Picture a server hall in Ulanqab, Inner Mongolia, a few hundred kilometers northwest of Beijing. Racks going in by the thousand. The plan on paper calls for at least 160,000 Huawei Ascend 950DT chips. Somebody on that floor is holding a tablet, checking a manifest, and none of the part numbers say Nvidia.
That image is the actual story. Not the chip count, not the data center, not even the Huawei logo. The story is that DeepSeek and Huawei have teamed up to build software tools for advanced chips, and that is the part of Nvidia’s business that nobody has managed to copy.
Hardware is the easy half
I review AI tools for a living, which mostly means watching teams discover that the hard part was never the thing they budgeted for. Chip design is brutally difficult, but it is a known problem with known steps. You can throw money and fabs and engineers at it. China has been doing exactly that.
The thing that keeps Nvidia’s customers from walking away is not transistor density. It is the fifteen-plus years of accumulated software underneath every model anyone trains. The kernels. The compilers. The libraries nobody reads but everybody imports. The layer that turns theoretical chip performance into performance you can actually bill for. The verified reporting says this partnership targets software that can extract the best performance from chips. That is the moat, described in plain language, by the people trying to cross it.
If you have ever ported a training run to unfamiliar accelerators, you know the gap between a chip’s spec sheet and what you get on the first attempt. It is not small. It is often humiliating. Closing it takes years of unglamorous work that no press release ever covers.
Why DeepSeek specifically
DeepSeek is an interesting partner for Huawei because DeepSeek has a track record of squeezing results out of constrained conditions. That is a software discipline, not a hardware one. Pairing a team known for efficiency engineering with a chipmaker that needs its silicon to look good in the real world is a reasonable fit. Huawei brings the Ascend parts. DeepSeek brings the people who have reason to care very deeply about every wasted cycle.
DeepSeek has raised over $7 billion in venture funding to finance this infrastructure, which is the other half of the equation. Software optimization work at this depth is expensive and slow and produces nothing demo-able for a long stretch. You need patient money. Seven billion is patient money.
What I am not going to tell you
Reuters has reported, citing three people, that DeepSeek is developing its own AI chip. I have seen that framed as confirmed in several places this week. The sources do not provide a definitive answer on whether DeepSeek is building its own silicon, so I am leaving it there. Unverified is unverified, and the story is strong enough without it.
I also cannot tell you how the Ascend 950DT performs against current Nvidia parts on real training workloads, because nobody outside those buildings can. Treat any confident benchmark comparison you see as marketing until someone publishes numbers you can reproduce.
What this means if you build things
For most readers of this site, the near-term effect is close to zero. You are not buying Ascend chips. Your cloud provider is not switching. But the medium-term effect is worth tracking for a few reasons.
- A second viable software stack for AI training means Nvidia’s pricing power has an actual ceiling, eventually. Competition at the compiler and kernel layer is the kind that matters.
- Nvidia has been optimizing for Chinese models like DeepSeek and Qwen, and has flagged in an SEC filing tied to its second-quarter earnings that potential White House restrictions on AI developed in China are a business risk. Both things are true at once, which tells you how tangled this has become.
- If the tooling from this partnership ships publicly, it becomes something you can evaluate. Open optimization work tends to leak useful ideas regardless of who funds it.
The honest read
This is not a Nvidia-is-finished story. Nvidia’s position is built on a software ecosystem that took over a decade to accumulate and that thousands of teams depend on without thinking about it. Two companies deciding to attack that is meaningful. Two companies succeeding at it is a different claim, and we have no evidence for the second one yet.
What we have is a serious attempt aimed at the right target. Most challengers to Nvidia have gone after the chip and lost on the stack. DeepSeek and Huawei appear to have read the problem correctly. That alone makes this worth watching more closely than the usual run of Nvidia-killer announcements, which have a 100% failure rate so far.
Check back when there are numbers. Until then, the manifest in Ulanqab is the most interesting document in the industry.
🕒 Published: