't be the answer, and that the eventual winners may not yet be obvious.

The strategic fork for model labs maps directly onto the moat framework: do you compete down the stack on capital (chips, data centers, power — the way hardware and infrastructure industries work) or up the stack on network effects, product, and go-to-market (the way software has always worked)? What you can't do is sit in the middle with a commodity model burning hundreds of billions. OpenAI's position illustrates this tension: massive mindshare but a commodity product with no differentiation, no platform, and no infrastructure of its own — scrambling to bundle everything from app platforms to browsers to e-commerce on someone else's balance sheet.

The absorption → innovation → disruption deployment cycle matters for timing: most enterprises are still in phase one (absorbing AI into existing workflows — code, marketing, customer support). Only about a third of large companies have even one generative AI product in production. The disruptive phase — where AI enables fundamentally new business models and market structures — is still ahead.

The Five Moats

1. Compounding Proprietary Data

Not static datasets (those get synthesized or worked around). The moat is living data: proprietary information continuously generated through defensible operations. Example: Orchard AI mounts cameras on farm equipment tracking billions of fruit across multiple growing seasons. You can't replicate it by training a model on public data.

2. Network Effects

Every user makes the product more valuable for every other user. DoorDash: every driver → faster delivery, every restaurant → more choice, every customer → better economics. "Clone the app overnight. The drivers, restaurants, customers in ten thousand cities don't come with it." Cold start problem may get harder as AI makes it trivial to build competitors.

3. Regulatory Permission

Governments move at the speed of politics, not technology. Bank charters take years. FDA approval takes years. Surface area of regulation is expanding because higher AI capability → higher stakes. Example: Anduril needs procurement clearances and classified contracts.

4. Capital at Scale

The endgame is physical. Chip fabs cost $20B. Nuclear plants cost $10B. "There's a reason Elon is raising $75B even while saying money might not matter in 15 years." Capital access = institutional trust + track record + relationships built over decades.

The scale is staggering: in 2025 the big four platform companies spent close to $400 billion on AI infrastructure, up 4x from two years prior, with growth rates expected to increase further source(https://www.youtube.com/watch?v=FtG8fMGHbNY). Getting access to electricity is now a bigger constraint than getting chips from Nvidia. Microsoft alone spent 45% of revenue on capex last quarter and is adding ~$50B in leasing; Meta has done two $30B infrastructure deals; Oracle may need to borrow 100% of revenue to meet commitments. This capital intensity is itself a moat — the money is coming from cash flow of enormously profitable companies, not capital markets, making it inaccessible to new entrants.

5. Physical Infrastructure

Factories, power plants, battery networks, data centers. Example: Base Power deploying thousands of battery units across Texas homes while building own manufacturing. "You can design the system in a week with AI. You cannot manufacture, install, and interconnect thousands of units in a week."

What's NOT a Moat Anymore

• Workflow embeddedness — Switching cost is really just engineering time in disguise
• Ecosystem lock-in — AI can rebuild integrations as fast as you describe them
• Software scale — Spreading engineering costs across millions of users stops mattering when engineering costs approach zero

These were moats against the scarcity of intelligence. "That's the one form of scarcity we know is ending."

Evans' benchmark data reinforces this: the top 10 foundation models cluster within 5-10% of each other on general-purpose benchmarks, with a new leader every week that quickly converges back to the pack. The models are commoditizing in capability even as usage concentrates around distribution advantages — OpenAI leads on users, others lead on benchmarks, and the gap between them narrows continuously.

Open Questions

• Human attention as moat? — When content creation cost drops to zero, brands that already hold attention may have the most compounding advantage.

Two Paths for Software Companies

David George's framework (a16z, Apr 2026): the comfortable middle is over for software. Public markets have repriced the sector. Only two credible paths to durable equity value:

Path 1 — Accelerate growth (+10pp revenue): Build genuinely new AI-native products within 12-18 months. Not bolt-on copilots — products that move the total growth rate. Requires: four-person pods (collapse design/product/eng), 50% of R&D on net-new, token/consumption pricing (not seats), and finding the ~5 people in the org who will deliver 100x value.

Path 2 — Rebuild for 40%+ true margins: Including stock-based compensation as a real expense. Requires flattening management, killing committees, standardizing implementation, raising prices where you own the workflow. The Broadcom/VMware playbook: radical cost discipline + product simplification → 61% adjusted EBITDA.

Key insight: The new growth sits in tokens, consumption, automations, and machine-driven workflows. Seat-based revenue is where customers look to cut costs. "If you are not in the token path, you are not standing in the fastest-growing part of the budget."

