Huawei Ascend

Reports emerged today, first from the Chinese financial outlet Star Market Daily, confirmed by TrendForce citing Kai Wen.

that DeepSeq has ordered approximately 160,000 Huawei Ascend 950DT chips for a planned 1GW data center in Ulaan Kaab, Inner Mongolia.

The stated goal? To pre-train its next-generation V5 model from scratch on those chips, with deliveries ramping through the fourth quarter of this year and the full training run likely in 2027.

This is a significant escalation.

Until now, Huawei's Ascend chips have mainly been used to run, to do inference on, existing DeepSeek models.

Training a frontier model on them is a different and much harder problem.

DeepSeek experienced stability and performance issues training on earlier Ascend generations.

And Daniel, the geopolitical subtext here is loud.

The first of these is what the actual yield rate of new logic DAI is.

is going to be, right? Well, let's take a step back and just to quickly review, you know, to make a modern AI accelerator, whether it's an NVIDIA Blackwell or a Huawei Ascend, you need a few parts.

You need a logic chip capable of conducting processing at scale, and then you need a memory stack to hold the model weights.

And then you need, finally, a third component, which is the packaging process.

Because I think right now, we see fairly minimal efforts.

The analogy I often use is you're kind of selling a car without the engine if you are trying to export the broader AI stack, but you don't have the most crucial component.

And so we see like little efforts by, say, Huawei to offer, you know, really a small number of Huawei Ascend 950 chips to a country like Egypt.

But then that's quickly sort of overwhelmed by a counteroffer from the U.S. because we have that difference in compute scale today and production scale today.

One of the big advances here is that with leadership from, for example, DeepSeek, There now exists a lot of good infrastructure for migrating from the NVIDIA platform to other chip platforms.

For example, the Huawei Ascend series.

So running the models to service customers, the post-training of models, and at least part of the pre-training of models here.

has been done on the Huawei hardware.

The other drop this week.

DeepSeek open-sourced DeepInfer, the engine behind its production v4.1 flash serving under Apache 2.0. It's built mostly on VLLM and targets NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.

By early Wednesday, it was the top story on Hacker News, past 2,000 points.

Read the readme, though.

While DeepInfer applies mainly to DeepSeq's own models, other developers could study the open-source software and adapt its techniques for their own inferencing systems." DeepInfer includes optimized kernels, parallelism methods for attention and mixture of experts layers, and a decoding method that uses draft tokens and confidence estimates to schedule verification.

It also offers asynchronous scheduling that overlaps CPU batch scheduling with device execution batch invariant inference deterministic sampling and structured outputs batch invariant inference and deterministic sampling work on nvidia hardware only no outside party has yet benchmark deep infers efficiency gains according to bloomberg the framework applies to deep seek's own models but outside developers might read through the code and borrow its techniques for their own systems bloomberg said DeepInfer's kernel list includes custom TileLang kernels and covers Huawei's Ascend 950 chips.

Last month, DeepSeek announced a partnership with Huawei to build programming tools for Ascend chips.

As part of that effort, it open-sourced six software modules for the Ascend platform, including a version of its TileLang chip programming language adapted to the Ascend 950." Inception has debuted Mercury 3, a diffusion LLM that it says tops GPT-6 Luna, Claude Haiku 5.5, and GLM 5.3 Flash on agentic benchmarks while being four to eight times faster.

Quoting the Deep View.

While the AI industry has bet heavily on LLMs and the massive data centers needed to scale them, one of the Neo Labs betting on a contrarian AI just released another new model that challenges the incumbent thesis.

This is when Huawei began to actually enter the global top 10 of chip designers in the world and which was unprecedented.

They were quickly growing as one of the leading chip design companies in the world and They had released an AI accelerator chip called the Huawei Ascend 910, which measured by raw performance was actually more powerful than the then leading edge NVIDIA chip, the NVIDIA A100.

Obviously, there's been questions about the maturity of Huawei software stack on net, I would say, probably at that time.

The NVIDIA chip was still more useful or AI because of the maturity of the software stack.

It's over $10 billion.

And more critically...

From a strategic perspective, when we're thinking about AI and AI chips, I think a number that is worth keeping in mind is that we mentioned that these 1980 systems, they enable seven nanometer logic, right? That's the logic that's used for Huawei's Ascend chips.

And one of the fundamentals in this paper that sort of has allowed us to draw the conclusions that we have drawn is we've modeled out not just their yields, but also their throughputs.