# Selection - which startups survive Game 3 ← [Startups](README.md) · [III.A](../README.md) · [Index](../../../README.md) --- [Formation](formation.md) says attempts get cheap. This page says which attempts can hold value. The filter is not "AI startup vs not" - almost every new firm will use models. The filter is [Game 3](../../../02-games/3-firms.md): **does the company own an inelastic complement, or is it selling cognition that will be free in ~24 months?** ## The category map Read each row as a claim about *where rent can sit*, not about moral worth or technical impressiveness. | Category | Complement held | AI-era prognosis | |---|---|---| | Pure feature / wrapper SaaS (no data lock-in) | None | Dies or becomes free; Red Queen for buyers | | Vertical SaaS with workflow embed only | Switching cost (temporary) | Eaten from below as internal tools cheapen → [software](../software.md) | | Proprietary real-time / flow data | Data | Durable if the data is *generated by use* and not scrapable | | Marketplace / network | Liquidity and default placement | Durable if chicken-egg already solved; hard to start cold | | Regulated / licensed product | License + liability capacity | Margin possible until rules reprice; slow entry is the moat | | Balance-sheet / risk-bearing | Capital that can be sued | [Insurance](../insurance.md) logic; hard for thin startups | | Physical product + software | Atoms, supply chain, service network | Capex and ops are the moat; software is table stakes | | Energy / land / interconnection adjacency | Powered sites, rights | Substrate-scarce; not a "software" story → [energy](../../../01-substrate/energy.md) | | Robotics / autonomy hardware | Manufacturing + deployment data | [Robotics](../../physical/robotics/) clocks; capital-heavy | | Developer tools / infra | Defaults, standards, community | Transient unless you become the default or the data plane | | Open-core / commodity model layer | None durable | Fast-follow and inference collapse → [inference economics](../../../01-substrate/inference-economics.md) | | Trust / brand / compliance infrastructure | Verification, attestation, audit | Under-supplied public good; [compressed](../../../00-overview/compressed.md) priority | | Talent / services with product skin | Relationships | Often a services firm in disguise; scales poorly, dies slowly | The uncomfortable cut: **most of what the 2010s trained founders and VCs to fund sits in the top rows.** Seat-priced software with nice UX and no unique data was a wonderful business when building it was hard. When building it is easy, it is a commodity process sold into a market that can rebuild it. ## Why "AI company" is not a category Calling a firm an "AI startup" predicts almost nothing about survival. The model layer is [not a durable moat](../../../01-substrate/inference-economics.md) on a ~24-month lag; open weights compress differentiation further ([Game 1](../../../02-games/1-labs.md)). What matters is the *application-layer* complement: - **Does use generate proprietary data the next model release cannot erase?** Telemetry from industrial equipment, claims outcomes, clinical workflows with consent, marketplace transaction graphs - yes. Generic chat logs of public knowledge - no. - **Does distribution sit inside the product?** Default placement in a workflow OS, exclusive channel partnerships, regulatory filing requirements that create switching friction - yes. "We have a landing page and ads" - no. - **Can the firm be the liability principal?** If customers need someone to sue when the agent is wrong, a startup with no capital and no insurance is a worse product than a mediocre incumbent with a balance sheet - accuracy is secondary ([insurance](../insurance.md), [law](../law.md)). - **Is the hard part atoms or permits?** Then AI is an input cost decline, not the product. > **Prediction:** among venture-backed software companies founded 2025–2028, those whose Series B narrative still centers "model quality / AI features" rather than a named complement (data, license, distribution, balance sheet, physical lock-in) are **<25%** as likely to reach a >$100M ARR durable outcome as those with an explicit complement thesis (~60% confidence; resolves on 2032 outcomes). The number is less important than the direction: **feature narrative underperforms complement narrative.** ## Selection pressure inside the customer Customers are under Red Queen too. They must adopt AI tools to keep costs in line with competitors, and they capture little surplus ([Game 3](../../../02-games/3-firms.md)). That produces a specific buying pattern startups misread as product-market fit: 1. **Pilot theater** - many tools evaluated; few integrated deeply. 2. **Build vs buy flips** - when internal rebuild costs $5k–50k instead of $500k, the long tail of vertical vendors loses the "too hard to rebuild" defense. 3. **Platform consolidation** - IT and security prefer fewer vendors with indemnities; a sea of wrappers loses to the suite that is already on the master service agreement. 4. **Outcome pricing demand** - buyers ask to pay for results ([B5](../../../07-indicators/diffusion/economy.md)); startups without reliability or capital to stand behind outcomes cannot reprice and churn when seats stop correlating with value. So "we have design partners and LOIs" is even weaker evidence than it was in 2019. The selection filter is **paid production retention after the first model-generation cycle**, not the pilot. ## Where new firms still have an edge over incumbents Entrants win when the incumbent's structure is the liability: | Incumbent drag | Entrant edge | |---|---| | Seat-revenue addiction | Can price outcomes or usage from day one | | Legacy codebase and process | Greenfield agents without the partial-rewrite tax | | Channel conflict (sales force protects old product) | Sells the threat the sales force refuses to | | Regulatory capture that *blocks* product change | Only if the entrant can get licensed - often they cannot | | Cultural inability to cannibalize | Ships the cannibal | The edge is real and temporary. Incumbents that hold distribution and licenses can buy or copy once the pattern is clear - which is why [exits](exits.md) skew toward acqui-hire and feature absorption rather than independent IPO for pure software. The durable independent outcomes cluster in rows of the category map where **buying the complement is harder than buying the team** (regulated licenses, physical networks, exclusive data rights, entrenched marketplaces). ### Complement erosion as the silent killer Even the "good" rows are not permanent. [Uncertainty 7](../../../06-uncertainties/complement-erosion.md) says the scarce list has a half-life: robotics may eventually attack physical lock-in; license regimes may reprice; proprietary data may be synthetic-replicated. Selection is therefore **dynamic** - a startup that wins 2028 on workflow data may lose 2034 when the model layer internalizes that workflow. The correct founder move is not "find a moat and relax" but "find a moat and keep moving to the next one," which is operationally the same as building a real company rather than a feature. ## Failure modes - **If distribution platforms intermediate everything**, selection still happens but the winner is the platform's ranking algorithm, not the independent complement owner - rents centralize one layer up. - **If learned verification ([U5](../../../06-uncertainties/learned-verification.md)) opens currently hard domains**, categories that look "accountability-safe" for humans compress and the license/liability rows thin. - **If internal rebuild stays hard** (enterprise context debt dominates forever), vertical SaaS lives longer and this page's eat-from-below claim was early - score with churn and build rates, not with demo culture. **Name the complement in one sentence or you do not have a company - you have a demo with a burn rate.** That is the selection test a diligence process can actually run. --- **Related:** [Game 3](../../../02-games/3-firms.md) · [Software](../software.md) · [Inference economics](../../../01-substrate/inference-economics.md) · [Assets](../../../09-macro/assets.md) · [Uncertainty 7](../../../06-uncertainties/complement-erosion.md) · [Formation](formation.md) · [Exits](exits.md) · [Ideas](ideas.md) (constructive list) **Next:** [Venture](venture.md)