Token Price Discrimination

Anish Acharya's observation (Feb 2026): despite claims of model commoditization, consumers pay $200-300/month for ChatGPT Pro/Claude Max/Gemini Ultra, and 75% of public SaaS companies have raised prices since ChatGPT launched. Value of a token varies enormously: a token unlocking a drug design vs a weather query, despite similar generation costs. Per-token price discrimination may be as important a business model innovation as "renting software" was 15 years ago.

Anthropic Growth Case Study

Anthropic's trajectory illustrates how moats compound in practice: $0 → $100M ARR (2023) → $1B (2024) → $19B+ (early 2026) — 10x year-over-year growth sustained across three years. Head of Growth: "Historically, we were very much the smallest, least well-funded player in this space. We didn't have the free cash flow or distribution of a Meta or Google. We didn't have the first mover advantage of an OpenAI."

Their internal growth program "CASH" (Claude Accelerated Sustainable Hypergrowth) uses Claude itself to automate growth experimentation. The meta-lesson: the product is its own growth engine when it's genuinely useful — "Claude is growing itself at this point." This echoes the token-based moat thesis: Anthropic's moat isn't software lock-in but compounding usage data, research capability, and the flywheel between model quality and adoption.

The Enterprise Data Risk: "The Big Rug"

The inverse of the compounding data moat: what happens when enterprises feed their proprietary data into AI lab tools?

goodalexander's thesis (Apr 2026): AI labs are executing what he calls the largest vampire attack in corporate history. The mechanism: enterprises adopt closed-source AI tools (models, coding assistants, agents) for productivity gains, generating millions of interactions and workflows. Those interactions are training data. Labs use that training data to build models that eventually outcompete the enterprises that fed them.

The arithmetic: Productivity at major enterprises is growing at 1.2-1.5% YoY. Token consumption is growing 400%+. The productivity gains are real but modest. The data leakage is enormous. Enterprise AI is characterized mostly by "vibe coding and slop" — low-value output generated primarily to test tools, not to create durable IP. Meanwhile, the high-value workflows and edge cases are exactly what model providers need for the next training run.

The trust architecture problem: This is the flip side of the regulatory permission moat. Enterprises with the most valuable proprietary workflows — in law, finance, pharma, defense — are also the most exposed if their process data leaves the perimeter. The enterprises that can maintain genuine data sovereignty (on-prem deployments, air-gapped models, contractual data isolation) may command a structural advantage as this concern becomes mainstream.

Counterargument: Harvey AI's approach suggests a middle path — proprietary process data stored inside the firm, with models rented and rotated based on performance (their "Model Selector"). The moat becomes orchestration and workflow IP, not the model itself. The data moat is genuinely valuable only when you control both the data and the model training.

See also: Vertical AI for Harvey's model U-turn and how vertical AI companies are navigating this.

The Opinionated Perspective Moat

fintechjunkie's counterpoint to the pure-infrastructure moat thesis: when anyone can build your product in a weekend, the last real moat standing is an opinionated perspective on the solution.

"Being able to build and understanding the best way to solve a problem aren't remotely the same thing." Building is mechanical now. But having a genuinely informed opinion about the right inputs, workflows, and outputs takes years of pattern recognition and listening to customers.

Why opinions compound into defensibility:

• Great product people ship constantly, plugging holes before users report them — copying them is hitting a moving target while reading their old blog posts
• Over time, hooks build: memory that doesn't port, context about preferences, integration work and muscle memory, data that makes the product smarter for your specific use case
• These switching costs don't show up on any spreadsheet but are real

The new competitive landscape: "Your competition isn't only other startups, it's also your user deciding they could probably just do this themselves on a Saturday." In that world, only one thing matters: having a perspective worth paying for.

This aligns with the "taste as moat" observation from Ann Miura-Ko's AI-native company visits: "When execution is nearly free, taste becomes the moat."

The Untrainable: Private Correctness as Moat

Sarah Guo's framework (Jun 2026) reframes defensibility around a single filter: can you train against it? Anything whose correctness can be cheaply verified — a compiler check, a test suite, a benchmark score — gets ground down to commodity by models iterating against that free verifier. What survives is work whose correctness is private, expensive to establish, and locked inside systems you can't access from outside.

The 2x2: Cross task saturation (commodity vs. frontier) with answer visibility (public vs. private). Saturated work with public answers is commodity tokens — open models own it. Frontier work with public answers (coding benchmarks) is where labs win, because owning the free eval counts for nothing. The prize is the last corner: frontier work whose correctness exists only in private. The best AI-native companies already show this — the vast majority of their inference tokens come from custom models, not generic open ones.

