DeepSeek's parent company earned 5 billion yuan last year, which is enough to support 2,380 R1s.

One year after the debut of R1, DeepSeek still hasn't secured new financing.

While other large - model players are either going public or raising funds, DeepSeek remains aloof and has made few commercialization moves.

Even so, it hasn't lagged behind in AGI research —

It has continuously produced high - level papers, and the list of authors has remained quite stable. One author even "returned" for the new version of the R1 paper.

Actually, there's no need to worry about whether DeepSeek has enough resources. After all, the latest news is...

Magic Square Quant earned 5 billion RMB last year.

The Soaring Magic Square Quant

Liang Wenfeng's main business has taken off.

According to Private Equity Ranking Network, in 2025, almost every fund under Magic Square Quant had a yield of over 55%.

It's worth noting that last year was a bumper year for Chinese quantitative funds, with an average yield of 30.5%, more than twice that of global competitors.

Even in such a booming market, Magic Square Quant still stood out —

With an average yield of 56.6%, it ranked second among China's 10 - billion - level quantitative funds, only behind Lingjun Investment with a 70% yield.

Considering that Magic Square Quant manages assets worth over 70 billion RMB, this astonishing yield undoubtedly allowed the company to rake in huge profits.

According to Bloomberg, an investment director of a Shanghai private equity fund believes that:

Magic Square Quant, this cash - cow, might have helped Liang Wenfeng earn over 700 million US dollars (about 5 billion RMB) last year alone.

Just stating this figure might not give you a clear idea. Well, MiniMax, which has been soaring recently, is pre - raising over 600 million US dollars for its IPO.

For Magic Square Quant to earn 700 million on its own is quite remarkable.

Since DeepSeek's research funds come from Magic Square Quant's R & D budget, this huge sum can undoubtedly provide more resources for DeepSeek.

What's even more crucial is sustainability.

According to DeepSeek, the V3 training cost only 5.576 million US dollars, and R1 cost only 294,000 US dollars.

If calculated based on these figures, Magic Square Quant's income from last year alone could support the reproduction of —

125 V3s and 2,380 R1s.

Considering that DeepSeek is further improving training efficiency, with 700 million US dollars in cash...

There's simply too much money to spend.

The Real AGI Player

In 2025, large - model players racked their brains to raise funds.

Take OpenAI for example:

First, there were early signs in the first half of the year. Altman, who used to hate ads, suddenly had a 180 - degree turn on his podcast, saying "it's worth a try".

Sure enough, a series of money - making moves followed.

The most notable one was the "capital internal circulation" around October last year. OpenAI used its huge computing power demand as a bargaining chip to attract investments from chip and cloud providers.

With this "brilliant move", OpenAI has squeezed over 1 trillion US dollars from giants like Nvidia, AMD, and Oracle.

That's not enough. OpenAI is also actively promoting a wide range of product portfolios, including Sora, Codex, ChatGPT Health, and a bunch of small features like shopping.

Although they are still in the early stages, they are all potential money - making channels.

Comparing OpenAI's tactics in Silicon Valley with DeepSeek, the latter is extremely pure.

When it says it wants to do AGI, it focuses solely on AGI research without any intention of making money.

Powerful models like R1 are open - sourced right away without any thought of making a profit from them.

After its success, it didn't rush to turn the model into products but mainly focused on API services.

It even hardly maintains its client - side, and most consumer users have migrated to other domestic models.

This might be the manifestation of its concept in resource allocation:

With limited computing power, it doesn't divert resources to application scenarios requiring high - concurrency inference but continues to fully invest in underlying training.

Since then, DeepSeek has remained extremely low - key. From Liang Wenfeng to the core research team, few people have made public statements.

However, it has been very active in the academic circle. It has been continuously releasing important research throughout the year, leaving people dazzled.

In the second half of the year alone, well - known ones include OCR and V3.2, and there were frequent moves at the end of the year. More than 60 pages of valuable content that could have been published as new papers were added as supplements to R1.

Recently, it open - sourced the memory module.

In an increasingly tense and anxious industry environment, DeepSeek can always adhere to its original intention of AGI research, thanks to the support of Magic Square Quant.

