Kunlunxin

Um, they say, "Last, one more technical contribution lies in the fact that from the first day, GLM-V is full stack adapted to Chinese GPU ecosystems.

We have successfully completed deep optimization, spanning from underlying kernels to upper level inference frameworks across seven mainstream domestic chip platforms, including Huawei Ascend, MoreThreads, Higon, Cambracon, Kunlunxin, MetaX and Enflame." When you're talking about the model helping bring up the infrastructure for serving GLM-V, it's not necessarily the case that this is just Nvidia GPUs and kernels and the serving stack that we're talking about for SGLang on the AMD or Nvidia GPUs that we test in InferenceX.

They're serving this model, at least they claim, for many months on seven different domestic Chinese accelerators.

What do you guys make of that?

Huawei's AI accelerator output reached 572,000 units, nearly double the 300,000-plus recorded the previous year, giving it a 28% market share.

T-Head, the chip design arm of Alibaba, accounted for 7%, Cambricon captured 5%, Baidu's Kunlunxin unit drew 4%, and Haigon registered 3%.

Of 11 domestic chip manufacturers monitored by IDC, nine posted year-over-year gains.

Globally, AI accelerator shipments topped 2.008 million units across the first two quarters of the year.

In mainland China, where authorized GPU inventory remained constrained, GPU deliveries dropped 26% to 991,000 units.

Non-GPU accelerator volumes climbed roughly 80% to just over 1 million units, pulling ahead of GPUs on a unit count basis.

Four vendors—Huawei, T-Head, Kunlunxin, and Cambricon—held a combined 95% share of that non-GPU category, which IDC noted had drawn uptake from government bodies, internet firms, and telecoms operators.

IDC says the competitive focus among chipmakers has shifted beyond raw processor performance toward how effectively chips integrate and operate within larger computing clusters.