The legibility trap: Measurable work is what's leaving. The MIT study (Demirer et al.) quantified the gap: across 100,000+ developers, coding agents lifted code written by ~180% but code that actually shipped by only ~30% source(https://open.substack.com/pub/saranormous/p/the-untrainable). Writing got cheap; the rest — deciding whether a change is right for a decade-old codebase with undocumented dependencies — still runs through a person. As Noam Brown noted, the only sure way to evaluate an agent over a one-year horizon may be to run it for a year.

Permission and accountability, not intelligence: A model can be far smarter than any person and still has to be let in the door, and someone still has to put their name on what it does. Intelligence isn't the bottleneck — permission is, and so is liability. This extends the regulatory permission moat beyond government: every enterprise has its own trust architecture of security reviews, integration contracts, and named accountability.

Trust as the deadbolt: A majority of American doctors now open OpenEvidence daily. No amount of compute buys that habit. A lab can train a flawless medical model and still have no path into the physician's decision flow or UCSF's systems, because trust is built slowly, on relationships, not gradient descent. This resolves the "trust as moat?" open question from Bloch's framework: trust isn't speculative — it's already the binding constraint in high-stakes verticals.

The absorption frontier: Labs are pulling scaffolding into weights — retrieval, routing, tool use, reasoning policy — so the wrapper becomes the model. But a general agent must be ready for anything (expensive), while a focused application can tune one workflow to run on a fraction of the token spend and keep the difference. Margin pressure cuts both ways.

Private evals as moat: The evaluation that decides real money is private and per-firm. Harvey publishes benchmarks for law, Sierra for voice agents. You earn the right to define what "good" means in a field by being the one the field already uses. A foundation lab can't author that standard however smart it gets — the standing only exists inside the domain. Defining "resolved" or "safe clinical answer" falls to whoever already holds that authority.

Offense is harder than defense: Choosing what to build is the scarcest capability. The model will do whatever you point it at but can't tell you what's worth pointing at — and you can't benchmark that, so you can't train it. This is also why incumbents don't take everything: they keep existing ground, and the next thing comes from someone who finds a use before the rest.

The Perennial "What If Big Co Builds This?"

Andrew Chen's historical framing: every technology wave produces the same objection — "what if IBM builds this?" (1980), "what if Microsoft builds this?" (1995), "what if Google builds this?" (2010), "what if <huge AI lab> builds this?" (today). Reality: when these waves happen, new markets are so large there will be tens of thousands of new viable companies, hundreds of unicorns, and a few iconic generational companies. Big cos play a role but can never compete with the open market.

"Pessimists ask 'what if they build it.' Founders ask 'what if I build it?'"

The AI SaaS Squeeze (Tyler Tringas) {#ai-saas-squeeze}

See also: Services as Software — Sequoia's thesis on the services-first pivot.

Tyler Tringas (Calm Fund) wrote a thesis in June 2025 (published Apr 2026) arguing AI-assisted coding is eroding the fundamental SaaS moat — and the predictions proved "very directionally correct," with effects faster and larger than expected.

The core thesis: SaaS's golden-era margins rested on one moat: it was hard to hire good developers and you needed lots of them. AI-assisted coding erodes this. As the difficulty and time to build shrink, fast-follow competition becomes dramatically easier. The result: multiple compression — SaaS valuations converging toward normal business multiples (5-9x profit vs. the historic 20x revenue).

Evidence: Public SaaS multiples returned to 2016-2017 levels (median ~6.7x ARR per SaaS Capital). Private acquisition offers are down significantly from 2022 peaks, with typical decent SaaS businesses seeing 4-6x revenue from non-strategic buyers. "Merely being a SaaS company is no longer a ticket to premium ARR multiples."

The Red Queen effect: Businesses may double revenue YoY but end up worth the same or less because the valuation multiple decreased over that time.

Three headwinds:

1. Price pressure — More competitors building faster and cheaper
2. Customer overwhelm — Buyers bombarded with AI pitches, slower to close, lower willingness to pay ("any day now I'll do this on my $20/mo ChatGPT subscription")
3. LLM platform competition — OpenAI, Anthropic, Google can release competitive vertical features on massive user bases

The Fractal of AI Panic: The Red Queen dynamic operates at every level simultaneously. Individual workers race to become AI-fluent before a colleague replicates them. Companies sprint to avoid being disrupted by leaner AI-native competitors. Public companies face activist investors and analyst pressure to show AI wins. The fractal cascades downward: board → CEO → direct reports → individual contributors. Each level runs so hard they have no bandwidth to notice the structural pattern — that they're competing on a dimension that is being commoditized, not a dimension where they can actually win.