In fact, DeepSeek is probably the only AI Lab in the world that has not received external financing and is not affiliated with any large company.

Moreover, the main business of its parent company, Magic Square Quant, has no conflict with AI.

As a result, there are no internal or external obstacles on DeepSeek's AGI path.

Externally, DeepSeek's funds come entirely from its parent company.

Magic Square Quant stopped accepting external funds several years ago.

This means that DeepSeek is not restricted by any equity structure or profit - loss expectations and doesn't need to consider investment returns in the short term.

Internally, AI was originally a "side business" to assist quantitative investment.

For Magic Square Quant, the AI business line is not forced but a complementary relationship, so there won't be the problem of a large ship being difficult to turn around.

In summary, Magic Square Quant not only shelters DeepSeek from the wind and rain but also continuously supplies it with nutrients.

This ability to cross - subsidize based on existing businesses is often underestimated by the market.

It's important to note that in the context of highly uncertain AI application scenarios, having a mature and market - proven business model as a backing is quite an advantage.

Take Google for example. Its search business has been criticized for dragging down Gemini, and there were frequent strategic mistakes in the early stage.

However, even in such a situation, Gemini has survived until now and even outperformed ChatGPT.

Behind this is the support of advertising revenue from the search business.

OpenAI doesn't have a traditional business, which is why it was able to take the lead. But without a stable cash flow, it has little room for error.

Once its model capabilities are surpassed and the brand effect of being the state - of - the - art weakens, every subsequent move will be as risky as walking on a tightrope.

In contrast, DeepSeek, which has both a business model and is AI - native, combines the advantages of Google and OpenAI.

A Pure Company Attracts Pure - Hearted People

Of course, the deeper impact of this internal self - sufficiency ability lies in "people".

Perhaps because it doesn't have to worry about money, DeepSeek's R & D team can fully immerse themselves in AGI research, maintaining a pure research environment internally.

This can be seen from the author list of the recently updated R1 paper —

Nearly a year after the paper was published, all 18 core contributors are still on the DeepSeek team.

Among the more than 100 authors in total, only 5 are marked with an asterisk (indicating they have left the team).

In last year's author list, there were 6 asterisks — one more than this year.

The missing asterisk belongs to Ruiqi Ge, who has now returned to the team.

In the highly competitive AI talent war, not only has DeepSeek's team not lost many members, but one has even "returned".

More and more idealistic people are joining DeepSeek, and Magic Square Quant's huge income allows DeepSeek to provide top - level resources and generous compensation packages for these researchers — There's no need to work for passion alone on the path of AGI research!

In summary, having a full wallet brings peace of mind.

Next, just sit back and wait for DeepSeek to unveil R2/V4.

After all, DeepSeek, which now has over 700 million US dollars in hand —

It has sufficient supplies before the battle.

One More Thing

While Magic Square Quant is making huge profits, more observant netizens are also benefiting from DeepSeek.

Many stock - trading friends treat DeepSeek's technical papers as research reports.

Previously, when DeepSeek released the V3 technical report, there was a section called "Hardware Design Suggestions", which mentioned the chip standards for the model design.

Since then, every time DeepSeek releases a new model, even if it doesn't mention it in the paper, it will publish a blog to announce the latest hardware developments.

You can probably imagine what happens next. As soon as the news comes out, many domestic chip companies will adapt their products to DeepSeek.

Then... their stock prices skyrocket.

For example, just four minutes after the release of V3.2 at the end of September last year, Cambricon announced that it had completed the adaptation of the DeepSeek framework around DSA.

Sure enough, when the market opened the next day, the stock price jumped by nearly 5%.

Many netizens who targeted this as an investment direction also made a fortune.

So, what DeepSeek provides is not just "hardware design suggestions"...

It's simply the investment guide endorsed by Buffett.

Reference links:

[1]https://www.bloomberg.com/news/articles/2026-01-12/deepseek-founder-liang-s-funds-surge-57-as-china-quants-boom

[2]https://x.com/i/trending/2010694416898629671

[3]https://finance.sina.com.cn/stock/t/2025-09-30/doc-infseunc2994903.shtml

This article is from the WeChat official account "QbitAI", author: Jay, published by 36Kr with authorization.