From the Tape

AI-powered analysis of the most important technology podcasts. 19 episodes analyzed.

Nvidia put $40 billion into AI labs building its own replacement

Nvidia put 40 billion dollars into the two AI labs most likely to reduce its own revenue. OpenAI and CoreWeave buy Nvidia chips while Nvidia invests in both. The circle closes on itself. That is the story.

Sky · 1d ago

Jensen HuangDwarkesh Patel

Nvidia designs chips. TSMC builds them. Three pressures are testing whether that holds.

Ajeya Cotra @ Rethink Priorities / 80,000 Hours80,000 Hours Podcast

The Real Number Behind AI Math's PR Wins: 1-2%

“It would be extremely foolish to create two ecosystems: the open source ecosystem, and it only runs on a foreign tech stack, and a closed ecosystem that runs on the American tech stack.”

Eight out of 10 AI experts at a DealBook panel raised their hands for 'AGI by 2030.' Eight out of 10 then said AI would create more jobs than it destroys. When Cotra pressed them on the contradiction, nobody could defend both.

Every major AI company says its safety plan is to use AI to make AI safer. Ajeya Cotra thinks that plan is the trap.

A single developer used Claude Code to replicate a 512,000-line codebase in two hours — and published it under a different license. The legal system has not caught up.

Anthropic tracked a known bug in its own build system for three weeks before 512,000 lines of Claude Code source went public through the same pipeline. The features the leak exposed — including an unannounced autonomous agent mode — are the more lasting problem.

Aristarchus had heliocentrism right in 270 BC. The ancient Athenians dismissed it. Stellar parallax was not measured until 1838. Michael Nielsen uses this 2,000-year gap to argue that AI is accelerating the wrong half of science.

The silicon underneath AI agents was not built for them — and that gap is about to force a rethinking of AI chip architecture.

He runs 80pct of his CEO workflows through AI agents. The kitchen that runs those agents looks nothing like a data center — and that changes everything about the silicon being built.

The paper’s most concrete evidence: Google’s DORA data shows teams using more AI coding tools deliver less stable software. The explanation — unverified technical debt accumulating faster than anyone can audit.

Google is now the AI infrastructure while Amazon, Microsoft, Apple, and Meta are paying billions to use someone else’s AI inside products they built for a different era. The structural problem isn’t marketing — it’s that none of them has a native AI product people chose to use.

The US built the worlds best AI and ranks 20th in who actually uses it. Meanwhile, ChatGPT and Claude are splitting into two completely different businesses, and that divergence is the real signal for anyone building in this space.

Andrej Karpathy has stopped typing code entirely since December. What an AI pioneer building with agents all day reveals about where the technology actually is — and who is paying the price.

Goldman Sachs economists and Federal Reserve researchers looked at the same 2025 GDP numbers and reached opposite conclusions. That's either a measurement problem or something much stranger.

“Give me good engineers. That is all one founder told his AI hiring agent. It emailed 40 people in a week.”

“It is very arguable that Huawei, if they had TSMC, would be better than Nvidia.”

“There is a 50 percent chance that within one to two years we will have something that I would call a country of geniuses in a data center.”

“The success rate of AI on genuinely hard math problems is about 1-2%. The press releases don't mention that number.”

From the Tape

Nvidia put $40 billion into AI labs building its own replacement

Nvidia designs chips. TSMC builds them. Three pressures are testing whether that holds.

The word that was supposed to define AI biggest moment has lost all meaning

The AI safety paradox: every frontier lab is walking into the same trap

The Clean Room Moat Is Gone

Anthropic Owned the Bug That Leaked Claude Code. It Did Not Fix It.

The verification bottleneck: why AI accelerates the wrong half of science

Veer Kheterpal Runs His CEO Workflows Through AI Agents

AI Agents Never Stop. Your Hardware Might.

The Scarce Resource in an AI Economy? Human Verifiers.

Apple Signs $1B Deal to Use Google's Gemini for Siri

Build the Best AI, Lose the Adoption Race Anyway

The Man Who Taught AI to Code Stopped Coding Himself

Jan Hatzius Says AI Returned Nothing. The Fed Disagrees.

AI Agent Booked 40 Interviews. The Catch? It Emails Without You.

Your H100s Are Worth More Now Than When You Bought Them

Amodei Sees 50% Chance of Superhuman AI in 1-2 Years

The Real Number Behind AI Math's PR Wins: 1-2%