The Founder Playbook — 6 strategic responses:

1. Lean into AI — Table stakes. Use AI for development and integrate it into the product. But alone, this won't protect against the headwinds.

2. Expand and bundle — Take more products through the "zero to good enough" phase using AI, sell as add-ons. Commoditize your complement. Every.to's approach: cross-bundling software with a paid media subscription.

3. Micro-acquire the competition — Multiple compression hits weakest performers hardest. Acquire smaller competitors at depressed multiples to drive growth. Especially effective for things with built-in distribution (plugins, Chrome extensions, newsletters).

4. Sell solutions, not software — The conventional wisdom flip. Use your own software, don't just sell it. Add done-for-you services at 5-20x the ARPU. "Customers would rather pay $300-500/mo for 'the books are done' than $30-50/mo for bookkeeping software." Blend humans and AI now; replace humans gradually as models improve.

5. Consider exiting — If an exit would be life-changing, take it seriously before multiple compression becomes consensus. "Get out of the business of only selling software as soon as possible."

6. Ignore me — Building a company isn't entirely about maximizing equity value. If you enjoy it and you're making money, carry on.

Key insight on AI coding reality: "Vibe coding is basically BS" — but acceleration for someone with programming skills is staggering. Work that required entire teams can now be done by one person running multiple AI agents in parallel. The project of making models all-purpose better is thorny, but models will keep getting phenomenally better at coding specifically.

The conventional wisdom flip: SaaS dogma said be hyper-specialized and avoid services revenue. Tringas argues the opposite: bundle aggressively, add services, cross-sell non-software products (community, media, events). The strategic responses that work are precisely those that traditional SaaS wisdom warned against.

Niche Construction: Escaping the Commodity Race

The Red Queen hypothesis (Van Valen, 1973) explains a recurring pattern in evolution and competitive markets: when all competitors adopt the same fitness-improving advantage, nobody gains ground — they just run faster to stay in the same place. AI adoption in 2025–26 shows this dynamic clearly: every company deploys AI, announces productivity gains, and ends up in the same relative competitive position.

The winning move isn't to adopt the commodity fastest. It's niche construction: rather than adapting to your environment, you modify the environment in ways that shift the competitive rules in your favor. Organisms (and companies) that do this don't just survive the commodity wave — they define what the next competition is about.

Three historical examples from the BuccoCapital thesis:

• Amazon vs. the internet: While retailers optimized their websites and Google ad spend, Amazon asked: What happens when digital distribution goes to zero? They built physical warehouses, logistics networks, and last-mile delivery — a niche nobody could replicate quickly, and one that mattered precisely because the digital layer became commoditized. Amazon's website is still ugly; their moat is physical infrastructure.

• GM (Alfred Sloan) vs. the Model T: Ford made cars a commodity. Sloan asked: If everyone can own a car, why not one that says something about you? GM built laddered brands (Chevrolet → Pontiac → Oldsmobile → Buick → Cadillac), added annual model changes, and invented consumer auto financing. The niche: identity and aspiration layered on top of the commodity.

• Toyota vs. mass production: When Ford-style scaled production became the norm, American automakers optimized inside the existing paradigm. Toyota asked: If scale is commoditized, what isn't? They constructed a niche around waste elimination and quality control (the Toyota Production System) — winning on a dimension where their competitors had grown complacent.

The pattern in every case: the winner recognized when the critical input had been commoditized, stopped optimizing for that input, and invested heavily in a different scarce variable. The loser kept running the old race harder.

Applied to AI: Companies racing to adopt AI fastest are optimizing for intelligence as a scarce input. That input is becoming a commodity. The five durable moats above (data, network effects, regulatory permission, capital, physical infrastructure) are all things that can't be parallelized by AI and can't be commoditized quickly. Building those is the niche construction play. Optimizing AI token usage metrics and winning "most AI-native company" press releases is the Red Queen trap.

The Terminal Value Collapse Thesis

Chamath Palihapitiya inverts the entire moat framework with a thought experiment: what if AI lowers the cost of disruption so dramatically that no company can credibly project free cash flow beyond five years? If moats become temporary by default, equities should be priced not as discounted streams of future cash flows but as short-duration multiples of current earnings — the same way markets priced taxi medallions right before Uber.

The repricing math: Start from the risk-free rate (~4.5% on 10-year Treasuries), add the equity risk premium (4–5%), and the required return on a stable, no-growth equity lands at roughly 9% — implying a baseline multiple of 10–12x FCF. Now add a 20% annual probability of AI-driven obsolescence. The expected business lifespan drops to ~5 years, and the rational multiple compresses to ~3.9x FCF. At 30% disruption probability: ~2.8x. At 10%: ~6.5x. The S&P 500 currently trades at ~22x earnings; repricing to the midpoint (5x FCF) would imply a drawdown from $58T to $14T in aggregate market cap.

