A Battery Giant Is Co-Funding China's $12B Bet on DeepSeek

CATL, the world's largest battery maker, doesn't train models. It builds the lithium-ion cells, gigafactories, and grid-scale storage systems that any serious AI compute cluster sits on top of. So when China Money Network reported on October 6 that the Chinese AI lab DeepSeek is closing in on a roughly $12 billion funding round backed by Tencent and CATL, the more interesting fact was not the dollar figure. It was which firm was writing the check.

CATL's involvement, if confirmed, would be an unusually direct signal that industrial energy capital is moving up the stack into frontier AI. DeepSeek, the Hangzhou-based lab that broke out internationally with its R1 reasoning model in early 2025, is reportedly seeking around 80 billion yuan, roughly $12 billion at current rates, in a round that Phemex, Quartz, The Next Web, and InsideAI also link to Tencent and CATL. Investing.com attributes the lead-investor structure to Bloomberg. None of the reporting includes an on-the-record statement from DeepSeek, Liang Wenfeng, or either named lead investor, and the round's closing status is not yet confirmed: Gurufocus's headline calls it "secured" while the other outlets describe it as nearing close or still being marketed.

Frontier model training has become bottlenecked on power more than on chips. A 100,000-GPU training cluster draws on the order of 100 to 200 megawatts of continuous power, by widely cited industry estimates, and the next generation will draw more. NVIDIA sells the silicon. But someone has to deliver the electrons, the cooling, and the firm grid capacity that makes a 24/7 training run possible. In China, the firms best positioned to do that are the same ones that already finance the country's lithium supply, battery manufacturing, and grid-scale storage, and CATL sits at the top of that stack.

When a battery giant co-leads a frontier AI round, it is pre-positioning around the constraint that will decide which labs get to train the next generation of models. The bet is that CATL wants skin in the model layer because compute, power, and storage are converging into a single industrial pipeline. The company that controls the upstream energy stack benefits more from being on the right side of the AI buildout than from selling batteries to whoever happens to win the model race.

Tencent, the Shenzhen-based internet and cloud conglomerate, has backed multiple Chinese model efforts and runs the GPU cloud business that many Chinese AI labs train on. Its involvement fits the standard pattern of a hyperscaler-adjacent investor underwriting domestic frontier compute. CATL's involvement does not. It looks more like a vertically integrated power play, where the backer that already controls the physical layer underneath a 100,000-GPU cluster wants a direct claim on the model that gets trained there.

Two practical questions follow. The first is whether $12 billion, even at a Chinese lab's cost structure, buys enough compute to train a frontier model for the next 18 months. The second is whether the round actually closes on the reported terms. The dollar figure is large by Chinese AI-lab standards but small relative to the compute commitments U.S. frontier labs have already locked in. And the timing matters: Phemex and Quartz both reference a planned 2027 IPO tied to the raise, which would put DeepSeek on public markets inside two years of its R1 moment. Whether the round closes as marketed, who is actually on the cap table, and how the cash gets split between compute, data, and talent are the facts that will decide whether this is a structural read on energy capital or a well-sourced capital headline.

DeepSeek has not publicly commented on the round, and CATL has not confirmed its participation. Watch the closing announcement and the cap table. If CATL is named as a strategic backer tied to compute or power, industrial energy capital has begun financing the model layer directly. If CATL shows up as a passive LP, the energy-gatekeeper read collapses into a routine mega-round.