Context

a16z's Jennifer Li and David George sit down with Aram Verdiyan of Accolade Partners, a longtime limited partner in a16z's funds, to discuss how AI is reshaping venture capital's power law dynamics and, along the way, how to tell real product traction from hype. Most of the conversation is about fund construction and capital allocation, but a recurring thread runs through it that matters directly for product people: how do you know if usage of an AI product is genuine demand versus a sugar high from early buzz? George's account of Harvey, the AI legal product, and how its usage pattern fundamentally changed once reasoning models arrived, is the clearest example, and offers a concrete lens for evaluating any AI product's real traction.

The Big Idea

For the first time, capital itself can directly reinforce a company's product advantage, because dollars convert into compute, and more compute measurably makes AI products better, which is a mechanism that didn't meaningfully exist in prior technology cycles and is part of why today's power law in AI is more extreme than anything seen in the last two decades of venture investing.

George argues this changes the basic logic of how much capital should chase category leaders, while Verdiyan's data on venture fund performance shows how unevenly that capital advantage compounds across the industry.

Key Insights

Capital now compounds product advantage directly

George's central observation is that AI breaks a long-standing pattern where throwing too much money at a startup usually backfires, since hiring too many people too fast creates coordination overhead that slows a company down. With frontier AI companies, extra capital converts into more compute, and more compute demonstrably makes the product better, so money becomes a direct lever on product quality rather than an indirect one mediated by headcount. This is why George sees today's power law as structurally different, not just larger, than prior cycles.

Usage data, not early revenue, revealed Harvey's real inflection point

George describes watching Harvey, an AI product for lawyers, land impressive early customers at high-profile law firms, but actual usage looked mediocre next to other AI software companies at the time, something you could only see by examining deployment data rather than headline signings. He attributes this partly to hallucination concerns making lawyers cautious. After reasoning models improved, usage and engagement visibly took off, and the signal flipped from "law firms are nervous about this" to "clients are actively demanding their law firms use it." For George, that shift, from headline wins to real engagement to demand pulling adoption forward, is the pattern he looks for to distinguish genuine product-market fit from early buzz.

Only 20 of 3,000 US venture firms hit consistent 3x returns

Verdiyan cites Accolade's own research: across roughly 3,000 US venture capital firms, only 20 achieved consistent 3x net returns over the past two decades, meaning three to four separate funds each returning 3x net over a 20-year span, not just one lucky fund. He found the defining trait of those 20 firms was consistent access to category-defining companies in nearly every fund vintage, not any other strategy difference. The average venture return across the broader industry over the last decade was only 1 to 2x net, which he notes underperforms both private equity and public markets without the decade-long capital lockup venture requires.

"Everything might work" replaces the winner-take-all framing

George pushes back on a thesis that frontier labs would inevitably cannibalize application-layer companies, or that open source winning would necessarily be bad for closed labs. His observed pattern instead: essentially every layer of the AI stack is growing simultaneously, because the addressable market is so large that competition between layers isn't zero-sum the way it might be in a mature, fixed-size market. He still believes in power-law dynamics within any single category (the leader captures the majority share, second place gets a much smaller outcome), but expects a much larger number of distinct categories to emerge than existed in prior technology cycles, similar to how CRM went from a niche product category two decades ago to a massive one today.

AI's addressable market dwarfs software because it targets labor, not IT budgets

Verdiyan frames AI's total addressable market using healthcare as an example: healthcare spends $60 to $100 billion per year on healthcare IT, but AI is going after the value of the actual labor tasks it can perform (claims processing, billing, administration), a market worth over a trillion dollars. He argues this pattern, AI capturing a share of the economic value of the task itself rather than just the IT spend around it, means AI's total addressable market can be an order of magnitude larger than traditional SaaS or healthcare IT ever was, and that historically, incumbents in each new technology wave have ended up roughly ten times smaller than the market ultimately became.

AI adoption intensity is wildly uneven across companies today

Verdiyan cites data showing the median US company spends about $12 per employee per month on AI, while the top 1% of companies in the same dataset spend around $7,000 per employee per month, a roughly 580x gap. George adds that even the most AI-forward banks are only spending about 1% of headcount cost on AI tools currently. Despite this, the fastest-growing AI companies are already adding more revenue per month than mega-cap tech companies historically did, based on adoption from what George estimates as only 10 to 30 million truly engaged users out of roughly 1.5 billion knowledge workers in the US alone. The gap between top-decile and median adoption suggests most organizations, and most product teams building for them, are still extremely early relative to where usage intensity is heading.