Historical precedent: Markets have already applied this logic sector by sector. Newspapers compressed from 12–15x EBITDA to 2–4x between 2005 and 2015 as digital advertising destroyed the print model. Department stores fell to 3–6x FCF as Amazon dismantled brick-and-mortar economics. Oil majors traded at 4–6x FCF when markets began pricing stranded reserves. NYC taxi medallions collapsed from over $1M to under $100K. In every case, the market correctly identified duration risk: real cash flows today, uncertain survival tomorrow.

The self-defeating paradox: AI infrastructure requires $300–500B per year in long-duration capital expenditure — investments that only make sense over 7–15 year horizons. But if markets reprice to 2–7x FCF, that capex becomes unfinanceable. The disruption engine disrupts itself. This creates an oscillating dynamic rather than a permanent regime: compression → capex drought → slower disruption → moats re-harden → multiples recover → capex resumes → cycle repeats.

Capital rotates to the physical world: Money flows toward assets insulated from AI disruption — energy infrastructure, farmland, toll roads, water rights, commodity producers, short-duration sovereign bonds. This reinforces the physical infrastructure and capital-at-scale moats: the same assets that are "hard to get" in Bloch's framework are also the assets that retain terminal value in Chamath's.

Nation-states fill the void: If private capital can't finance long-duration projects, sovereign capital steps in. Countries with high savings rates, large borrowing capacity, or patient state investment vehicles (US, China, Gulf states, Norway, Singapore) gain a structural advantage. Industrial policy stops being fringe; strategic infrastructure becomes a national security question rather than an ROE question.

The likely outcome is not a permanent 5x-FCF world but structurally higher equity risk premiums, shorter capital cycles, fatter tails, and periodic crises of confidence in terminal values. Even a partial compression — 30–40% rather than 90% — would represent the most significant structural shift in capital markets since the postwar era.

The tension with the five durable moats above is productive: Chamath's framework doesn't invalidate them so much as compress the timeline over which they compound. The meta-question becomes whether any moat can accumulate fast enough to outrun the accelerating disruption cycle — or whether the cycle itself is self-limiting.

The Meta-Moat

"Time that can't be parallelized." Network density takes years of human adoption. Regulatory approval takes years of political process. Infrastructure takes years to build. Data takes years to compound. Capital relationships take decades to earn.

Sources

• "The Only Moats That Matter" — Michael Bloch (tweet, Mar 2026) (link)
• "There are only two paths left for software" — David George (a16z, Apr 2026) (link)
• "Notes on AI Apps / Feb 2026" — Anish Acharya (tweet, Apr 2026) (link)
• "Claude is growing itself at this point" — Head of Growth, Anthropic / Lenny's Podcast (video, Apr 2026) (link)
• "The Big Rug" — goodalexander (Apr 2026)
• "ok this startup is cool but..." — andrew chen (tweet, Apr 2026)
• "The Last Moat Standing" — fintechjunkie (tweet, Apr 2026)
• "The AI-pilled compounding startup" — Ann Miura-Ko (tweet, Apr 2026)
• "The AI SaaS Squeeze" — Tyler Tringas (Calm Fund, Jun 2025 / published Apr 2026) (link)
• "Running Faster to Go Nowhere: The AI Adoption Trap" — BuccoCapital / Educated Guess (Apr 2026) (link) — Red Queen dynamics in AI adoption, Fractal of AI Panic, niche construction historical case studies (Amazon, GM/Sloan, Toyota)
• "AI Eats the World" — Benedict Evans / Slush (video, 2025) (link) — Platform shift framing, model commoditization data, capital-vs-network-effects fork for model labs, $400B infrastructure spend, absorption-to-disruption deployment cycle
• "The Untrainable" — Sarah Guo (essay, Jun 2026) (link) — Private correctness as moat framework, the legibility trap (measurable work → commodity), permission/accountability > intelligence, trust as deadbolt, absorption frontier, private evals as defensibility, MIT coding agent data (180% written / 30% shipped)
• "The Collapse of Terminal Value" — Chamath Palihapitiya (tweet, Jun 2026) — Disruption repricing framework: if AI makes moats temporary, equities compress to 2–7x FCF; historical precedents (newspapers, retail, energy, taxi medallions); self-defeating paradox of AI capex; capital rotation to physical assets and sovereign investors