The Biggest Revolution of His Life
Marc Andreessen doesn't mince words: AI is the biggest technological revolution of his life. Not the internet. Not mobile. The comparisons he reaches for are the microprocessor, the steam engine, electricity—and the wheel.
The backstory matters: In the 1930s, before computers existed, scientists debated whether to build machines in the image of calculators or in the image of the human brain. The first neural network paper was published in 1943. But the industry took the calculator path—IBM, Intel, the entire 80-year computer industry. The neural network path became a backwater, decade after decade of "excessive optimism followed by disappointment."
That was Christmas 2022. Less than three years ago. And now the alternate path—the human cognition path—is delivering on 80 years of promise.
Why AI Spreads Faster Than Any Technology in History
The internet took decades to build out: fiber in the ground, cell towers everywhere, billions of smartphones shipped. The original iPhone in 2007 didn't even have broadband—it ran on 2G. Mobile broadband didn't really arrive until 2010.
But that infrastructure now carries 5 billion people. And AI can deploy to all of them as fast as they want to adopt. You couldn't download electricity. You couldn't download indoor plumbing. But you can download AI.
The result: AI consumer products are growing revenue at rates Andreessen has never seen. And they're monetizing well, including at higher price points. The new AI companies are more creative on pricing than SaaS ever was—$200-300/month tiers are becoming routine for consumer AI.
The Price of AI Is Collapsing
Per-unit AI costs are falling faster than Moore's Law. Every input is getting cheaper. GPUs that were scarce 18 months ago are being extended to 7+ year lifespans. And the small model revolution means frontier capabilities get shrunk down to run on laptops within 6-12 months.
This hyperdeflation is driving massive demand. Tokens by the drink are going to get a lot cheaper—and that means a lot more demand.
US vs. China: The Only Race That Matters
AI is being built in two places: the United States and China. Everyone else either can't build it or won't (Europe's AI Act has "killed AI development in Europe to a large extent").
The Chinese ecosystem exploded in the last year. DeepSeek was the "supernova moment"—an open-source model from a hedge fund that surprised everyone with how good it was and how cheap it was to run. Now there are 4-6 major Chinese AI companies: DeepSeek, Qwen (Alibaba), Kimmy (Moonshot), plus Tencent, Baidu, and ByteDance.
The common understanding: China's government has instructed DeepSeek to build the next version only on Chinese chips—a forcing function to accelerate their chip ecosystem.
The Trillion-Dollar Questions (Not Answers)
Andreessen is refreshingly honest about what's still uncertain:
Open source vs. closed source? Still open. Big models keep getting better—people at the labs say progress is continuing rapidly, they have "800 new ideas." But open source keeps catching up within months. XAI caught up to OpenAI/Anthropic level in less than 12 months from a standing start.
Incumbents vs. startups? The "GPT wrapper" critique is wrong. The best AI application companies (like Cursor) end up building their own models, using dozens of different models, backward integrating into serious technology. They're not wrappers—they're full-fledged deep tech companies.
Why VC Has an Advantage
When a company faces open strategic questions, it's often a big problem—they need to pick one strategy and be right. But venture capital can bet on multiple strategies at once.
a16z is "aggressively investing behind every strategy that we've identified that has a plausible chance of working." Foundation model startups, application companies, both open source and closed source plays, US and international.
State Regulation: The New Threat
Federal AI regulation risk has dropped—there's little appetite in DC to do anything that would prevent beating China. But the attention has shifted to states. a16z is tracking 1,200+ AI bills across all 50 states—not just blue states, red states too.
California's SB1047 would have assigned downstream liability to open-source developers. If your model gets built into a nuclear power plant that melts down 5 years later, you're liable. "Completely insane. It would have completely killed open source, killed startups, killed academic research in its entirety." Governor Newsom vetoed it at the last minute.
Colorado passed a draconian bill and is now trying to reverse it. The EU AI Act has been so bad that even the EU is trying to unwind it.
Revealed Preferences vs. Stated Preferences
Andreessen makes a sharp observation about AI sentiment: if you poll Americans, they're in a "total panic"—AI is terrible, it's going to kill all the jobs. But if you watch their revealed preferences? They're all using AI.
That gap between what people say and what they do is the story of technology adoption. And the revenue numbers show which side is winning.