Latest / Minds, Bodies, and Terawatts / Chinese AI Models Now Take Up to 46% of US Companies' Token Usage as Domestic Costs Surge
Transcript
- 0:00[Host] This is Minds, Bodies, and Terawatts, where the smartest software in America was quietly built somewhere else. Sam, CNBC dropped a number this week that stopped me cold.
- 0:10[Guest] The one about who US companies are actually renting their intelligence from. Go ahead.
- 0:16[Host] CNBC reported on July 7th that Chinese AI models now handle up to 46 percent of the tokens American companies run through the routing service OpenRouter. A year ago that figure was around 11.
- 0:29[Guest] And it's not a blip. It's cleared 30 percent every single week since February. The cheap option stopped being the fallback and became the default.
- 0:39[Host] Why the stampede? These are American firms choosing Beijing's software.
- 0:43[Guest] Money, plainly. The open Chinese models run 60 to 90 percent cheaper than the top tier from OpenAI or Anthropic. When a task doesn't need genius, you route it to whatever's good enough.
- 0:55[Host] Give me a name. Who's winning this?
- 0:58[Guest] A model called GLM 5.2, from a lab called Z dot AI, out in June. Vercel says it's the fastest-growing model on its gateway since launch, daily usage up about 27-fold in the first week.
- 1:12[Host] Good enough, though. That's the safety net, right? These are the cheap seats, not the frontier.
- 1:18[Guest] That's the comforting read, and it's wrong. GLM 5.2 benchmarks fifth in the world, right behind the American labs. It's not a budget knockoff. It's near-frontier at a fraction of the price.
- 1:30[Host] So the frontier labs win the headlines, and someone else wins the invoice.
- 1:35[Guest] Exactly. OpenAI ships GPT-5.6 to the public this week, brilliant stuff. But the boring middle of the market, the millions of routine calls, is leaking overseas where the price is a rounding error.
- 1:48[Host] And Washington noticed.
- 1:49[Guest] On July 8th, the House Homeland Security Committee opened a probe into how much American business now leans on Chinese models. The worry isn't only competitiveness. It's that your customer-service bot and your code assistant run on weights you don't control.
- 2:05[Host] Here's what unsettles me. If everyone's routing to Beijing's models, doesn't that hand Beijing a lever?
- 2:11[Guest] That's the twist. Reporting this week says China is weighing whether to restrict access to its most powerful models, now that US dependence is visible. The seller just counted how many buyers it has.
- 2:25[Host] So the cheap thing everyone got hooked on can be switched off.
- 2:29[Guest] That's the shape of it. We keep score by benchmarks, who has the smartest model. But adoption gets decided by a purchasing manager staring at a bill.
- 2:40[Host] Which is a very different scoreboard.
- 2:42[Guest] It's the oldest pattern in geopolitics. When the resource that decides power shifts, and it's shifting from oil to compute, the smartest producer matters less than whoever supplies the cheap, plentiful version.
- 2:55[Host] And in that framing China isn't a laggard trailing us.
- 2:58[Guest] It's a competing architect, building the same future on a cheaper foundation. Oil taught us the pattern, the barrel doesn't care who pumped it, but the country holding the spigot cares very much.
- 3:11[Host] So the takeaway isn't that Chinese models are better. It's that they're cheaper, and cheaper is what actually spreads.
- 3:19[Guest] The best model wins the demo. The affordable one wins the world, until the supplier decides to close the tap.
- 3:26[Host] That's the shift Patrick Deglon calls the geopolitics of abundance in Unscarcity, where power flows not to whoever's smartest, but to whoever controls the supply.
- 3:36[Guest] For more on how compute becomes the chokepoint that redraws the map, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify, and this has been Minds, Bodies, and Terawatts.