Written by: Xiaosuan

In 2013, Google's engineers calculated an arithmetic problem.

The question was simple: If each user uses 3 minutes of voice search every day, how much does Google need to expand its global data centers?

The answer left everyone gasping: It would double.

Trying to fill that gap by purchasing NVIDIA graphics cards, Google would be crushed by the bills first. So this search company made a decision that seemed unconventional at the time: to build its own chips. The subsequent story is well-known; that chip was called TPU, which is now the strongest leverage Google has against the "NVIDIA tax."

Thirteen years later, this arithmetic problem reached the hands of the Chinese.

On the evening of July 7, Reuters cited three informed sources, stating that DeepSeek is developing its own AI chips, with the project starting a year ago and already in talks with chip design firms, foundries, and storage manufacturers. A few hours later, The Information added that Zhipu is also evaluating self-developed custom chips and is in contact with local chip design companies.

Within 24 hours, the two leading Chinese model companies were revealed to be taking the same action:

Making chips.

1.

DeepSeek's chip has a thought-provoking qualifier: it is aimed at inference, training not included.

Training is about teaching the model, which is incredibly expensive, but it's a one-time payment; inference is when the model goes to work, and every time a user asks a question, it burns electricity in the data center. The more users, the more it burns, and it never stops.

Training is like buying a house; inference is akin to paying rent. The true cost black hole in the AI industry has never been in the down payment but in the rent.

The priority problem that DeepSeek needs to solve can be translated into one sentence:

How much does it cost to serve each user.

The founder of this company, Liang Wenfeng, is one of the few who has treated chips as a matter of life and death from day one. He comes from a quantitative fund background and made a name for himself in the industry by stockpiling graphics cards even before the big model craze. In 2023 and 2024, he gave two interviews to Duyun, where he stated a quote that has been repeatedly cited:

Our real challenge has never been funding, but the export ban on high-end chips.

What he said verbally, he is also doing. DeepSeek's R1 model was trained on NVIDIA H800 and then shifted to Huawei Ascend; the engineering team designed the UE8M0 FP8 data format in the model, which is widely recognized in the industry as tailored to the hardware characteristics of the next generation of domestic chips.

By June this year, the ammunition was also ready. This company, which has refused external investment for many years, completed its first round of financing, raising about 51 billion yuan, with a post-money valuation of 52 to 59 billion USD. The publicly disclosed use of funds was clearly stated: to expand the domestic computing power center and develop AI chips in-house.

In recent months, DeepSeek has been hiring chip design engineers, with all positions not appearing on any public recruitment platforms.

2.

Zhipu is another solution to the same arithmetic problem.

This company, which emerged from Tsinghua Laboratory, went public this year, proudly wearing the title of "first stock in large models," with a market value that briefly surpassed 1 trillion HKD. Behind the glory is a tight financial report, with a loss of 2.958 billion yuan in 2024 and another loss of 2.358 billion yuan in the first half of 2025, burning through 5.3 billion in a year and a half.

In February of this year, GLM-5 was released, becoming a sensation overseas, with programming capabilities close to those of top closed-source models. With massive traffic pouring in, Zhipu's first response was to raise prices, increasing its Coding package prices by at least 30%; the second action was to issue a "Computing Power Partner" recruitment announcement, openly inviting chip manufacturers to collaborate on optimization.

A newly listed star company publicly posted looking for computing power. The business was so good that they had to rely on price increases to deter users, which is rare in business history.

Thus, The Information's report was not surprising. Zhipu's evaluated approach is cooperative customization, where it provides the model architecture and requirements, and local chip design companies contribute engineering capabilities.

DeepSeek builds its factory to make cars; Zhipu uses blueprints to find car manufacturers for modifications. There is no distinction of high or low in the route; there is a difference in the bill.

3.

The most noteworthy aspect of this chip-making movement is a line from Reuters:

DeepSeek's chip-making aims to reduce dependence on NVIDIA and Huawei.

The first half of that sentence is almost trivial. Under export control, NVIDIA's share in the Chinese data center market has nearly gone to zero; the second half is the real news.

In the past two years, the term "domestic substitution" in the context of computing power has roughly equated to "switching to Ascend." DeepSeek itself is one of the most proactive practitioners, with its V4 series completing Ascend adaptation, and Huawei confirming that its processor participated in some training. Zhipu has gone further, with the GLM architecture being adapted for more than 40 domestic chips; on the day the new model was released, Haiguan, Moore Threads, and Muxi queued up to announce that they had completed adaptations.

The deeper the embrace, the clearer one realization becomes. A company with inference bills in the billions cannot bet its lifeline on any single supplier.

Even if that supplier is a member of its own team.

Embracing Ascend resolves the question of "whether there is." Developing self-made chips addresses the question of "whose advice to listen to." As the narrative of domestic substitution enters its fifth year, internal stratification has begun.

4.

Chip-making by model companies is already standard practice across the Pacific.

Last month, OpenAI announced a custom inference chip in collaboration with Broadcom, codenamed Jalapeño; Anthropic was reported to be evaluating the same matter. Coupled with earlier efforts from Google, Amazon, and Microsoft, any company in Silicon Valley with significant inference bills has its own self-developed chip or at least a self-developed chip PPT.

For China's chip industry, this is a double-edged sword.

On one side, the custom orders from model companies are the dream income for local chip design companies; Zhipu's cooperative customization model is almost written according to their script; storage manufacturers also benefit, as inference chips are highly dependent on bandwidth, and the demand curve for high-bandwidth memory will only steepen.

On the flip side, today's big customers are learning how to break away from you tomorrow. Google was once a quality client for chip suppliers, but later, it became the master of TPU.

Of course, the cards have just been dealt. A competitive AI chip usually requires years of time and billions of investment, and no one guarantees success; Meta's self-developed chip plan once had to start over from scratch. Even more subtly, custom chips gamble on the stability of model architecture, while DeepSeek and Zhipu’s next-generation models have just begun using new mechanisms like sparse attention. Today, the blueprints sent for wafer production may be outdated by the time the chips roll off the line in two years.

In 2013, the answer to the question calculated by Google was TPU.

In 2026, this arithmetic problem for Chinese model companies has just begun; the person posing the question has changed, but the logic of solving the problem remains:

The longer the rent is paid, the more one wants to have their own house.