Self-developed chips: The arithmetic problem for DeepSeek and Zhipu

Original Author: Xiao Suan

In 2013, engineers at Google worked on a simple arithmetic problem. If every user spent three minutes per day using voice search, how much would Google's global data centers need to expand?

The answer left everyone stunned — they would need to double their capacity. Relying solely on buying NVIDIA GPUs to fill this gap would have bankrupted Google with the bills. So the search giant made what was then considered a radical decision: build its own chips. The rest is history — that chip became the TPU, now Google's strongest bargaining chip against the "NVIDIA tax."

Thirteen years later, this same arithmetic problem has landed in the hands of the Chinese.

On the evening of July 7th, Reuters, citing three sources, reported that DeepSeek is developing its own AI chip. The project started a year ago, and the company has already been in contact with chip design firms, wafer foundries, and memory manufacturers. A few hours later, The Information added another piece: Zhipu is also evaluating custom chip development, and is in talks with domestic chip design companies.

Within 24 hours, two of China's leading AI model companies were revealed to be pursuing the same strategy:

Building their own chips.

1.

DeepSeek's chip comes with a telling qualifier: it's designed for inference, not training.

Training is the process of teaching the model, incredibly expensive but a one-time cost. Inference is the model doing its job, burning electricity costs in the data center every time a user asks a question. 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 is never in the down payment, but in the rent.

The core problem DeepSeek prioritizes solving boils down to one question:

How much does it cost to serve each user?

Liang Wenfeng, the founder of DeepSeek, is one of the few people who saw chips as a life-or-death issue from day one. Coming from quantitative finance, he was known for stockpiling GPUs even before the AI model boom. In two interviews with Waves in 2023 and 2024, he said a sentence that has been repeatedly quoted since:

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

What he said, he put into action. DeepSeek's R1 model was trained on NVIDIA H800s, then switched to Huawei's Ascend; the engineering team designed the UE8M0 FP8 data format into the model, widely regarded in the industry as custom-tailored for the hardware characteristics of the next generation of domestic chips.

By June this year, the ammunition was ready. The company, which had refused external funding for years, completed its first financing round, obtaining approximately 51 billion RMB, with a post-investment valuation of 52 to 59 billion USD. The publicly stated use of the funds was clear: expanding domestic computing 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 represents a different approach to the same arithmetic problem.

This company, emerging from a Tsinghua University lab, was listed on the Hong Kong Stock Exchange this year under the title "First AI Model Stock," at one point reaching a market cap of over one trillion Hong Kong dollars. Behind the glitz lies a tense balance sheet: a loss of 2.958 billion RMB in 2024, another 2.358 billion RMB loss in the first half of 2025, burning through 5.3 billion RMB in a year and a half.

In February this year, GLM-5 was released, becoming a hit overseas, with its coding capabilities rivaling top-tier closed-source models. A flood of users poured in, and Zhipu's first move was to raise prices, starting with a 30% increase for its Coding package. Its second move was to issue a "Computing Partner" recruitment notice, publicly inviting chip manufacturers to collaborate on optimization.

A newly listed star company publicly posting for computing power. A business doing so well it has to raise prices to deter users is a rare sight in business history.

So The Information's leak was hardly surprising. Zhipu is evaluating a collaborative custom route: providing the model architecture and requirements itself while relying on domestic chip design companies for engineering capabilities.

DeepSeek builds its own factory to make cars. Zhipu takes blueprints to a car manufacturer for modifications. One route isn't inherently better than the other, but the difference shows up on the ledger.

3.

The most thought-provoking part of this chip-building movement might be a direct quote from Reuters:

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

The first part is almost a truism. Under export controls, NVIDIA's share of the Chinese data center market has already neared zero. The second part is the real news.

Over the past two years, "domestic substitution" in the context of computing power has practically meant "switching to Ascend." DeepSeek itself is the most active practitioner — the V4 series is fully adapted for Ascend, and Huawei confirmed its processors participated in part of the training. Zhipu has gone even further, adapting its GLM architecture to over 40 domestic chips; on the release day of its new model, Haiguang, Moore Threads, and Muxi lined up to announce complete compatibility.

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

Even if that supplier is domestic.

Embracing Ascend solves the problem of "availability." Self-developed chips solve the problem of "who's in control." As the narrative of domestic substitution enters its fifth year, internal differentiation has begun.

4.

Model companies building chips is already standard practice across the Pacific.

Last month, OpenAI announced a custom inference chip developed in collaboration with Broadcom, codenamed Jalapeño; Anthropic has been reported to be evaluating the same thing. Along with earlier moves by Google, Amazon, and Microsoft, any company in Silicon Valley with a large enough inference bill either has its own custom chip or at least a presentation on one.

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

On the one hand, custom orders from model companies are the dream revenue for domestic chip design firms. Zhipu's collaborative custom model almost seems written from their playbook. Memory manufacturers also benefit, as inference chips demand high bandwidth, ensuring a steeply rising demand curve for High Bandwidth Memory (HBM).

On the other hand, today's big customers are learning how to eventually operate independently. Google was once an excellent customer for chip suppliers too, until it became the master of the TPU.

Of course, the cards are only just being dealt. A competitive AI chip typically takes years and billions of dollars in investment, with no guarantee of success — Meta once completely scrapped its in-house chip project. More subtly, custom chips rely on the model architecture remaining relatively stable, yet the next-generation models from DeepSeek and Zhipu are just adopting mechanisms like sparse attention. By the time the chips sent for tape-out today come off the production line in two years, the model architecture may have already evolved.

In 2013, Google's arithmetic problem led to the TPU.

In 2026, these Chinese model companies have just begun writing their solution. The person asking the question has changed, but the logic of the solution remains the same:

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