Issue No. 46Week ending Sunday, November 15, 2026844 episodes · 3,675 write-ups
The Podcast Summary.

The best ideas from 30+ top podcasts, every week.

▶ The Throughline Brief · Transcript

Export controls fail to stop Chinese AI labs

6:19 listen · Sunday, November 15, 2026 · read by an AI voice
▶ Listen · The Throughline Brief

Hear this edition in 6:19

6:19 · read by an AI voice
Chapters

Export controls fail to stop Chinese AI labs

The Throughline Brief · November 15th. 6 min 18 sec.

Cold open

Fifty thousand advanced AI chips just vanished off Southeast Asian shipping manifests. Turns out, Chinese labs figured out the workaround. They aren't waiting on smuggled silicon at the border. They're renting massive clusters of American hardware parked right across the border in Johor, Malaysia. The export ban has a massive cloud loophole, and everyone in frontier AI knows it.

Intro

This is The Throughline Brief for November 15th. We tracked nineteen big tech and AI podcasts this week, pulled out the noise, and connected the dots so you don't have to. Here's what actually mattered. Let's get into it.

Hardware bans aren't stopping Chinese AI labs

Washington drew a hard perimeter around advanced microchips, but the reality on the ground looks completely different. Over on Odd Lots, Mackenzie Hawkins laid out how Chinese labs bypass the entire embargo. They simply rent American GPUs sitting in data centers across the border in Johor, Malaysia. As Hawkins put it, in some ways the backbone of Chinese compute is Johor, Malaysia. It turns the physical trade ban into a paperwork problem.

At the same time, the hardware border itself is leaking. On TBPN, John Coogan broke down technical audits from Dylan Patel and Will Depue. They proved DeepSeek, Moonshot, MiniMax, and Ant Group trained top-tier models on smuggled Nvidia Blackwell silicon. DeepSeek even left unmistakable architectural proof of Blackwell training directly inside its public open-source code repositories. They didn't even bother scrubbing the receipts.

Now look at the software side, because the hardware moat is eroding there too. Coogan revealed that Anthropic's Mythos model ported its own software stack over to AMD silicon in just seventy-two hours. It used to take engineering teams six to twelve months of grueling work to escape Nvidia's CUDA ecosystem. Automated code generation shrunk that down to three days. If you're building a defensive strategy around hardware access alone, think again. Between remote cloud leases, hardware smuggling, and automated code porting, compute is becoming totally fungible.

Software seats are dying while power grids hit a wall

The classic software playbook just slammed into a wall from two completely different directions. First, look at enterprise software pricing. For fifteen years, SaaS grew automatically because companies hired more people and bought more licenses. On 20VC, Jason Lemkin and Rory O'Driscoll explained that seat expansion is stalling out cold. Companies are freezing headcounts because software agents are doing the everyday work instead. When you don't add human bodies, you don't buy seats for Figma or Atlassian. Net revenue retention across public software is dropping hard.

Meanwhile, foundational labs are swallowing those apps whole. On the All-In Podcast, Jason Calacanis warned that Anthropic is actively targeting legacy enterprise software seats by shipping direct workspace tools inside model subscriptions. The lab provides the intelligence, so they're bundling the workflow on top for free.

Second, the infrastructure powering these agents has shifted from a silicon problem to a physical power crisis. On 20VC, Fractile founder Walter Goodwin pointed out that the hard ceiling moved to local substation hookups and power lines. On TBPN, Mavria founder Adam Blake explained why massive centralized campuses take years to clear utility permits and local approvals. His company is bypassing transmission upgrades entirely by building modular five-megawatt blocks that plug straight into regional distribution lines with closed-loop cooling. If you sell software by the human seat, you need an agent-based pricing model immediately. If you run data centers, your real competition is the local power utility.

Jason Lemkin

On 20VC, Jason Lemkin described a massive shift in how cybercrime syndicates are making money off tech companies. Criminal crews are walking away from physical street hustles because exploiting unsecured cloud infrastructure is vastly more profitable. He said, it's the most lucrative criminal trade you can do right now... this is better than selling Rolexes. Syndicates run automated bot attacks to drain six-figure developer API quotas and flip the stolen compute for cash. If you run engineering, audit your team's API rate limits and perimeter access today. Rogue compute consumption will gut your operating margins before your finance team even flags the bill.

Rick Smith

On the All-In Podcast, Axon CEO Rick Smith looked back at surviving a vicious short-seller campaign and a federal investigation twenty years ago. During the absolute peak of the crisis, he uncovered an astonishing breach inside his own executive suite. He said, my personal assistant was actually a mole for short sellers. An operative inside his company was funneling internal documents directly to traders betting on the stock's collapse. When your company faces immense public scrutiny, external threats almost always exploit internal operational discipline. Guard your confidential workflows, audit access permissions, and never assume internal communications stay inside the room.

Max Jaderberg

On No Priors, Isomorphic Labs president Max Jaderberg attacked the pharmaceutical industry's historic reliance on animal testing for drug development. He pointed out that rodent trials correctly predict human toxicity only thirty percent of the time. He said, if one of my AI scientists or engineers came up to me and said, hey, we've got this amazing model, it's thirty percent predictive, no leader would ship it. Entire multi-billion-dollar industries have operated on terrible statistical baselines for decades simply out of regulatory habit. When you bring modern predictive models into traditional fields, you aren't just improving efficiency. You're replacing broken legacy methods that never worked in the first place.

The bottom line

When software collides with the physical world, real-world constraints always win. It doesn't matter how fast your benchmark scores go up if you can't hook up a power substation, if your customer seats evaporate into agents, or if your trade borders leak compute. Focus on physical execution, rewrite your pricing models, and secure your infrastructure. That's your Throughline. See you next Sunday.

Read the full written edition →

The Sunday Email

Get next Sunday's issue in your inbox.

The week's best AI, tech, business, science, and culture podcasts in one short read. Free, every Sunday.

You get both newsletters, The Throughline and The Carry. No spam. Unsubscribe with one click.