Coding's AI success may be a misleading signal for other knowledge work

George cautions against assuming coding's rapid AI transformation predicts an equally fast transformation for other categories of knowledge work. His reasoning: code has three properties most business tasks lack. It is perfectly documented, verifiable (you can run it and see if it works), and simulatable (you can test it in a sandbox before it matters). Most other knowledge work, from legal analysis to strategic decisions, lacks at least one of those properties, which is why he expects diffusion of AI into non-coding knowledge work to take meaningfully longer than coding did, even though he remains highly bullish on AI overall.

Mental Models & Frameworks

Access, selection, sizing as the venture allocator's toolkit

Verdiyan frames a limited partner's entire job around three levers: access (can you get into the funds or deals that will contain the category winners), selection (can you correctly identify which firms or companies will be winners), and sizing (once you have access and made a selection, is your position large enough to actually move your portfolio's overall return). He stresses that access alone is not sufficient. An LP that gets into a winning fund but allocates only 1% of their portfolio to it will see that fund 10x in value without the overall portfolio outcome moving meaningfully, since the position was too small to matter.

Diffusion of a technology into the real economy as a demand signal

Both George and Verdiyan use "diffusion" as their core lens for judging whether AI's growth is durable: not how much buzz a technology has, but how deeply and broadly it has actually been adopted into real, paying, ongoing use across the economy. George explicitly separates hype-stage traction (a viral demo, a headline customer signing) from diffusion-stage traction (sustained usage that expands over time, like Harvey's post-reasoning-model adoption curve), and treats the latter as the only reliable signal of a durable business.

Trade-offs & Nuance

Fast early revenue can be a false signal, not a real one

Verdiyan describes a specific pattern he now watches for skeptically: an early-stage company reporting an extremely fast jump in annualized revenue (his example: zero to five million dollars in a month) with no renewal cycle yet to prove the revenue is durable, sometimes because cohorts of similar companies compare notes on how metrics are being reported. He contrasts this with a much rarer pattern: a company generating a more modest few million dollars in genuine, verified annual recurring revenue that goes on to become a category-defining outcome. His practical takeaway is that in the current environment, headline revenue growth numbers need much deeper verification than they used to, and founder judgment built from direct, sustained relationships matters more than financial analysis alone when a company has only been selling for a few months.

Late-stage venture math changed, but the underlying risk did not disappear

George notes that "fund-returning" outcomes (a single portfolio company large enough to return an entire fund) used to only be realistically possible at the early stage, but is now achievable in late-stage growth investing too, because top-decile outcomes have grown from roughly $10 billion to potentially $100 billion as companies stay private longer. He is clear this only works if a late-stage fund can concentrate 5 to 10 percent or more of its capital in a single category-defining company, which in turn depends on having an early-stage relationship with that company's founders built years earlier. A late-stage firm without that early-stage access, attempting to write a large check into an already-hot company cold, faces much worse odds.

Legacy software companies without AI-native features face a real reckoning

Verdiyan describes a 2016-2021-vintage software company growing at a merely healthy 30 percent, without AI-native features, as now facing a genuinely uncertain future: it may be priced at 10 to 20 times revenue on paper in venture portfolios, but neither public markets nor private equity buyers currently want to acquire that kind of asset, because AI resilience has become the dominant lens investors use to judge a software company's terminal value. Public SaaS valuations have compressed sharply as a result, with only 15 to 20 companies trading above 10 times revenue today, down from dozens a few years ago, and nearly all of those remaining premium companies are the ones showing visible AI-driven growth acceleration.

Practical Application

Judge product traction by usage depth, not headline signings

When evaluating whether an AI product has genuine market pull, look past which prominent customers signed on and instead examine actual usage and engagement data, the way George describes doing with Harvey. A prestigious early customer logo is a weak signal on its own; a rising trend in how deeply and how often that customer's own end users engage with the product, especially when a competing product's clients start demanding they switch, is the stronger one.

