a16zHow AI Is Rewriting the Power Law of Venture Capital
a16z
The Extreme Power Law of AI Venture 0:00
David George and Ron from Adia open by noting that out of 3,000 venture capital firms in the US, only 20 have delivered consistent 3x net returns over the past two decades. The power law, the pattern where a tiny number of winners capture almost all the value, is now more extreme than at any point in the last 10 to 20 years of technology investing. For the first time in George's career, capital itself compounds a company's advantage: pouring money into compute directly improves products, rather than creating the coordination problems that used to come from overhiring.
Frontier Labs Reshape the Landscape 1:01
SpaceX, OpenAI, and Anthropic together represent an estimated three and a half to five trillion dollars of potential enterprise value. Notably, before SpaceX went public, most institutional allocators had little exposure to it. This sets up the episode's core question of how portfolio construction and asset allocation may need to change now that the power law extends beyond venture capital into the broader market.
AI Touches the Entire Economy 4:01
Ron points out that AI reached 100 billion dollars in revenue in just four years, compared to 15 years for SaaS, even though demand penetration is still early. AI is attacking every part of GDP, including transportation, labor, services, capital, and coordination, making it the first technology paradigm to touch roughly 30 trillion dollars of GDP at once. Because of this, he argues AI should be a core or super core allocation for investors, not a side bet. George adds that private markets now represent 5 to 6 trillion dollars in value, and that companies staying private longer is a permanent shift, not a temporary trend. Outcomes that once topped out around 10 billion dollars now reach 40 billion, with Anthropic and OpenAI potentially pushing that to 100 billion.
Estimating a Market Nobody Can Size 6:00
Using healthcare as an example, Ron notes that healthcare IT spending is only 60 to 100 billion dollars a year, but AI competes for the value of entire labor tasks like claims, billing, and administration, a trillion dollar opportunity. This means AI's addressable market can be ten times larger than traditional software markets, though nobody yet knows the eventual capture rate. Both investors admit to consistently underestimating how large these outcomes can become. George adds that labor spending dwarfs software spending by roughly 40 times, and that AI will not eliminate labor so much as reinvent the tasks people do, making comparisons to a simple next stage of software far too limiting.
Everyone Might Win, Not Just One Layer 8:01
Rather than framing AI as a zero sum fight between frontier labs and application layer startups, George argues the market may be big enough that most layers succeed, even though many individual companies will still fail. He distinguishes this from true winner take all dynamics, noting that within any given category the top company will still capture most of the market share while second place gets scraps. He expects a massive expansion in the number of viable categories, similar to how CRM software grew from a small niche into a huge market over 20 years. Andreessen Horowitz tolerates a 60 percent loss rate on early stage investments and 10 to 20 percent at growth stage, treating this as the acceptable cost of backing category leaders.
Consistency Requires Category Access 10:00
Ron explains that firms achieving consistent 3x returns almost always had access to category defining companies in every fund vintage, and that simply having a logo in the portfolio isn't enough since ownership size matters too. He notes fund returning math now exists at the late stage in a way it didn't before, since a single company can represent 5 to 10 percent of a large fund. Without access to top companies, investors get the average venture return, which Cambridge data puts at only 1 to 2x over the last decade, worse than private equity or public markets.
Death of the Middle 13:31
The conversation turns to why highly specialized niche funds and large multi-stage firms both thrive while mid-sized generalist firms struggle, a pattern they call death of the middle. George explains that founders choose partners who can derisk their outcome, which is why large firms invest heavily in operating resources and hundreds of employees to help portfolio companies win. This creates a flywheel: strong domain expertise and support lead to better outcomes, which attract future founders and generate references, reinforcing the firm's advantage over time.
How Small and Large Firms Coexist 16:30
Ron suggests that small pre-seed and seed funds can carve out a niche a round or two before large firms enter, since big firms often prefer to wait until there is more certainty before leading a category winner's A or B round. George agrees, noting Andreessen Horowitz's own seed investments tend to be larger checks aimed at serial entrepreneurs, and points to their Speedrun program as evidence the firm still wants visibility into the earliest stages even when it doesn't lead those rounds. Both agree that a strong late stage franchise depends on having an early stage franchise behind it, since early relationships make it possible to secure meaningful ownership later, something a brand-new late stage entrant writing a 500 million dollar check cannot easily replicate.
Traction Is Harder to Read Than Ever 20:32
Ron describes the venture landscape as split into four categories: pre-seed and seed, the messy middle, large multi-stage firms, and dedicated late stage funds, and notes his firm invests in the large firms and select seed funds but avoids dedicated late stage. He argues AI has made judging traction unusually confusing, since some companies jump from zero to five million in annual recurring revenue within a month with no renewal cycle yet to prove durability, sometimes selling to each other within the same accelerator cohort. For every nine companies with this kind of inflated traction, he says, there is one genuinely special company doing real revenue that will become the next breakout success. He points to Cursor as an example, recalling that people were still calling the company dead the morning its acquisition was announced.
