Author: Xiaosuan
In 2013, Google engineers did a simple arithmetic problem.
The problem was simple: if each user used voice search for 3 minutes a day, how much would Google's global data centers need to expand?
The answer made everyone gasp: double.
By buying Nvidia graphics cards to fill this hole, Google would be overwhelmed by the bills first. So the search company made a decision that seemed unconventional at the time: to manufacture its own chips.
The rest is history. That chip was called TPU, and it's now Google's strongest bargaining chip against the "Nvidia tax." Thirteen years later, this arithmetic problem fell into the hands of the Chinese. On the evening of July 7th, Reuters, citing three sources familiar with the matter, reported that DeepSeek is developing its own AI chip. The project started a year ago and it has already been in contact with chip design companies, wafer foundries, and memory manufacturers. A few hours later, The Information added that Zhipu is also evaluating its self-developed custom chip and is in contact with local chip design companies. Within 24 hours, two of China's top model companies were exposed for the same move: chip manufacturing. DeepSeek's chip has an intriguing description: it's inference-oriented, training is not a concern. Training is about teaching the model, which is incredibly expensive, but paid for upfront. Inference is about the model actually doing the work; every time a user asks a question, the server room burns through its electricity bill. The more users there are, the more electricity is consumed, and it never stops. Training is like buying a house; inference is like paying rent. The real cost black hole in the AI industry is never in the down payment, but in the rent. DeepSeek's primary concern, which can be summarized in one sentence: How much does it cost to serve each user? The company's founder, Liang Wenfeng, is one of the very few who has considered chips a matter of life and death from day one. He comes from a quantitative fund background and was known in the industry for hoarding graphics cards long before the big model craze. In two interviews with Dark Surge in 2023 and 2024, he said something that has since been repeatedly quoted: "Our real challenge has never been funding, but the export ban on high-end chips." They said it, and they did it. DeepSeek's R1 model was trained on NVIDIA H800 and then switched to Huawei Ascend; the engineering team designed the UE8M0 FP8 data format in the model, which is widely recognized in the industry as being tailor-made for the hardware characteristics of the next generation of domestic chips. By June of this year, they had also prepared their ammunition. This company, which had refused external investment for many years, completed its first round of financing, raising approximately 51 billion yuan, with a post-investment valuation of 52 billion to 59 billion US dollars. The publicly disclosed use of funds is clearly stated: to expand domestic computing power centers and to develop AI chips. In recent months, DeepSeek has been recruiting chip design engineers, but none of these positions have appeared on any public recruitment platforms. 2. Zhipu is another solution to the same arithmetic problem. This company, which originated from a Tsinghua University laboratory, went public on the Hong Kong Stock Exchange this year, boasting the title of "the first stock in large-scale modeling," and its market value once exceeded one trillion Hong Kong dollars. Behind the glory is a tight financial statement: 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 yuan in just one and a half years. In February of this year, GLM-5 was released and became a hit overseas, with its programming capabilities approaching those of top-tier closed-source models. Faced with a massive influx of traffic, the first thing Zhipu did was raise prices, increasing the price of its coding packages by at least 30%; the second thing was to issue a "computing power partner" recruitment notice, publicly inviting chip manufacturers to cooperate in optimization. A newly listed star company publicly advertised for computing power. Its business is so booming that it has to raise prices to deter users—a rare occurrence in business history. Therefore, The Information's exposé is not surprising. Zhipu's evaluation strategy is collaborative customization: it provides the model architecture and requirements, while local chip design companies provide engineering capabilities. DeepSeek, on the other hand, builds its own factory to manufacture cars; Zhipu takes the blueprints and finds car manufacturers to modify them. There's no inherent hierarchy in their approaches, only differences in their billing. 3. In this chip-making movement, the most noteworthy statement is a direct quote from Reuters: DeepSeek is making chips to reduce its dependence on Nvidia and Huawei. The first half of the sentence is almost redundant. Under export controls, Nvidia's market share in China's data center market has almost disappeared. The second half is the real news. Over the past two years, in the context of computing power, the phrase "domestic substitution" has been roughly equivalent to "switching to Ascend." DeepSeek itself is the most active practitioner, with its V4 series completing Ascend adaptation, and Huawei confirming that its own processors participated in some training. Zhipu has gone even further, with its GLM architecture adapted to more than 40 domestic chips. On the day of its new model release, Hygon, Moore Threads, and Muxi lined up to announce that they had completed adaptation. The deeper the embrace, the clearer one understands one thing: a company with annual inference bills in the billions cannot stake 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" chips; developing our own chips solves the problem of "who to listen to." The narrative of domestic substitution has entered its fifth year, and internal stratification has begun. Model companies developing chips is already standard practice across the Pacific. Last month, OpenAI announced a custom inference chip developed in collaboration with Broadcom, codenamed Jalapeño; Anthropic was reportedly evaluating the same. Adding to the earlier examples of Google, Amazon, and Microsoft, virtually every Silicon Valley company with a large enough inference bill has at least one self-developed chip, or at least a PowerPoint presentation showcasing it. For China's chip industry chain, this is a double-edged sword. On the one hand, custom orders from model companies represent the coveted revenue of domestic chip design companies; Zhipu's collaborative customization model is almost a carbon copy of theirs. Storage manufacturers also benefit, as inference chips are extremely reliant on bandwidth, and the demand curve for high-bandwidth memory will only steepen further. On the other hand, today's major customers are learning how to outpace you tomorrow. Google was once a high-quality customer of chip suppliers; later, it became the owner of TPUs. Of course, the cards have only just been dealt. A competitive AI chip typically requires years of development and billions of dollars in investment, and there's no guarantee of success. Meta's self-developed chip plan was completely scrapped and restarted. More subtly, custom chips rely on the model architecture stabilizing, while DeepSeek and Zhipu's next-generation models have just adopted new mechanisms like sparse attention. The blueprints sent for tape-out today might have completely changed architecture two years later when the chip rolls off the production line. In 2013, Google solved that problem with the TPU. In 2026, the Chinese model company had just begun posing this question. The person setting the question had changed, but the underlying logic remained the same: The longer you pay rent, the more you want to own your own house.
Kikyo
Weatherly
Anais
Catherine
Joy