Public transcript

Marc Andreessen's 2026 Outlook: AI Timelines, US vs. China, and The Price of AI

By: a16z
Length: 1 h 21 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

In this a16z AMA, Marc Andreessen argues AI is the biggest technological revolution of his life — bigger than the internet — with new AI companies growing real revenue at an unprecedented takeoff rate. He walks through the economics of consumer and enterprise AI, the falling price of intelligence, big versus small models, and the chip competition forming around Nvidia. He then covers the US–China AI race sparked by DeepSeek, why he thinks federal rather than state regulation should govern AI, and open questions like usage-based versus value pricing and open versus closed source. He closes with thoughts on a16z's AI reorganization, his partnership with Ben Horowitz, society's revealed preferences around AI adoption, and a lightning round.

Key points

Marc Andreessen calls AI the biggest technological revolution of his life, comparable to the microprocessor, electricity and the steam engine, and bigger than the internet.
New AI companies are growing real customer revenue at an unprecedented takeoff rate, though he is skeptical today's product forms will look the same in 5 or 10 years.
The price of AI is falling much faster than Moore's law, and he expects tokens-by-the-drink to get much cheaper, driving enormous demand.
He expects a pyramid structure: a few giant 'god models' plus a cascade of smaller models, with small models matching big model capabilities within 6 to 12 months.
DeepSeek's release was the year's 'supernova moment,' and China now has roughly three to six serious AI companies, making AI a two-horse US vs China race.
He sees roughly 1,200 state-level AI bills across the 50 states as a serious problem, and argues the federal government should assert itself as the regulator of AI.
He calls usage-based vs value-based pricing a trillion-dollar question, and says pricing by value rather than cost is the goal, with high prices often good for customers.
Open source vs closed source and incumbents vs startups remain open questions; venture's advantage is betting on multiple contradictory strategies at once.

Questions it answers

01:22The AI revolution's scale

How big is the AI revolution and what inning are we in?

He calls it the biggest technological revolution of his life, bigger than the internet, comparable to the microprocessor and electricity. He traces an 80-year path from the 1930s neural network theory through the ChatGPT moment, and says new AI companies are growing real revenue at an unprecedented takeoff rate.

AI is compared to the microprocessor, electricity and the steam engine.
New AI companies are converting real demand into revenue faster than anything he has seen.
He is skeptical today's product forms will persist in 5 or 10 years.

09:04Business models and falling prices

Why isn't he worried about AI expenses keeping pace with revenue?

The internet is already built, with 5 billion people on mobile broadband, so consumer AI can proliferate at light speed. Enterprise value comes from injecting intelligence into business processes, and the core model is tokens by the drink, with the price of AI falling much faster than Moore's law.

Consumer AI tiers of $200–$300 per month are becoming routine.
Per-unit AI costs are collapsing, driving elastic demand growth.
Shortages of GPUs and data centers will become gluts as trillions go into the ground.

16:12Big models vs small models

Will big models or small models dominate?

He expects a pyramid like the computer industry: a few supercomputer-scale "god models" plus a cascade of smaller models down to embedded chips. Small models chase big model capabilities within 6–12 months, as with the Chinese model Kimmy reportedly replicating GPT5-level reasoning on one or two MacBooks.

Many economic tasks don't require Einstein — a capable smaller model suffices.
The smartest models stay at the top, but volume goes to smaller models.
Nvidia's profits are a bat signal drawing AMD, hyperscalers, startups and China into AI chips.

24:09China and the two-horse race

How worried should the US be about Chinese open source AI?

AI is essentially only being built in the US and China, making it a two-horse race. DeepSeek's open-source release was the year's supernova moment, and China now has three to six serious AI companies including DeepSeek, Alibaba's Qwen, Moonshot's Kimmy, Tencent and ByteDance.

DeepSeek came from a hedge fund, not a national champion, and was reportedly a surprise even to the Chinese government.
Cynics in DC say China is dumping open source to commoditize the field.
Awareness of the two-horse race has improved the US policy landscape.

33:39State and federal regulation

What is the state of AI regulation in the US and Europe?

