Top VCs on How AI Is Rewriting Investment Logic: The 'Middle Trap' and Flywheel Effect

Several partners at a16z believe AI is upending traditional investment logic: capital is directly converted into computing power, and once the flywheel of the strong getting stronger starts, it is hard to reverse. Among 3,000 VC firms, only 20 have achieved sustained 3x returns, and the "middle" is dying. AI revenue reached $100 billion in four years; SaaS took 15 years. Traditional SaaS asset valuations have been halved, and companies that cannot become AI-native are being eliminated. The real bottleneck is not demand, but energy and data center supply.

AI is pushing the "power law distribution" of tech investing to its extreme: for the first time, capital can be directly converted into computing power and strengthen a company's absolute moat, while traditional SaaS assets that cannot become AI-native face the severe risk of being left behind by the times.

Silicon Valley's top venture capital firm Andreessen Horowitz (a16z) released an in-depth interview podcast on September 10. a16z partner Jennifer Li, growth fund partner David George, and long-time LP and Accolade Partners partner Aram Verdiyan engaged in a rare deep conversation about investment logic in the AI era. The three's core judgments are highly consistent: AI is turning the "power law effect" from a characteristic of the venture capital industry into a systemic rule of the entire capital market; capital allocation frameworks, fund structures, and even the boundaries of asset classes are being rewritten.

Power law refers to a specific functional relationship between two variables: the change in one variable is a multiple of some power of the other variable. Its core characteristic is that "a very small number of individuals occupy the vast majority of resources or share," which is what people often call the "long tail effect" or the "80/20 rule."

Several core data points in the conversation are sobering: among 3,000 VC firms in the United States, only 20 have achieved sustained 3x net returns over the past two decades, accounting for less than 1%; AI revenue reached $100 billion in just four years, while SaaS took 15 years to cover the same distance; currently, the average US enterprise spends only $12 per employee per month on AI, while the top 1% of companies spend as much as $7,000—meaning AI diffusion is still in its very early stages.

"Capital Equals Compute": The "Extreme Flywheel" Changes Industry Moats

David George made one of the most disruptive judgments in this AI wave during the interview:

"Now, you can throw capital at a company, and it will directly compound and amplify that company's advantage. This is the first time in my career."

He explained that in the traditional software era, pouring too much money into a startup was a disaster—hiring a thousand people, creating coordination chaos, and forming overlapping priorities would cause the company to "mess up." But the AI era is different: "You can throw money into computing power, and computing power can directly improve products and business."

This means economies of scale are real in the AI market and will continue to strengthen. The three frontier model companies—SpaceX, OpenAI, and Anthropic—have a combined potential enterprise value of $3.5 trillion to $5 trillion. Jennifer Li pointed out that it is shocking that before SpaceX's IPO, most institutional investors had almost no exposure to it.

Aram Verdiyan gave his judgment from a more macro perspective:

"AI is attacking every corner of GDP—transportation, labor, services, capital, coordination. Historically, no technological paradigm has ever hit $30 trillion of GDP at the same time."

He also put forward a disruptive view on AI's TAM (total addressable market): the healthcare IT industry spends about $60 billion to $100 billion annually, but AI is actually attacking the healthcare workforce itself—billing management, administrative processes—which is a trillion-dollar industry. "AI's TAM could be more than 10 times that of traditional SaaS or healthcare IT."

David George cited labor data to further support this: the ratio of labor spending to software spending in the US economy is about 40 to 1. "This doesn't mean labor will disappear, but labor will be reinvented. Equating AI with 'next-generation software' underestimates it too much."

"Death of the Middle": Structural Reshaping of the VC Landscape

A phrase repeatedly appeared in the interview: "death of the middle."

Aram Verdiyan shared a core data point: among 3,000 venture capital firms in the United States, only 20 have achieved sustained 3x net returns over the past two decades, accounting for less than 1%. He particularly emphasized that "sustained" is the key word—you don't need seven or eight funds, but over a 20-year cycle, do you have 3 to 4 funds that achieved 3x net TVPI? Only 20 firms nationwide meet this condition.

In contrast, Cambridge data shows that over the past decade, the average return of venture capital has been only 1 to 2 times net return—not only worse than private equity, but even worse than public markets, and with capital locked up for ten years.

Against this backdrop, the VC industry is polarizing: on one end are truly "small and refined" firms—focused on niche verticals or entering at earlier stages than large institutions, complementing them; on the other end are full-lifecycle large institutions—covering from seed to IPO, with 700 employees and strong operational resources.

Jennifer Li clearly expressed a16z's flywheel logic:

"Our business is a flywheel. It starts with whether we are domain experts and can impress founders; if we win deals and help companies achieve better outcomes, a brand effect forms, new founders want to be with winners, and winning founders endorse other founders—that's how the flywheel works."

David George added: the success of growth funds fundamentally depends on the early-stage business—"I always say, our business starts with early stage and ends with early stage."

For institutions that want to participate in AI by "only doing late stage" or "skipping early stage," Aram Verdiyan was direct: "The strongest late-stage franchises all have a huge early-stage franchise behind them. If you are a brand-new late-stage fund and want to write a $500 million check, it's really hard." David George agreed: "I've been in that world, and it's indeed very hard."

