DeepSeek, chasing Moutai

In the first half of the year, China's A+H share market saw the consecutive emergence of three AI industry chain companies with a market value exceeding one trillion yuan: Zhipu, Cambricon, and Zhongji XuChuang.

A one-trillion-yuan market value has become the standard for top global AI companies. It is not a ceiling, but an admission ticket. Previously, in the wealth circle, this domain was entirely dominated by Moutai. With a market value of 1.5 trillion yuan, it steadily generates hundreds of billions in profits every year. It represents certainty, a competitive moat, and the compounding effect of time.

As China's equivalent of OpenAI, DeepSeek is rumored to list on the Science and Technology Innovation Board of the Shanghai Stock Exchange in 2027. It recently launched its second round of financing, with a latest valuation of approximately 71 billion US dollars, equivalent to about 480 billion yuan.

The pre-listing valuations of OpenAI and Anthropic have already touched the one-trillion-US-dollar threshold. It is worth imagining when DeepSeek will also reach a valuation of one trillion yuan.

When global AI capital votes in units of trillions, does China truly have its own "OpenAI-style answer" at the table, and can it become the new Moutai in the AI era?

With an annualized revenue of 500 million US dollars, why does it support a valuation of 480 billion yuan?

DeepSeek's business structure is becoming increasingly clear.

Annualized revenue of 400 to 500 million US dollars (ARR) — mainly from enterprise API call services.

Gross profit margin exceeding 50% — mainly from the V4 API. Even though the API price is not particularly high, the gross profit margin remains strong.

71 billion US dollars (approximately 480 billion yuan) — this is the pre-money valuation for DeepSeek's second round of financing. Only six weeks have passed since the first round's post-money valuation of 52 billion US dollars, marking a nearly 40% surge in valuation.

148 times — this is DeepSeek's latest price-to-sales ratio (P/S) relative to its annualized revenue, far higher than most AI startups.

36 billion US dollars — this is the personal net worth of founder Liang Wenfeng. According to the Bloomberg Billionaires Index, he has surpassed Anthropic CEO Dario Amodei (around 15.5 billion US dollars) and OpenAI President Greg Brockman (around 28 billion US dollars) to top the list of the world's wealthiest AI company founders.

These numbers paint a highly striking picture: DeepSeek generates less than 500 million US dollars in annual revenue, yet it is being priced at over 140 times its revenue multiple.

What does this multiple mean in practical terms?

According to publicly disclosed information, Anthropic has a valuation of approximately 965 billion US dollars, with annualized revenue exceeding 47 billion US dollars and a P/S ratio of around 21 times.

OpenAI has a valuation of approximately 852 billion US dollars, with total revenue of 13.07 billion US dollars and a net loss of 38.5 billion US dollars in 2025, resulting in a P/S ratio of around 65 times.

DeepSeek's valuation multiple has already far exceeded that of the two major US giants OpenAI and Anthropic, even though its annualized revenue is only a small fraction of Anthropic's.

Looking at Chinese peers, as of July 16, Zhipu's closing market value on the Hong Kong Stock Exchange was approximately 794.8 billion Hong Kong dollars (equivalent to about 102.4 billion US dollars). Its revenue for the 2025 fiscal year was approximately 724 million yuan (about 107 million US dollars), giving it an extremely high P/S ratio of 957 times. Zhipu's P/S ratio is even higher than DeepSeek's revenue multiple, leaving the two US giants far behind.

Although there is a gap in their P/S ratios, both are more than sufficient when compared across sectors to Moutai, which has a P/S ratio of just over 10 times. DeepSeek, with 500 million US dollars in annual revenue, supports a valuation nearly one-third that of Moutai, while Moutai's annual attributable profit exceeds 80 billion yuan.

This reveals a "Chinese logic" to investors: The Chinese capital market's pricing of so-called "scarce AI assets" is unique worldwide, even breaking away from traditional valuation frameworks.

The reason Zhipu can be so "outstanding" lies in its technological scarcity, scarcity of tradable shares (small circulating share capital), and strategic scarcity (aligning with China's national narrative).

It positions itself in the "first tier" globally. Its GLM-5.2 model approaches Anthropic's performance in scenarios such as coding and AI agents, representing one of the domestic large model technological peaks.