Weight a numbers-only sales pitch against unverifiable revenue claims

If evaluating a young AI company's fast early revenue growth, ask specifically whether that revenue has survived a renewal cycle, and be skeptical of extremely fast month-over-month annualized-revenue growth claims from very early-stage companies, particularly ones that came out of the same accelerator cohort and may be comparing notes on how they report metrics. Verdiyan's framework is to look for real, verified annual recurring revenue even if the absolute number is smaller, since that is a stronger signal than a large but unverified headline figure.

Match your task type to how mature AI diffusion actually is for it

Before assuming an AI capability that works well in one domain will transfer quickly to another, check whether your domain shares coding's three enabling properties: is the work well-documented, verifiable, and simulatable before it matters. George's framework suggests that tasks lacking these properties (most strategic, judgment-heavy, or relationship-dependent knowledge work) will see AI diffusion happen more slowly than coding did, so a product roadmap assuming coding-speed adoption elsewhere may be overly optimistic.

Questions to Consider

• If you are assessing an AI product's traction, are you looking at engagement and usage depth, the signal David George says flipped Harvey from a mediocre-adoption product to a client-demanded one, or only at which customers signed up?
• Does the AI feature or product you are building share coding's three properties (well-documented, verifiable, simulatable) that George argues made coding an unusually fast case for AI diffusion, or does it more closely resemble the slower-to-diffuse categories of knowledge work?
• If your organization's AI spend per employee is closer to the median (roughly $12 per month in Aram Verdiyan's data) than the top 1% of companies (roughly $7,000 per employee per month), what would meaningfully higher, more intentional AI investment actually unlock for your team?
• For any fast-growing early revenue number your team or a company you're evaluating is reporting, has it survived a renewal cycle yet, or is it still an unverified growth claim from the first few months of selling?

Bottom Line

David George and Aram Verdiyan argue that AI has made capital itself a direct product-quality lever for the first time, which is reshaping venture's power law to be more extreme than in prior cycles. But the same distinction that matters to venture investors, telling real, diffusing usage apart from early hype, is exactly what any product leader needs to judge whether their own AI feature or product has genuine traction, using signals like sustained engagement and client-driven demand rather than headline wins.

Case Studies Mentioned

Harvey's usage inflection after reasoning models arrived

Harvey, an AI product for lawyers, signed high-profile law firm customers early on and executed a smart go-to-market combining legal expertise with AI research capability. Despite the strong early signings, actual product usage looked mediocre compared to other AI software companies at the time, which George attributes partly to lawyer caution around AI hallucinations. Once reasoning models improved, usage and engagement visibly took off, and the market signal flipped from law firms being nervous about the tool to their own clients actively demanding the firms use it. George treats this as the clearest example of the gap between early headline traction and real, durable product-market fit.

Cursor's acquisition despite persistent "it's dead" narrative

Cursor was acquired by SpaceX in a deal reported in this conversation as roughly $300 million in annual recurring revenue with a large accompanying valuation, yet Verdiyan notes that people were publicly claiming "Cursor is dead" even on the morning the acquisition was announced. The example illustrates how disconnected public sentiment about a product can be from its actual underlying business performance, reinforcing the episode's broader point about verifying real usage and revenue rather than trusting prevailing narrative.

Intercom's founder-led AI-native rebuild

Intercom brought its founder back to lead a full revamp of the business into an AI-native product, described in the conversation as an example of a company willing to essentially disrupt its own existing revenue base to rebuild around AI, something Verdiyan says is extremely difficult for private-equity-owned companies to pull off because it requires board and investor alignment to intentionally sacrifice near-term performance for a long-term rebuild.

People to Follow

David George

General partner at Andreessen Horowitz, focused on growth-stage investing. George argues that AI companies benefit from a direct capital-to-product-quality feedback loop that didn't meaningfully exist in prior technology cycles, and that the firm's strength in late-stage growth investing depends on relationships built through its early-stage business.

Aram Verdiyan

An investor at Accolade Partners and a longtime limited partner in Andreessen Horowitz funds, previously an employee at the firm a decade earlier. Verdiyan brings a data-driven, allocator's perspective on venture fund performance, citing original research on consistency of returns across thousands of US venture firms.

Notable Quotes

"We've looked at the data of 3,000 venture capital firms in the U.S. Only 20 have achieved consistent 3x net returns over the last two decades." (Aram Verdiyan)

"Just because you put Sears on a website didn't make it Amazon. You have to have the benefit of building Amazon from the studs logistically to make it Amazon." (Aram Verdiyan)