Venture - power laws under flooded deal flow#
Contents
Venture capital is not "funding innovation" as a public service; it is a portfolio strategy for power-law equity outcomes under high uncertainty. That strategy still works when a tiny fraction of companies produce nearly all returns. What cheap founding changes is everything upstream of the return: how many things look fundable, how expensive diligence must be to tell them apart, and which partners have an edge.
The classical machine, restated#
A fund raises a fixed pool, takes 2-and-20 (or tighter), writes a large number of checks, and needs a handful of outcomes that return the fund. The model assumes:
- Asymmetric upside - equity in a winner pays for many zeros.
- Selection skill or access - partners see better deals or pick better.
- Governance and follow-on - ownership maintained into the rounds that mint the outcome.
- Illiquidity premium - LPs accept decade-long capital locks.
None of those four is killed by AI. What breaks is the ease of assumption 2 when the left tail of the quality distribution floods with demos that look like 2018 Series A products and are worth a weekend of inference spend.
Signal collapse at seed#
Formation collapses product-build cost. The seed deck, the demo, the landing page, the synthetic user research, and the polished technical architecture all get cheaper to fake - not always fraudulently; often as honest overconfidence with better tools. That produces:
| Old seed signal | New status |
|---|---|
| "Working product" | Near-worthless as differentiator |
| "Technical team shipped fast" | Common; speed ≠ judgment |
| "Design partners love it" | Inflated; pilots are cheap to run |
| "Proprietary model / secret sauce" | Usually vapor or 24-month moat |
| "Real revenue with retention" | Rises in relative value |
| "Own the data / license / channel" | Rises in relative value |
| "Founder has domain scars" | Rises - harder to fake than a demo |
The rational response is to reprice diligence toward the residual hard signals: unit economics under load, contractual distribution, regulatory position, data rights with actual exclusivity, and founder history in the domain's ugly parts. Funds that keep optimizing for demo quality and narrative fluency will systematically overpay the Red Queen layer.
by 2030, top-quartile seed funds will show systematically higher fraction of first checks into companies that already have either (a) ≥$1M ARR with net revenue retention >100%, or (b) a named non-software complement (license, exclusive data, physical network), relative to 2020–24 vintages (~70% confidence). The "pre-product seed on a slide" does not disappear - it concentrates further into pedigree and brand-name founders, which is a different market.
What happens to check sizes and ownership#
Two equilibria are compatible with the same thesis; the data will decide which dominates:
Equilibrium A - smaller seeds, more of them. Formation is cheap, so capital need per attempt falls; funds spray more tiny checks and accept worse ownership, hoping power law still saves them. This is the "spray and pray scaled" world. It fails if follow-on discipline is weak and the fund is diluted out of winners - or if winners are so rare among pure software that ownership in the median attempt is worthless.
Equilibrium B - fewer seeds, harder gates, larger ownership in complement-native companies. Capital concentrates in the capital-heavy and regulated categories that still need real money (selection), and pure software is expected to reach revenue before a meaningful check. This is the "venture looks more like classic growth + specialty" world.
Base case of this corpus: A at the very early stage (angels, micro-seeds, platform-tied programs), B at institutional seed and above. The industry dual-tracks. Confusing the tracks - applying spray metrics to capital-heavy deep tech, or growth-equity diligence to a weekend wrapper - is the operational failure mode for both GPs and LPs.
The platform and corporate diverticulum#
Not all "venture" is independent funds. Corporate VC, hyperscaler startup programs, and model-provider marketplaces increasingly offer distribution as the capital: credits, co-sell, default placement. That capital is cheap and strategically priced - it buys optionality for the platform, not optimized LP returns. Founders trading equity for distribution may be rational; they should model it as selling the complement early (assets), because the platform often ends up owning the scarce layer while the startup owns the replaceable feature. Score these deals on who holds the customer relationship at year five, not on the press-release valuation.
Fund returns and the Red Queen#
If most software adopters do not capture surplus (Game 3), and most pure-software startups sell into that layer, then aggregate VC returns in pure application SaaS compress even while a few distribution or data winners print the decade. That is compatible with:
- Record outcomes at the extreme (power law intact or steeper).
- Miserable mean and median outcomes for the category.
- LP disappointment in "AI funds" that were secretly diversified feature portfolios.
The 2027–29 capital / compute correction - if it arrives - hits venture through the second order: exit multiples compress, dry powder pauses, and only funds with dry powder and selection discipline buy the survivors. A capital winter is not a capability winter (winters); it is a sorting event that makes selection visible in the marks.
LP behavior and the narrative cycle#
LPs allocate to stories. "AI" as a sleeve will be over-allocated relative to complement-native underwriting until a vintage marks badly in public. The base-rate pattern is cycles: railway and electrification capital both funded the right technology and still destroyed late-arriving equity. The analogous mistake now is funding exposure to the theme rather than ownership of the constraint. Watch whether new AI-specialist funds' portfolios cluster in wrappers or in data/license/physical - the clustering is the diligence quality signal before TVPI is knowable.
Failure modes#
- If outcome pricing and agentic reliability make software margins expand industry-wide (not only in licensed niches), Game 3's application to startups was too harsh and pure SaaS venture works again - watch B3/B5 jointly.
- If a few platforms capture all distribution, venture returns concentrate in companies that are the platforms or have exclusive access - independent selection skill matters less than access relationships.
- If formation collapse is overstated (enterprise software remains expensive to ship safely), seed signals degrade less and classical product diligence retains value longer.
Underwrite the complement, not the demo. Any process that cannot name what the company owns that intelligence cannot manufacture is not diligence - it is theme investing.
Related: Capital · Game 3 · Selection · Exits · Cycles · Inference economics