· AI
DeepSeek Is Raising $12 Billion. The Chips It Bought Still Can't Train Its Models.
Investors just doubled down on DeepSeek. The chips at the center of that bet are still doing the one job Huawei built them for, just not for DeepSeek.
DeepSeek's first outside funding round has grown fast. In June, Technode reported the company was targeting 50 billion yuan, about $7.4 billion, at a post-money valuation of $52 to $59 billion, with founder Liang Wenfeng putting in 20 billion yuan of his own money alongside Tencent, CATL, NetEase, JD.com, and China's national AI fund. By October, Bloomberg's reporting had the round at a minimum of 80 billion yuan, roughly $12 billion, demand pushing it toward 100 billion, ahead of a planned 2027 listing. That round outgrew its own target by more than 60 percent in four months. Investors are piling in harder than the company first asked for, not getting cold feet.
Some of that money has a specific destination. Technode also reported that DeepSeek plans to deploy at least 160,000 of Huawei's new Ascend 950DT accelerators at a gigawatt-scale data center in Ulanqab, Inner Mongolia, for inference only, running finished models to answer users rather than building the next one. Huawei designed and marketed the 950DT as a training chip. DeepSeek is buying it and pointing it at the other job anyway.
That is the most honest number in the whole story, more honest than the valuation and more honest than the chip count on its own. It says where Huawei silicon actually stands against Nvidia on the hardest workload in AI, a year after we first learned why.
Steelman the bull case
The bull case here is not naive. A round that grows 60 percent past its own target, backed by Tencent and CATL, two of the most sophisticated corporate investors in China, is a real market signal, not hype dressed up as diligence. Routing new Huawei chips to inference while keeping training on proven hardware is also, read generously, discipline rather than defeat: use each chip for the job it is actually good at, and do not bet a frontier training run on silicon that has not earned that trust yet. That is a defensible engineering call, not a cover story.
Why the inference-only choice is the tell
Here is why I do not read it that way. The Financial Times' reporting, via The Register, found that when DeepSeek tried to train its R2 model on the prior generation of Ascend hardware in 2025, every attempted run failed before finishing. The chips kept faulting mid-run, the links between them could not move data fast enough to keep a large training job synchronized, and the software stack was too immature to work around it, missing support for a numeric format DeepSeek's training process depended on. The fix was a split: Nvidia H20 chips for training, Huawei Ascend demoted to inference. A year later, with a new Huawei chip built specifically to compete on training, DeepSeek is running the identical split. If the 950DT had closed the gap, training is exactly where DeepSeek would want to prove it, especially with a government and an investor base ready to reward the headline. It has not happened. That does not prove the gap is permanent, but it proves it has not closed, on DeepSeek's own roadmap, which tells you more than any investor deck does.
The valuation is getting ahead of the numbers DeepSeek has actually shown
Now the money. The Information's reporting put DeepSeek's annualized revenue at approaching $500 million, the first real figure the company has disclosed, on 70 to 80 percent gross margins. I want to flag plainly that this is a single-source, unaudited number, so treat it as directional, not confirmed financials. Even taking it at face value, a company with roughly half a billion in disclosed revenue saw its pre-money valuation jump from the $52 to $59 billion range in June to around $71 billion in July, before this latest round pushes the capital raised itself toward $12 billion or more. That math requires betting heavily on growth nobody has shown yet, not growth already delivered. None of that makes DeepSeek's low prices fake. Its per-token pricing is real, independently benchmarked, and has genuinely forced frontier labs to compete on cost, which is the free market doing exactly what it is supposed to do. But a round this size is pricing in a chip and revenue story the company's own recent history has not backed up yet.
What this means if you are evaluating these models
This does not change the advice I gave when V4.1 Flash launched: if an open-weight DeepSeek model fits your workload, get it through a US-based host running it on proven hardware, not DeepSeek's own app. If anything, hold that line harder. The economics you can bank on today run on Nvidia-class chips with a real production record. The chip-sovereignty bet is real money, and it may pay off eventually. It has not paid off yet, and the clearest evidence is sitting in Inner Mongolia, doing inference. If you want a second opinion on what an open-weight model actually costs once you account for where and how it is hosted, that is worth a conversation.
Sources
References used in this article. Links also appear alongside the relevant claims.
Let's make it happen.
You don't need a polished brief. A couple of lines about where you want your company to go is plenty, and we'll come back with what we'd tackle first.
We scope the work around your goals and timeline before quoting anything, so you know exactly what you're getting.