If DeepSeek lists on the Science and Technology Innovation Board in the future, its scarcity may be even higher than Zhipu's, with a positioning as "the peak of China's AI underlying technology."

Its V4 Pro model has 1.6 trillion parameters, far exceeding Zhipu GLM-5.2's 0.7 trillion and MiniMax M3's 0.4 trillion. In the Vals AI benchmark, its V4 model ranks first in the open-source category for coding capabilities and ninth globally.

Yet its costs are extremely low: V4 Flash output costs only 0.28 US dollars per million tokens, roughly 1% of Claude Opus's price. Its open-source ecosystem also carries influence in global communities, backed by its positioning as "China's OpenAI" — a combination of advantages rarely seen domestically.

To reach a trillion-yuan valuation, what extreme efforts does DeepSeek still need to make?

A massive "Mariana Trench" lies between DeepSeek's 500 million US dollars in annualized revenue and its 71 billion US dollar valuation.

DeepSeek's current revenue level only proves the "potential" of its business model to generate profits, but it does not yet demonstrate the "capability" to achieve rapid, large-scale growth in the future.

Rough estimates suggest that to reach a valuation of one trillion yuan, DeepSeek's revenue may need to increase at least tenfold to hit the 5-billion-US-dollar mark, which requires DeepSeek to transition from "technology experimentation" to "enterprise essential demand."

This sounds like a fantasy, but Anthropic grew its annualized revenue from 9 billion US dollars (December 2025) to 47 billion US dollars in May 2026 in just five months, multiplying more than five times. DeepSeek is fully capable of similar growth.

Therefore, the focus of our imagination is not "when DeepSeek will reach a trillion-yuan valuation," but "how it can achieve a trillion-yuan valuation."

It is widely believed that to cross this chasm, DeepSeek needs to undergo further in-depth transformation. Judging from its recent actions, Liang Wenfeng recognized this long ago.

Currently, DeepSeek's revenue mainly comes from API calls, with the V4 API boasting a gross profit margin exceeding 50% — slightly lower than that of the liquor industry, which is not a particularly remarkable story.

Additionally, APIs are standardized commodities that cannot withstand market price wars and customer poaching, and they cannot build a moat that guarantees DeepSeek's long-term stability.

To support a trillion-yuan valuation, DeepSeek is expected to upgrade from selling tokens (selling raw materials) to "selling more advanced solutions" — finished services that can directly solve specific business problems and create tangible commercial value.

As Zhou Bowen, Director and Chief Scientist of the Shanghai AI Laboratory, stated, "The next step for the AI industry is to find monetization opportunities in high-value scenarios."

So what is the high-value scenario transformation that Liang Wenfeng is pursuing?

Reports indicate that DeepSeek's new round of financing will fund self-built data centers and in-house developed AI chips. Some observers believe it is transforming from a "model-selling company" to an "AI infrastructure operator."

Why is the latter more promising? Because "selling models" is product-focused, earning profit margins, while "operating infrastructure" is platform-focused, earning "taxes."

For example, selling models and building products is like running a Michelin-starred restaurant that needs to launch new dishes regularly. Customers will still leave if they find a more cost-effective restaurant — this is the essence of API competition.

An AI infrastructure operator, however, is like building a highway, collecting tolls while operating gas stations, restaurants, and hotels at service areas. This is believed to create three revenue engines: selling computing power, selling platforms, and selling ecosystems, generating continuous, stable, and nearly unlimited "taxes." Clearly, the latter brings far greater returns.

If DeepSeek successfully completes this transformation and the market accepts this narrative, its valuation model will no longer be benchmarked against software companies, but against "Amazon AWS in the computing power era."

As we all know, after Amazon transformed from an e-commerce business to an AWS cloud service provider, it did not see a gentle growth curve, but an exponentially steep growth curve driven by a new business model.

Today, Amazon is no longer regarded as a "low-margin product-selling" company, but has transformed into a "high-margin technology-selling" enterprise. Although AWS only accounts for about 18% of the company's total revenue, it contributes over 60% of its operating profit, clearly highlighting AWS's core role as Amazon's "profit engine."

In comparison, Moutai's moat is the compounding of time, including its brand, craftsmanship, and geographical indication, which cannot be replicated for a century. For AI companies, their moat lies in technological generational gaps and ecosystem lock-in effects. Whether DeepSeek can catch up to Moutai depends not on who makes more money today, but on who can build an "irreplaceable" barrier like Moutai's in this decade.

