For two years, one idea held the entire AI trade together.

Chinese chips could run AI models. But building them was a different beast. And that beast had a name: Nvidia.

Running a model is called inference—the light lifting. Building one is called training—the heavy lifting. Washington drew its whole export-control strategy around that split. 

Choke off Nvidia's best chips, keep China dependent on American silicon for the hard part, and you buy the West a multi-year head start.

BUT… 

That idea could now have a problem.

The (Long) Cat's Out of the Bag

Meituan—think China's DoorDash, GrubHub, and Yelp fused into one—released a model called LongCat-2.0. 

On its own, a food-delivery giant shipping an AI model is worth a shrug. The interesting part is how it was built.

Meituan says it trained the model from scratch on a cluster of 50,000 chips. Every one of them domestic. No Nvidia.

It gets more specific. 

The model chewed through 35 trillion tokens of training data—and it did so without a single crash or meltdown along the way. 

In AI training, that kind of clean run is the hard part. Anyone can buy chips. Getting tens of thousands of them to work together for weeks without falling over is where most projects die.

Meituan didn't name its chip supplier. But the software fingerprints all point to Huawei's Ascend chips—the same line Washington has spent years trying to strangle.

And LongCat wasn't alone:

• A research group including Huawei finished training work on DeepSeek-V4-Pro, a 1.6-trillion-parameter model, on more than 1,000 Ascend chips—1,500 training rounds, zero interruptions.
  
  
• Huawei's own Pangu models were trained from scratch on thousands of homegrown chips.
  
  
• Baidu says a version of its flagship model was trained on chips from its in-house Kunlunxin unit, reportedly at a 97% efficiency rate.

Read all of them together and you see the shape of something forming. 

Domestic Chinese chips have moved into the part of the pipeline everyone said they wouldn't reach—and the pace is picking up.

Another DeepSeek Moment

The entire logic of the chip bans rests on a bottleneck. Training eats compute. So starve China of the fastest chips, and the compute-hungry job of building frontier models stays out of reach.

Two cracks are now showing.

First, the H20—the most powerful Nvidia chip the US actually allows into China—is already getting lapped. 

Morgan Stanley data cited by the Financial Times shows Chinese chips beating it on raw performance and memory bandwidth. The "allowed" American chip is no longer the best chip in the room.

Second, Huawei has started selling Ascend-plus-DeepSeek systems abroad, courting buyers in the Middle East and Central Asia. 

A company doesn't export a product it can barely make work at home.

Bernstein analysts called it another "DeepSeek moment"—their shorthand for the market waking up to Chinese capability it had written off. 

One analyst put it bluntly: the idea that export controls left China's chip industry dead in the water no longer holds.

But Don’t Panic Yet 

Here's where most of the coverage overreaches—and where you get to sound measured while everyone else panics.

Crossing the training line and closing the efficiency gap are two different questions.

The first one now has real evidence behind it. Yes, China can train frontier models on its own chips. Settled.

The second is wide open. Can they do it as cheaply and as fast as a team running Nvidia's newest hardware? That's the number nobody has yet. Post-training on 1,000 chips is a milestone. 

Pre-training from scratch on 50,000 is a different beast. 

The headline says "China solved training." The data says "China showed up—now let's see the bill."

That efficiency question is the one that actually moves money. 

It sets the ceiling on how high Chinese chip companies can be valued. It decides how much of Nvidia's China revenue is truly at risk. 

And it tells you whether the export controls are a genuine speed brake—or just an expensive detour that bought a couple of years and a lot of headlines.

The Takeaway For Your Money

The clean bull case for "nobody can train without Nvidia" is gone. 

That's the update.

But the trade isn't dead… it's more interesting. 

The bet now hinges on a single unknown: efficiency. 

Watch for the metric analysts call MFU, or "model FLOP utilization.” In plain English, it measures how much useful work a chip cluster actually squeezes out versus how much it wastes. 

When Chinese labs start posting reliable efficiency numbers that rival Nvidia's, that's your signal the moat has sprung a leak. 

Until then, the honest read:

China walked through the door. Whether it can run once it's inside is the trillion-dollar question still on the table.