Post-Launch Focus: Who Can Sell GPUs into Data Centers?

Just saw this article from TMTPost in a group chat, saying four domestic GPU companies—Moore Threads, MetaX, Biren, and Iluvatar—are starting to release financial reports. My advisor told me to try looking beyond valuations and see whose money is earned from real compute tasks. Whether chip companies are doing well shouldn't be judged just by launch event benchmarks; you have to see if labs and server rooms are willing to put them into training tasks.

In H1 2026, Moore Threads had the highest revenue scale, reaching RMB 1.736 billion.

This number doesn't necessarily mean they've won. But it at least proves someone sold cards, and sold them to places capable of generating revenue. Anyone doing reinforcement learning knows that between a model "running" and being "trainable," there's a gap filled by environments, rewards, scheduling, VRAM, and operator compatibility. If you tell me about high compute density, I'll first ask if my PyTorch tasks can be migrated with minimal code changes.

Biren's contrast is more direct. The material mentions H1 revenue was only RMB 58.9 million against a loss of RMB 1.6 billion. This number deserves separate attention. The most expensive thing for domestic GPUs right now is trust. Customers don't dare push training tasks, platforms don't dare switch inference traffic, and labs certainly don't dare gamble project stability on them. With paper deadlines looming, who wants to spend two weeks tuning communication libraries just to test a new card?

So in this round of financial reports, revenue scale, revenue quality, and whether the ecosystem minimizes hassle are all worth examining. High revenue doesn't equal good profit, but revenue at least indicates customers are voting with their feet. No revenue means no feedback. No feedback makes it hard to iterate on operators, compilers, and cluster scheduling.

The four paths differ, and stories can be told either way. Once financial reports come out, we can at least see who has stepped out of the "domestic substitution" slogan. Right now, I'm watching to see who lets a PhD student finish a reinforcement learning task without writing adaptation code.

Stable delivery of compute power is the true judgment standard.