自研晶片，DeepSeek和智譜的算術題

Author: Xiao Suan

In 2013, engineers at Google ran a calculation.

The problem was simple: if each user used 3 minutes of voice search per day, how much would Google's global data centers need to expand?

The answer left everyone stunned: double.

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

Thirteen years later, this same calculation has fallen into the hands of the Chinese.

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

In 24 hours, the two leading Chinese large model companies were revealed to be making the same move:

Chip-making.

1.

DeepSeek's chip has a telling qualifier: it's for inference, not training.

Training is about teaching the model; the cost is staggering but a one-time payment. Inference is the model going to work. 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 buying a house; inference is paying rent. The real cost black hole in the AI industry has never been in the down payment, but in the rent.

The problem DeepSeek prioritizes solving translates to one sentence:

How much does it cost to serve one user?

The company's founder, Liang Wenfeng, is one of the very few people who treated chips as a life-or-death issue from day one. Coming from a quantitative hedge fund background, he was well-known in the industry for hoarding GPUs even before the large model boom. Between 2023 and 2024, he gave two interviews to Dark Waves, where he said a line that has been frequently quoted since:

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

His actions matched his words. DeepSeek's R1 model was trained on Nvidia's H800 GPUs before shifting to Huawei's Ascend chips. Their engineering team designed the UE8M0 FP8 data format into the model, which the industry widely recognizes as being tailor-made for the hardware characteristics of the next generation of domestic chips.

By this June, the ammunition was ready. The company, which had refused external investment for years, completed its first funding round, securing approximately RMB 51 billion, with a post-money valuation between USD 52 and 59 billion. The public disclosure of the fund usage was crystal clear: expanding domestic computing power centers and developing proprietary 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 AI represents a different solution to the same calculation.

This company, spun out of a Tsinghua University lab, rang the bell on the Hong Kong Stock Exchange this year. Bearing the title of "First Stock of Large Models," its market capitalization once exceeded HKD 1 trillion. Behind the glory lay a strained balance sheet: a loss of RMB 2.958 billion in 2024, and another loss of RMB 2.358 billion in the first half of 2025, burning through RMB 5.3 billion in 18 months.

In February this year, GLM-5 was released and became a hit overseas, with its coding capabilities rivaling top-tier closed-source models. A flood of traffic poured in. Zhipu AI's first move was to raise prices, with its Coding plan prices increasing by at least 30%. Its second move was to issue a "Computing Power Partner" recruitment order, publicly inviting chip manufacturers to collaborate on optimization.

A newly listed star company publicly posting for computing power. It's rare in business history for demand to be so high that it relies on price hikes to discourage users.

So, The Information's scoop was hardly surprising. The route Zhipu AI is evaluating is collaborative customization: providing its own model architecture and requirements, while local chip design firms provide the engineering capability.

DeepSeek is building its own factory to make cars; Zhipu AI is taking blueprints to a car manufacturer for modifications. Neither route is inherently superior, but the difference shows up on the balance sheet.

3.

The most insightful part of this chip-making movement is a direct quote from Reuters:

DeepSeek is building chips to reduce its dependence on Nvidia, and on Huawei.

The first half is almost moot. Under export controls, Nvidia's market share in Chinese data centers is nearly zero. The second half is the real news.

For the past two years, the term "domestic substitution" in the context of computing power has been almost synonymous with "switching to Ascend." DeepSeek itself has been one of the most active practitioners. The V4 series completed Ascend adaptation, and Huawei confirmed that its processors participated in some of the training. Zhipu AI went even further; its GLM architecture has adapted to over 40 types of domestic chips, with Haiguang, Moore Threads, and Muxi lining up to announce successful adaptations on the same day a new model was released.

The deeper the embrace, the clearer one thing becomes. A company with an annual inference bill in the billions cannot stake its lifeline on any single supplier.

Even if that supplier is on home turf.

Embracing Ascend solves the problem of "having it." Developing proprietary chips solves the problem of "who's in charge." As the narrative of domestic substitution enters its fifth year, internal stratification has begun.

4.

Model companies making chips is standard practice across the Pacific.

Last month, OpenAI unveiled its custom inference chip, codenamed "Jalapeño," developed in partnership with Broadcom. Anthropic has been reported to be evaluating the same thing. Along with Google, Amazon, and Microsoft from earlier, any company in Silicon Valley with a sufficiently large inference bill now has its own proprietary chip, or at least a slide deck for one.

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

On one side, custom orders from model companies are a dream source of revenue for local chip design firms. Zhipu AI's collaborative customization model seems almost scripted for them. Memory manufacturers also benefit, as inference chips are heavily dependent on bandwidth, meaning the demand curve for high-bandwidth memory will only steepen.

On the other side, today's biggest customers are learning the skills to cut you out tomorrow. Google was once a premier customer for chip suppliers, and then it became the master of the TPU.

Of course, the cards have only just been dealt. A competitive AI chip typically takes years and billions in investment, with no guarantee of success. Meta's own chip project was once completely scrapped and restarted. More subtly, custom chips are a gamble that model architectures will stabilize, yet the next-generation models from DeepSeek and Zhipu AI have just begun using novel mechanisms like sparse attention. The blueprint sent for tape-out today might be based on an architecture that's obsolete by the time the chip comes off the line two years later.

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

In 2026, Chinese model companies have just started working on this calculation. The question setter has changed, but the logic of solving the problem remains the same:

The longer you pay rent, the more you want a house of your own.