a16zHow AI Is Rewriting the Power Law of Venture Capital
CHAPTERS
- 0:00 – 1:34
Power law goes systemic: AI makes capital compound advantage
The conversation opens with the claim that power-law outcomes are becoming more extreme—and no longer confined to venture as a niche phenomenon. AI changes the mechanics: capital can now be converted directly into compute, which can rapidly improve products and reinforce competitive moats.
- •Power law in tech investing is more extreme than in the last 10–20 years
- •In AI labs, money buys compute, and compute can directly improve capability/product
- •Traditional risk of “too much capital” (coordination/hiring overhead) is reduced in AI contexts
- •Scale advantages and compounding effects are stronger in frontier AI businesses
- 1:34 – 4:02
Venture-backed IPOs, bigger outcomes, and the expanding ceiling of value creation
Jen frames the question of where venture-backed IPO value goes next, as outcomes move from $10B to $40B+ and potentially $100B. The group argues that each tech wave creates more total market cap than the previous one, pushing more value into fewer, larger companies—often while they stay private longer.
- •Top-decile outcomes have expanded dramatically (e.g., $10B to $40B+)
- •Startups staying private longer appears structural, not cyclical
- •Prior cycle created enormous market cap; next wave expected to be larger
- •Frontier model companies concentrate trillions of potential enterprise value
- 4:02 – 7:40
From software TAM to labor TAM: why AI’s market size is hard to model
Aram and David explain why AI isn’t just “the next software wave.” Because AI targets tasks and labor costs—not just IT budgets—its addressable market can be an order of magnitude larger, making historical TAM frameworks misleading.
- •AI reached ~$100B revenue far faster than SaaS (4 years vs ~15)
- •AI attacks broad GDP categories simultaneously (transport, labor, services, coordination)
- •TAM shifts from software spend to economic value of tasks/labor
- •Labor spend dwarfs software spend, implying far larger upside even if work is reinvented
- 7:40 – 11:08
Stack debates vs reality: ‘everything might work’ and category power laws
They address the common “which layer wins?” debate (labs vs apps vs open source). While they believe in power-law outcomes within categories, they argue the overall market expands so rapidly that multiple layers can succeed simultaneously—though many individual companies will fail.
- •Zero-sum framing (labs vs apps vs open source) may be too limiting
- •Power law is strongest within each category: the leader captures most value
- •AI expands the number of categories, creating more ‘shots on goal’
- •Venture tolerance for losses enables broad category experimentation
- 11:08 – 12:51
Why consistency is now the core venture edge (and how rare it is)
Aram presents allocator data showing how few firms deliver consistent top-tier venture returns. The implication is that access to category-defining companies across vintages—and proper ownership/sizing—is what drives persistent outperformance, especially as dispersion increases.
- •From ~3,000 US VC firms, only ~20 achieved consistent 3x net performance over two decades
- •Consistency correlates with repeated access to category-defining companies
- •Logo access isn’t enough: ownership (early) and sizing (late) determine fund impact
- •Average venture returns (1–2x net) often don’t justify illiquidity vs other asset classes
- 12:51 – 20:32
Death of the middle: founder preference, platform firms, and seed complementarity
They define the “middle” in venture and why it struggles: founders prefer partners who de-risk outcomes via brand, resources, and lifecycle capital. Highly specialized early funds can coexist by entering earlier, while scaled platforms win hot deals through a services-and-reputation flywheel.
- •Founders select investors who reduce risk via resources, brand, and lifecycle support
- •Large platform firms invest fees into operating capability to win and improve outcomes
- •Specialized/pre-seed funds can win earlier when uncertainty is highest
- •A firm’s reputation and references create a flywheel for repeat access
- 20:32 – 22:53
Venture has structurally changed: four ‘modes’ and AI making diligence harder
Aram breaks venture into four segments (pre-seed/seed, messy middle, big firms, dedicated late stage) and argues AI accelerates pace and confusion. Rapid early “traction” can be misleading without renewal cycles, forcing investors to evaluate deeper market demand signals rather than surface ARR metrics.