He is now much less worried about ruinous federal legislation, but tracks roughly 1,200 state-level AI bills across red and blue states. He argues the federal government should assert itself since AI is interstate by definition, and points to the EU AI Act as a cautionary tale that has killed European AI development.

California's SB1047, modeled on the EU AI Act, would have assigned downstream liability to open source developers; the governor vetoed it.
Colorado passed a draconian bill and is now trying to reverse out of it.
The administration and most of Congress support federal rather than state regulation.

42:01Pricing AI

Is usage-based pricing the right way to price AI?

Usage-based pricing is a trillion-dollar question. Cloud providers proliferating AI by the drink is great for startups with no fixed costs, but the optimal model for applications is pricing by value — a percentage of business value or marginal productivity, not by cost.

High prices are often good for customers because they fund better products.
Startups are experimenting with pricing as a percentage of what would have been a person's job.
a16z has dedicated pricing experts working with its companies.

47:14Open questions: open source and incumbents

Will open source or closed source win, and incumbents or startups?

Both are still open questions. Big labs say progress continues rapidly, open source models keep shrinking frontier capabilities, and open source also spreads the knowledge of how to build AI. On incumbents, big tech plays hard, new incumbents like OpenAI, Anthropic and xAI emerged fast, and application companies like Cursor build their own models rather than remaining "GPT wrappers."

Once someone proves a capability, others catch up even with far fewer resources — xAI caught up in under 12 months.
Leading application companies use dozens of models and backward-integrate into building their own.
Venture's advantage is betting on multiple contradictory strategies at once.

58:32A16Z and Ben Horowitz

What do Marc and Ben disagree on, and how has the a16z AI reorg gone?

They debate everything but rarely disagree-and-commit; the biggest ongoing discussion is the firm's public footprint and how many controversial topics to touch. He defends the AI reorg and American Dynamo, says crypto is back, and expects intersections between AI, crypto, biotech and energy.

Being outspoken and controversial has proven a competitive advantage with founders.
Venture firms must jump on each new technology wave or peter out.
He doesn't feel the firm is missing a vertical right now.

Notes

The scale of the AI revolution

Marc Andreessen frames AI as the biggest technological revolution of his life, bigger than the internet, comparable to the microprocessor, electricity and the steam engine.
He traces the history back to the 1930s debate between adding-machine-style computers and neural networks, noting the first neural network paper appeared in 1943.
Roughly three years after the ChatGPT moment (Christmas 2022), he sees daily surprises in both research and products, and expects the current product forms to look very different in 5–10 years.

Economics and pricing

New AI companies are converting real demand into bank-account revenue at an unprecedented takeoff rate.
The price of AI is falling much faster than Moore's law, driving elastic demand growth; tokens by the drink will get much cheaper.
Consumer AI products can deploy instantly to the roughly 5 billion people already on mobile broadband internet — "you can download AI."
Consumer tiers of $200–$300 per month are becoming routine, and he praises AI companies for being more creative on pricing than SaaS companies.
Pricing by value rather than cost is the goal; high prices can benefit customers by funding better products.

Models and chips

He expects a pyramid: a few "god models" in giant data centers, cascading down to small models on embedded chips.
Small models catch big-model capability within 6–12 months; the Chinese model Kimmy reportedly replicates GPT5-level reasoning and runs on one or two MacBooks.
Nvidia's profits are a "bat signal" drawing AMD, hyperscalers, startups, Korea, Japan and China into AI chips; shortages become gluts.

China and policy

DeepSeek's open-source release was the "supernova moment" of the year; China has three to six serious AI companies (DeepSeek, Qwen/Alibaba, Kimmy/Moonshot, Tencent, ByteDance).
He tracks ~1,200 state AI bills and argues the federal government must assert regulation of AI as interstate commerce; the EU AI Act has killed European AI development.
California's SB1047 would have assigned downstream liability to open source developers and would have killed open source, startups and academic research; the governor vetoed it.

Open questions and a16z

Open vs closed source and incumbents vs startups remain "trillion dollar questions"; venture can bet on multiple contradictory strategies at once.
He and Ben Horowitz debate mainly the firm's public footprint and how many third-rail topics to touch.
Revealed preferences show people love and use AI even while surveys show panic; he expects the usual pattern — proliferation, freakout, then gratitude.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.