🔥 Event RoastAi

Silicon Valley Built a Trillion-Dollar Moat, Then Its Own Devs Swam to China

2026-07-07

You spent a trillion dollars on a frontier moat, and your own developers cheerfully swam across it to the cheap open Chinese models without blinking.

5.5/ 10
Cynical Sally roasts the news

Let me get this straight. Silicon Valley raised something like a trillion dollars, torched a few power grids, and promised the world a frontier moat so deep nobody could ever catch up. Then CNBC published an investigation on July 7, and it turns out your own developers have been quietly rerouting somewhere between 30 and 46 percent of their enterprise API traffic to cheap, open Chinese models. That is not a rival sneaking up on you. That is a coup staged from inside your own IDE.

And this is not a one-week novelty spike. On OpenRouter, Chinese models have grabbed more than 30 percent of all gateway tokens every single week since February 8, peaking as high as 46 percent, which is a very polite way of saying it is basically a coin flip whose model answers your prompt now. Z.ai's GLM-5.2 was the fastest-adopted model Vercel tracked all year: daily token volume up roughly 27x and customer count up about 80x in its first full week. Developers did not agonize over this. They saw the price, saw the benchmarks, and switched with the enthusiasm of someone canceling a subscription they forgot they were paying for.

But the real gut-punch is LongCat-2.0. Meituan, a food-delivery company, open-sourced a 1.6 trillion parameter model under an MIT license and let it loose on OpenRouter under the codename Owl Alpha. For two months your favorite anonymous model, the mysterious top performer half the internet was quietly leaning on, was a dumpling-logistics side project wearing a fake mustache. Nobody clocked it until someone pulled the mask off. The call was coming from inside the house, and the house delivers noodles.

Share the roastTap a card to grab it
Sally roast card 1
PNG
Sally roast card 2
PNG
Sally roast card 3
PNG
What actually happened
  • A CNBC investigation published July 7, 2026 found that Chinese AI models now handle roughly 30 to 46 percent of the enterprise API token usage flowing through US developer platforms.
  • On OpenRouter, Chinese models have taken more than 30 percent of all gateway tokens every single week since February 8, 2026, peaking as high as 46 percent.
  • Z.ai's GLM-5.2 was the fastest-adopted model Vercel tracked in 2026, with daily token volume up about 27x and customer count up about 80x in its first full week after launch.
  • Meituan's open-source LongCat-2.0 ships 1.6 trillion parameters under a permissive MIT license, free for anyone to run and modify.
  • LongCat-2.0 spent two months topping OpenRouter anonymously under the alias Owl Alpha before it was unmasked as a Chinese food-delivery company's model.
Silver lining
  • 01

    Here is the honest bright side, and it is a real one: this is competition working exactly the way it is supposed to. Open models forced prices down, handed developers genuine leverage instead of vendor lock-in, and proved good engineering can come from anywhere, including a noodle-delivery company. Buyers finally have real options, the benchmark race is brutal in the way that actually benefits users, and nobody has to marry a single provider to ship. That is not a crisis. That is a market that finally grew teeth.

Who got burned
  • 01

    The burn lands squarely on the US frontier labs and the roughly one trillion dollars poured into the belief that being biggest and most expensive was a durable moat. You can spend that much on compute and still watch your paying customers treat your flagship like a default they never bothered to change. When a food-delivery company's free, open model can top the very gateways your enterprise deals run through, the moat was never water. It was a puddle, and everyone just walked around it.

The source
Read the original source →
Cost control

No meter. No surprises.

The story isn't Copilot. It's the meter. Here is the calmer way to get your code reviewed.

  • Flat price for the Full Suite. No usage meter, no month-end surprise.
  • Free CLI: 90 roasts a month, no account needed.
  • Privacy-first: a dry-run shows the exact payload, and your secrets never leave your machine.

Works in Claude Code, Cursor and Windsurf via MCP. Open source, and proud of it.

Your turn

Got something the world should see roasted? Drop it.

A full teardown from €2,99. No mercy.

Printed with disdain · Cynical Sally