How to Identify Real AI Traction? "The Word ARR Is Being Redefined"

As AI bubbles and real demand coexist, distinguishing real traction has become one of the thorniest problems.

Aram Verdiyan described a typical scenario: "A company comes out of an accelerator and claims ARR went from zero to $5 million in a month. But there is no renewal cycle in between; they are selling to each other and multiplying monthly numbers by 12 to annualize ARR." He said sharply: "Out of every nine such companies, maybe one has real ARR of a few million and is the next Cursor."

Jennifer Li laughed: "I love it when people help 'redefine the meaning of ARR,' like a very helpful announcement."

David George gave his judgment framework:

"Whether the market is proactively demanding your product is always the core question. Financial analysis cannot tell you the answer; you must deeply understand customers and talk to them."

He used Harvey (an AI legal tool) as an example: early on, Harvey won some contracts at top law firms, but actual usage was mediocre, and retention data was not outstanding compared to other AI software companies. But after reasoning models landed, it completely flipped—the value lawyers derived from the product increased significantly, and usage and engagement soared; more critically, customers shifted from worrying about hallucination risks to proactively demanding that law firms use the product. "That's the market signal we look for."

The Plight of Traditional SaaS and the "AI Trap" of Private Equity

Compared with traditional software-as-a-service (SaaS), AI's penetration speed and potential total addressable market (TAM) show exponential explosion.

"Today, AI revenue has reached $100 billion. SaaS took 15 years to reach the same level, while AI took only 4 years." Aram Verdiyan gave a highly impactful data point. He pointed out that AI is attacking every layer of GDP—transportation, labor, services, and capital coordination—"never before has a technological paradigm been able to touch $30 trillion of GDP simultaneously."

However, AI's rapid advance has left old-era software assets and the private equity (PE) funds heavily invested in them on edge. Between 2021 and 2022, a large number of traditional software companies were acquired by PE through leveraged buyouts (LBOs) at high valuations of 25 to 32 times EBITDA, but these assets lacking "AI-native" capabilities now face valuation shrinkage and liquidity drying up.

Aram Verdiyan pointed out bluntly: "Software assets you acquired before 2021 at 25 to 32 times EBITDA may be worth only half today." He cited data: between 2021 and 2022, leveraged buyout software deals totaled about $200 billion to $300 billion, taking on more than $200 billion in debt; today those companies' valuations have contracted sharply, and leverage ratios have soared. The redemption wave in the private credit market is a direct reflection of this logic.

More severe is the reversal of growth logic: Aram Verdiyan pointed out, "In public markets, every 1 percentage point increase in growth rate is equivalent to a 3 percentage point increase in EBITDA." This means companies that cannot prove AI-accelerated growth will suffer brutal valuation discounts—"many PE-held targets have neither fast enough growth nor management teams capable of reshaping the business."

Aram stated bluntly:

"Look at today's public market, that asset (traditional SaaS) is now trading at 2 times price-to-sales. This is not just a valuation issue; more seriously, these companies may not find buyers at all. Because if you look at a software company today, the first thing you think is: what is its ultimate value? Can it withstand the impact of AI?"

Jennifer Li warned against the illusion of "AI-version private equity":

"Just putting Sears on a website won't turn it into Amazon. We see some companies whose first move is to launch AI customer service, but if you don't rebuild from the bottom of the workflow, customer churn actually accelerates. Every drop in NPS directly corresponds to a drop in revenue, and with debt on top, it spirals downward."

For attempts to save old businesses by simply introducing AI, Aram also poured cold water: "You can't just send an operating partner into a company and say 'let's add some AI on top,' that simply doesn't work." DG added that even simple AI customer service replacement, without deep workflow restructuring, will lead to a decline in Net Promoter Score (NPS) due to customer churn and trigger a downward spiral in revenue.

AI's Real Bottleneck: Not Demand, but Supply

At the end of the interview, the topic turned to an underestimated angle—where is AI's real bottleneck?

Aram Verdiyan asserted: "AI's bottleneck today is not on the demand side, but on the supply side." The supply chain from left to right is: energy → power grid → data centers → chips → frontier models → applications. The US has a strong advantage in chips and the chain to the right, and the VC ecosystem is also strongly supportive. But at the energy and grid infrastructure level, the problem is not insufficient power generation, but "speed to power"—permitting, transmission, regulation. Other countries add 10 times more renewable energy capacity annually than the US.

"This means data centers need to be completely reimagined. The density of old data centers can no longer support new AI facilities. What exists here is not a $5 billion or $10 billion opportunity, but a hundred-billion-dollar-level opportunity."

David George listed the areas he believes are most likely to create enormous value in the next 10 years: robotics, autonomous driving (fewer than 10,000 Waymo vehicles on US roads), healthcare (18% of GDP, AI has barely touched it), energy, manufacturing, and defense.

"It feels like we've done a lot in AI now. But looking back in ten years, those other major areas are just beginning to create value. The next SpaceX or OpenAI will most likely be born in these areas."