At the same time, considering China's national conditions for future development, whoever can close the loop of "ten-thousand-card clusters + large model training" within the domestic computing power ecosystem amid long-term GPU supply constraints will become the "critical infrastructure" for China's national AI strategy.

If DeepSeek successfully secures this position, it will undoubtedly enhance its strategic scarcity. In the absence of existing valuation anchors in the A-share market, the market can only complete its pricing through a "premium driven by imagination."

After DeepSeek's listing, it still needs to pass three layers of market tests

There is a significant gap in the price-to-sales ratios (revenue multiples) between DeepSeek, OpenAI, and Anthropic. Some point out that this is because the market is pricing them separately: The former is positioned as a "potential stock," while the latter two are being revalued as "blue-chip stocks with solid performance."

If DeepSeek lists in 2027, it will face institutional investors who compare it item by item with global peers. Whether this 148-times P/S ratio can be continuously validated will become a mandatory challenge for DeepSeek.

Behind this lie three core challenges for DeepSeek.

Can it maintain continuous technological leadership? Open-source models are catching up rapidly. The market will observe: Is DeepSeek's moat a generational technological gap in its next-generation model, or a cost advantage?

The answer the market craves may be DeepSeek's "engineering methodology" that continuously produces cost-effective models. This methodology will be amplified by the open-source ecosystem, solidified by domestic computing power, and eventually become a long-term strategic barrier that competitors can never catch up to in terms of cost and cannot bypass in the ecosystem.

Can profitability scale up? Currently, AI companies still face high costs for model training and inference. If losses expand while revenue grows, the capital market will lose patience.

Across the entire industry, DeepSeek's 50% gross profit margin is not the highest — Anthropic claims its direct sales business can reach a 60% gross profit margin. But DeepSeek's pricing is only a fraction of its major competitors', yet it achieves "low prices with high gross margins," demonstrating its extraordinary cost control capabilities.

Perhaps the real question for DeepSeek is not whether its gross profit margin can be even higher, but whether it can sustain this level. If gross profit margins continue to improve, its profit model will be validated; otherwise, the capital market will begin to raise doubts.

Continuous financing must be accompanied by sustained growth. In the AI industry, bubbles are the fuel for innovation. Without the capital market's "advance overdraft," AI infrastructure such as data centers could never be built in the short term.

But the flip side of the bubble is that a company's performance must outpace capital's enthusiasm. Anthropic took five years to grow from a 124-million-US-dollar seed round to its current near-trillion valuation, undergoing multiple rounds of financing (Series F, G, H) with each round backed by substantial revenue growth. Whether DeepSeek can replicate this path will be the biggest suspense in the next two to three years.

It is already clear that despite many uncertainties, DeepSeek has a core pillar that other competitors lack: its founder Liang Wenfeng.

According to Bloomberg's estimates, in DeepSeek's equity structure, Liang Wenfeng holds 78% of the shares and controls pricing power. With the support of external industrial capital, DeepSeek has formed a unique model of "Founder's large stake anchoring + Refusing dilution + Industrial capital support."

In contrast, Zhipu's model can be summarized as "Tsinghua's technology foundation + Industrial capital alliance + State-owned capital strategic backing," which is completely different from DeepSeek's approach of "founder absolute control, extremely cautious about equity dilution." Zhipu more closely resembles a carefully constructed "capital community" that binds the interests of multiple parties, with actual control locked in through a complex concert party agreement.

If Zhipu is a "federated entity," DeepSeek is a company led by a "spiritual and practical leader" with extremely strong execution capabilities — naturally, it also carries corresponding risks.

This means DeepSeek's entire development pace is fully controlled by its founder. He can secure valuations at the most favorable time and list during the best possible window. He determines all of DeepSeek's development directions and can even prevent the company from experiencing stock price volatility like Zhipu's after listing.

In essence, discussing DeepSeek's path to a trillion-yuan valuation is discussing how much capital will value the grand proposition of China's independent and controllable AI development.

Industrial trends, capital preferences, and national strategies will jointly determine the final price. At the table of China's AI competition in the next two to three years, whether DeepSeek can remain at the center will test its true capabilities.

This article is from the WeChat official account "Muhe's Tech Press Conference", written by Gong Zheng, and published with authorization from 36Kr.