- •Four venture approaches: pre-seed/seed, messy middle, big firms, dedicated late stage
- •AI drives faster, larger rounds and noisier early traction signals
- •Early ARR can be inflated or non-representative (cohort selling, redefining ARR)
- •Market-demand validation requires customer texture, not just cohort spreadsheets
- 22:53 – 25:34
How to underwrite in the AI era: founder judgment, customer pull, and post-model step changes
David explains why financial analysis alone can’t resolve early AI ambiguity. They focus on whether customers truly demand the product, with examples of step-function improvements when model capability changes (e.g., reasoning models improving real usage), turning skepticism into enterprise pull.
- •Founder judgment and close founder relationships matter more when data is immature
- •Key question: ‘Is the market demanding more of your product?’
- •Customer interviews and usage patterns can reveal true pull before renewals exist
- •Model breakthroughs can transform product utility and adoption dynamics quickly
- 25:34 – 33:39
LPs fear ‘catching a falling knife’: incentive misalignment and portfolio sizing
Jen and Aram discuss why LP behavior often lags technological shifts: LPs are penalized for visible mistakes more than for missed upside. They argue LP success hinges on access, selection, and especially sizing—because small allocations to big winners don’t move the needle.
- •Common LP concern: timing/overheating—‘are we catching a falling knife?’
- •LP incentives favor avoiding errors of commission over avoiding missed winners
- •LP job: access + selection + sizing/portfolio construction
- •Over-diversifying across many VC funds often results in average returns
- 33:39 – 38:35
The legacy SaaS overhang: what happens to pre-ChatGPT vintages
They confront the difficult question of older software assets bought or valued under pre-AI assumptions. Some may adapt, but many face valuation compression and fewer exit paths if they can’t show AI resilience via renewed growth acceleration.
- •Legacy SaaS companies may be stuck: high private marks, limited IPO appetite, fewer PE buyers
- •AI resilience increasingly judged by growth acceleration, not just profitability
- •Diffusion beyond coding is still early—creating uncertainty about pace of disruption
- •Portfolio exposure matters: top funds may have limited NAV in legacy, more in AI-accelerating assets
- 38:35 – 40:57
Why ‘AI private equity’ isn’t a shortcut—and private credit risk is rising
They argue you can’t simply “add AI” to a legacy company and expect transformation, especially under leverage. Private credit and LBO-era software deals face pressure as public comps reset and AI threatens terminal values, raising leverage ratios and refinancing risk.
- •‘Put AI on it’ fails without workflow redesign and aligned leadership/board
- •Leverage amplifies operational mistakes (e.g., churn spirals from poor AI CS deployments)
- •LBO software deals from 2021–22 at high EBITDA multiples are stressed after valuation resets
- •Private credit is exposed as SaaS valuations compress and AI disrupts durability
- 40:57 – 44:01
Liquidity and exits: long holding periods, but fast liquidity in true outliers
They explore the main rational critique of venture: time to liquidity and slow distribution even after IPOs. The counter is that category winners can deliver outsized returns and sometimes faster liquidity (M&A or early IPOs), and many LP types prefer compounding over early exits.
- •Key pushback: extended time-to-liquidity; IPOs aren’t immediate distributions
- •Counterpoint: for true category winners, compounding is valuable and worth waiting for
- •Some AI-era outcomes may reach liquidity faster via M&A or earlier IPO cycles
- •LP preferences vary: endowments/family offices may prefer compounding; others need distributions
- 44:01 – 48:18
Looking ahead: the path to $100T companies and the real bottleneck—power, chips, and data centers
The closing discussion imagines what could create the next $10T–$100T market cap outcomes: consumer-native AI, robotics, autonomy, healthcare, and physical-world reinvention. Aram emphasizes the limiting factor is supply-side infrastructure—energy, grid speed-to-power, data centers, and chips—creating massive investment opportunity and risk.
- •Future value creation likely in consumer-native AI (beyond chatbots), robotics, autonomy, healthcare, and physical-world industries
- •AI adoption is still early relative to global knowledge-worker base and enterprise spend
- •Primary bottleneck is supply: energy, transmission/permissioning, data centers, chips/memory
- •Infrastructure constraints create large new venture-scale opportunities and strategic urgency
