Self-developed chips, the arithmetic problem of DeepSeek and Zhipu

Author: Xiao Suan

In 2013, engineers at Google did a simple math problem.

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

The answer made everyone gasp: double.

Relying on buying NVIDIA's GPUs to fill this gap would have crushed Google with the bill first. So the search company made a decision that seemed heretical at the time: build its own chips. The rest of the story is well known—that chip was called the TPU, and today it is Google's strongest leverage against the "NVIDIA tax."

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

On the evening of July 7, Reuters, citing three sources familiar with the matter, reported that DeepSeek is developing its own AI chip. The project started a year ago and has already been in contact with chip design companies, foundries, and memory manufacturers. A few hours later, The Information added that Zhipu AI is also evaluating self-developed custom chips and has been in touch with local chip design companies.

Within 24 hours, two of China's top large model companies were revealed to be taking the same action:

Building chips.

DeepSeek's chip has an interesting qualifier: it is for inference, not training.

Training is teaching the model, with huge one-time costs; inference is the model working on the job. Every time a user asks a question, electricity is burned in the server room. The more users, the more electricity—and it never stops.

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

The priority problem DeepSeek wants to solve translates to one sentence:

How much does it cost to serve each user?

The company's founder, Liang Wenfeng, is one of the very few people who have treated chips as a life-or-death issue from day one. Coming from a quantitative fund background, he was known for hoarding GPUs long before the large model craze. Between 2023 and 2024, he gave two interviews to DarkWave, where he said a sentence that has been widely quoted since:

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

What he said, he also did. DeepSeek's R1 model was trained on NVIDIA H800s, then shifted to Huawei's Ascend; the engineering team designed the UE8M0 FP8 data format into the model, widely acknowledged in the industry as tailor-made for the hardware characteristics of next-generation domestic chips.

By this June, the ammunition was ready. The company, which had long rejected external investment, completed its first funding round, raising about 51 billion RMB, with a post-investment valuation of $52 billion to $59 billion. The publicly disclosed use of funds was clear: expanding domestic computing power centers and self-developing AI chips.

In recent months, DeepSeek has been recruiting chip design engineers, and none of these positions appeared on any public recruitment platform.

Zhipu AI is another approach to the same math problem.

This company, born from Tsinghua University's lab, went public in Hong Kong this year under the banner "First Stock of Large Models," with a market value once exceeding one trillion HKD. Behind the glory is a tight balance sheet: a loss of 2.96B RMB in 2024 and another 2.36B RMB in the first half of 2025, burning 5.3 billion in a year and a half.

In February this year, GLM-5 was released, going viral overseas with coding abilities rivaling top closed-source models. A flood of traffic came in. Zhipu's first move was to raise prices—the Coding package price increased by at least 30%; the second move was to release a "Computing Power Partner" recruitment order, publicly inviting chip manufacturers to collaborate on optimization.

A newly listed star company openly posting to find computing power. Business was so good that they had to raise prices to discourage users—rare in commercial history.

So the reveal by The Information was no surprise. The route Zhipu is evaluating is collaborative customization, where they provide the model architecture and requirements, and local chip design companies provide engineering capabilities.

DeepSeek builds its own factory and car; Zhipu takes the blueprint and finds a car factory to modify. There is no distinction between high and low routes—only the difference in the bill.

In this chip-building movement, the most noteworthy is a line from Reuters:

DeepSeek is building chips to reduce dependence on both NVIDIA and Huawei.

The first half is almost a cliché. Under export controls, NVIDIA's share in China's data center market has nearly dropped to zero. The second half is the real news.

Over the past two years, the phrase "domestic substitution" in computing power has essentially meant "switching to Ascend." DeepSeek itself is the most active practitioner—the V4 series has completed Ascend adaptation, and Huawei confirmed its processors participated in some training. Zhipu went further, adapting the GLM architecture to over 40 domestic chips. On the day the new model was released, Haiguang, Moore Threads, and Muxi lined up to announce completed adaptations.

The deeper the embrace, the more one thing becomes clear: a company with an annual inference bill in the billions cannot bet its lifeline on any single supplier.

Even if that supplier is one of their own.

Embracing Ascend solves the problem of "having or not having"; self-developed chips solve the problem of "who calls the shots." The narrative of domestic substitution has reached its fifth year, and internal stratification has begun.

Model companies building chips is already standard practice on the other side of the Pacific.

Last month, OpenAI revealed its custom inference chip, codenamed Jalapeño, developed in collaboration with Broadcom; Anthropic was reported to be evaluating the same thing. Plus the earlier moves by Google, Amazon, and Microsoft—every company in Silicon Valley with a large enough inference bill has its own self-developed chip, or at least a PPT for one.

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

On the positive side, custom orders from model companies are a dream come true for local chip design firms. Zhipu's cooperative customization model is almost written from their script. Memory manufacturers also benefit—inference chips are highly dependent on bandwidth, and the demand curve for high-bandwidth memory will only get steeper.

On the negative side, today's big customers are learning how to leave you behind tomorrow. Google was once an excellent customer for chip suppliers; later, it became the master of TPU.

Of course, the cards have just been dealt. A competitive AI chip typically takes years and billions of dollars, and success is not guaranteed. Meta's self-developed chip plan was entirely scrapped and restarted. More subtly, custom chips bet on model architecture becoming stable, while DeepSeek and Zhipu's next-generation models have just adopted new mechanisms like sparse attention. The blueprint sent for tape-out today might be outdated by the time the chip is ready two years later.

In 2013, the answer to Google's math problem was the TPU.

In 2026, this math problem for China's model companies has just begun. The questioner has changed, but the logic of the solution remains:

The longer rent is paid, the more one wants a house of their own.