Reading Real Market Demand 23:30
The clearest signal of a company's strength is whether the market is genuinely demanding more of its product, and you cannot answer that with financial analysis in the early months of a company's life. You have to talk to customers directly and understand the texture of the market. Harvey, the legal AI company, is the example given: early on it had smart positioning and signed up high-profile law firms, but actual usage looked mediocre next to other software companies. Once reasoning models arrived, usage flipped dramatically, lawyers got real value, and law firms went from fearing hallucinations to having clients demand they use the product. Catching that kind of shift early, ideally earlier than investors caught it with Harvey, is what the best early-stage investing is built on, since founders and early teams see this texture before anyone else does.
Why LPs Resist Change 25:31
The biggest pushback fundraisers hear from limited partners is fear of catching a falling knife, worrying that valuations and the market are overheated. But a deeper issue is that LPs face a structural incentive problem. A general partner can get fired for missing the next Facebook or Uber, an error of omission, while an LP rarely gets fired for missing out entirely, or even for investing safely in something like IBM. If an LP misses the frontier AI models, they still sit near the benchmark and keep their job. This misalignment means the upside of catching the next generational company often isn't compelling enough to change LP behavior, and the more useful role for LPs is learning how AI can improve their allocation across their entire portfolio, not just chasing the next power-law winner.
Access, Selection, and Sizing 28:01
An LP's job comes down to three things: access, selection, and portfolio construction or sizing. Historical data shows only about 20 firms out of 3,000 consistently perform well, so a portfolio holding 50 to 70 venture firms is unlikely to beat the average, since real outperformance requires concentration in that small consistent group. Sizing matters just as much as picking correctly. An LP who finds a great fund but only puts 1 percent of their capital into it might see that fund return 10x, yet it barely moves their overall portfolio. Compared to private equity, which might return 1.5 to 2x without the lockups or the roughly 60 percent loss ratio venture carries, venture only justifies its risk if allocators size their bets meaningfully.
AI Reshaping Every Asset Class 29:30
Public markets show the shift starkly: only 15 to 20 SaaS companies now trade above 10 times revenue, down from dozens a few years ago, and nearly every one of them shows growth acceleration tied to AI, whether in monitoring, security, or agent deployment. Private equity has changed too, now hunting for AI-native systems of record rather than workflow software growing 10 percent a year. Exits in venture have started exceeding private equity's biggest deals this year; EA and Medline, private equity's largest buyouts, were each around 50 billion dollars, while Cursor's acquisition by SpaceX (referred to loosely here as an M&A sale) was larger. Data shows that one percentage point of organic growth in public markets is now worth three percentage points of EBITDA, a complete reversal from 2021, when profitability mattered more than growth.
The Fate of Legacy Software 32:32
Software companies from the 2016-2021 vintage that grew steadily but never became AI-native face a real problem: assets once valued at 15 to 20 times EBITDA now trade closer to two times revenue, and buyers may not exist at all if a company isn't seen as AI-resilient. Reviving one requires something like what happened at Intercom, where the founder returned, rebuilt the product as AI-native, and scaled it before selling, an outcome described as nearly having to suicide the existing business, which is rare in private equity because boards and investors must all align behind such hard decisions. Even so, AI diffusion into the broader economy remains extremely early. The median US employee costs a company 12 dollars a month in AI spend, while the top 1 percent of companies spend 7,000 dollars per employee monthly, and even the most cutting-edge banks are only around 1 percent of headcount cost in AI tools, suggesting enormous room for growth ahead.
Liquidity Timelines and Future Giants 40:31
A common objection to venture investing is the long timeline to liquidity, since the average unicorn stays private over 10 years and even an IPO can take 12 to 24 months to convert into real distributions. The counterargument is that the top firms, roughly 20 out of 3,000, consistently deliver liquidity fairly quickly, sometimes faster than private equity, through means like early M&A rather than waiting on IPOs. Different LPs want different things too, some, like endowments, prefer to let winning positions keep compounding rather than cash out. Looking ahead, the next major value creation is expected not from chatbots, which are called a skeuomorphic interim step, but from robotics, autonomy (noting fewer than 10,000 Waymo vehicles currently operate in the US), healthcare (18 percent of GDP with almost no AI penetration yet in care delivery or drug discovery), and physical infrastructure like energy, data centers, and chips, where speed to power, not energy generation itself, is described as the real bottleneck holding back the next wave of trillion-dollar companies.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
