Self-developed chip DeepSeek and Zhipu's arithmetic problem

Author: Xiao Suan

In 2013, Google’s engineers calculated a math problem.

The question was simple: if each user uses 3 minutes of voice search every day, how much would Google’s global data centers need to expand?

The answer made everyone gasp: double.

Trying to fill that gap by buying Nvidia graphics cards would crush Google under the bills first. So the search company made a decision that, at the time, looked heretical: it would build its own chips. The story that followed is well known—those chips are called TPUs, and today they are Google’s toughest card in its fight against the “Nvidia tax.”

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

On the evening of July 7, Reuters cited three people familiar with the matter: DeepSeek is developing its own AI chips. The project was launched a year ago and has already been in discussions with chip design companies, wafer foundries, and storage vendors. A few hours later, The Information added that Zhipu is also evaluating self-developed custom chips and is in talks with domestic chip design companies.

In 24 hours, two of China’s leading large-model companies—both—were reported to have taken the same action:

Make chips.

--

DeepSeek’s chip comes with an intriguing qualifier: it is intended for inference, no matter what—training doesn’t matter.

Training is what you do to teach the model; the cost is astonishing, but you pay it once. Inference is when the model goes to work—each time a user asks a question, the data center burns a bill for electricity. The more users there are, the more it burns, and it never stops.

Training is buying a house; inference is paying rent. The real cost black hole in the AI industry is never in the down payment—it’s in the rent.

The problem DeepSeek wants to solve, translated into plain language, is just one sentence:

How much does it cost to serve one user?

The company’s founder, Liang Wenfeng, is among the very few people who treated chips as a matter of life and death from day one. With a background in quantitative funds, he was already famous in the circle for stockpiling GPUs long before the large-model boom. In 2023 and 2024, he gave two interviews to Dark Waves, saying a line that would later be quoted repeatedly:

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

He said it, and he was also doing it. DeepSeek’s R1 model was trained on Nvidia H800s, and then shifted to Huawei Ascend. In the model, the engineering team designed a UE8M0 FP8 data format—an industry-recognized choice custom-tailored to the hardware characteristics of the next generation of domestic chips.

By June of this year, the ammunition was ready. The company, which for years had refused outside investment, completed its first round of fundraising, raising approximately 510 billion yuan, with a post-investment valuation of 52 billion to 59 billion USD. The company’s externally disclosed intended use of funds is spelled out clearly: expanding domestic compute centers, and self-developedAI chips.

In recent months, DeepSeek has been hiring chip design engineers. However, all of the positions have not appeared on any public recruitment platform.

--

Zhipu is another solution to the same math problem.

This company, which grew out of a Tsinghua laboratory, rang the bell on the Hong Kong stock exchange this year under the banner of “China’s first large-model stock.” Its market value once exceeded 1 trillion Hong Kong dollars. Behind the glamour is a tense set of financial statements: in 2024 it lost 2.958 billion yuan, and in the first half of 2025 it lost another 2.358 billion yuan—burning 5.3 billion yuan in just 18 months.

In February this year, GLM-5 was released. It went viral overseas and its coding capability was close to that of leading first-tier closed-source models. As the flood of traffic poured in, Zhipu’s first move was to raise prices—Coding package pricing increased by 30% or more. Its second move was to issue a “compute partner” recruitment notice, publicly inviting chip manufacturers to collaborate on optimization.

A just-listed star company publicly posting to find compute resources. Business was so good that it relied on price hikes to deter users—this is rarely seen in business history.

So The Information’s scoop is hardly surprising. Zhipu’s evaluated approach is collaborative customization: it provides the model architecture and requirements, while domestic chip design companies provide engineering capability.

DeepSeek builds its own factory to make cars; Zhipu takes the drawings to the car factory for retrofitting. There’s no hierarchy in the route—only differences in the bill.

--

In this chip-making campaign, the most worth savoring is a direct quote from Reuters:

DeepSeek makes chips to reduce dependence on Nvidia—and also on Huawei.

The first half is almost boilerplate. Under export controls, Nvidia’s share in China’s data center market is close to zero. The second half is the real news.

Over the past two years, the phrase “domestic substitution” in the context of compute power has basically been equivalent to “switching to Ascend.” DeepSeek itself is the most proactive practitioner—its V4 series has completed Ascend adaptation, and Huawei confirmed that its processors participated in part of the training. Zhipu has gone even further: it has adapted the GLM architecture to more than 40 domestic chips. On the day the new model was released, Hygon, Moore Threads, and Muxi lined up to announce completion of the adaptation.

The deeper the embrace, the clearer one thing becomes: a company whose annual inference bills run into the billions can’t stake its lifeline on any single supplier.

Even if that supplier is one of its own.

Embracing Ascend solves the “whether you have it” problem; developing your own chips solves the “who’s in charge” problem. Now in its fifth year, the narrative of domestic substitution has begun to stratify internally.

--

For model companies making chips, this is already a standard move on the other side of the Pacific.

Last month, OpenAI announced a custom inference chip in collaboration with Broadcom, codenamed Jalapeño; it was reported that Anthropic is also evaluating the same. Add in the earlier moves from Google, Amazon, and Microsoft, and in Silicon Valley, any company with a sufficiently large inference bill has a self-developed chip—or at least a self-developed chip PowerPoint slide.

For China’s chip industry supply chain, this is a double-edged sword.

On the positive side, custom orders from model companies are a dream source of revenue for domestic chip design firms. Zhipu’s collaboration-and-customization model is almost written according to their script; storage vendors also benefit. Inference chips are extremely dependent on bandwidth, and the demand curve for high-bandwidth memory will only become steeper.

On the flip side, today’s major customers are learning the skills to ditch you tomorrow. Google was once an excellent customer for chip suppliers—later, it became the owner of the TPU.

Of course, the cards have only just been dealt. A competitive AI chip usually takes years and billions in investment, and no one can guarantee success. Meta’s self-developed chip plan was even completely rolled back and restarted. More subtly, custom chips bet that model architectures will stabilize, but DeepSeek and Zhipu’s new generation of models have just adopted new mechanisms such as sparse attention. The tapeout blueprints delivered today—by the time the chips come off the production line two years later—the architecture may already be a thing of the past.

In 2013, Google’s math problem had an answer: TPU.

In 2026, China’s model companies have only just started writing theirs. The person who sets the question has changed, but the logic of solving it hasn’t:

The longer you pay rent, the more you want to have your own house.