Valuation $480B: DeepSeek’s revenue is revealed for the first time—its IPO could happen as early as next year

DeepSeek has finally taken the step toward going public.

According to related media reports, DeepSeek has already kicked off early preparations for an IPO, with plans to list on mainland China’s capital markets.

The company is engaging with accounting firms, investment banks, and other institutions to prepare audit reports, corporate governance, and filing materials. It may submit an application before the end of 2026 and strive to be formally listed in 2027.

Insiders say the proposed listing location discussed by DeepSeek is Shanghai, with the fastest possible timing being in the second quarter of 2027.

Another report says DeepSeek’s annual recurring revenue generated through API Token calls has reached $400 million to $500 million, with a gross margin of over 50%.

Just closed a Series A, and a Series B is coming

From late May to early June this year, DeepSeek completed its first round of external fundraising since its establishment, raising about $7 billion to $7.4 billion, with a post-money valuation exceeding $50 billion.

Just a little over a month later, DeepSeek began contacting potential investors again. The pre-money valuation discussed for the new round is about $71 billion (about 480 billion yuan), up roughly 37% from the previous round’s post-money valuation of about $52 billion. Based on the current discussions, the company may raise several more billion dollars.

An important use of DeepSeek’s new funding round is to build its own data centers, purchase more AI chips, and stock up on supply-chain resources in advance.

First, DeepSeek wants to develop AI Agents that can run autonomously. Such products clearly increase demand for compute resources.

Training a model is a relatively concentrated capital expenditure. After the model training is completed, it can be reused repeatedly. But when DeepSeek starts providing services to tens of millions of users, enterprise customers, and Agent application scenarios, compute spending turns into a cost that occurs every day.

Second, compared with relying entirely on third-party cloud vendors, building infrastructure in-house allows DeepSeek to better plan chip procurement, network, storage, and inference workloads—and also avoids being constrained by external compute supply during business peak periods.

But this is a heavy-asset business. Servers, chips, data halls, electricity, liquid cooling, and network equipment—anything requires upfront investment. It is understood that DeepSeek is developing its own AI chips to reduce reliance on Nvidia and Huawei.

In this just-completed funding round, the value brought by investors such as Tencent, CATL, and the National AI Industry Investment Fund is not only cash. Tencent has cloud computing, traffic, and application scenarios; CATL has energy, batteries, and industrial resources; and the national fund may help DeepSeek enter a broader ecosystem of domestically produced compute and policy support.

Third, the reason DeepSeek’s fundraising can never get around is retaining talent.

In April 2026, DeepSeek first sought external capital to establish a credible valuation for employee options, preventing core research personnel from being poached by other companies. At that time, some employees had already moved to ByteDance, Xiaomi, and other companies with stronger financial muscle.

After completing the funding, DeepSeek gained a market valuation of more than $50 billion. Employee options that had previously been a vague promise became computable wealth. DeepSeek also set up an employee stock ownership plan, distributing shares based on actual valuation.

But stabilizing the original team is not enough. DeepSeek has planned to expand teams covering product, data, compute clusters, and commercialization, and is preparing to double the size of part of its core teams.

Using low prices to crush the market

A background that cannot be ignored in DeepSeek’s preparation for going public at this time is Zhipu’s astonishing performance in the capital markets.

Zhipu listed on the Hong Kong Stock Exchange on January 8, 2026, at an issue price of HK$116.2 per share. On the first day of listing, the company raised about HK$4.35 billion, with a market capitalization of about HK$500-plus billion.

Over the next half year, Zhipu became one of the most sought-after AI stocks in Hong Kong shares.

On June 22, Zhipu’s share price hit a highest intraday level of HK$2,980, about 25 times the issue price. Its market cap once briefly exceeded HK$1 trillion. The company successively surpassed a number of major tech companies with far larger revenue and profit scales than itself, becoming one of the highest market-cap companies in the Hang Seng TECH Index.

After that, Zhipu’s share price saw a clear pullback. On July 15, its intraday price fluctuated sharply between HK$1,535 and HK$1,867, still around 14 times the issue price, with a total market cap of about HK$770 billion.

After Agent products such as OpenClaw exploded earlier this year, China’s average daily Token consumption rose from about 100 trillion at the end of 2025 to 140 trillion in March 2026.

Zhipu hit this exact moment.

The company disclosed that as of March 2026, its MaaS platform registered users had exceeded 4 million. Even if the API call price rose by 83% compared with the end of 2025, the market still could not get enough supply. In the first quarter of this year, the number of API calls grew by about 400% year-on-year. This is not simply “revenue contributed by higher prices,” but both prices and call volume rising at the same time.

Zhipu proved Token can be sold, and DeepSeek is preparing to “sell Token” like tap water.

In May 2026, DeepSeek V4 Flash’s monthly Token usage reached 184 trillion, surpassing GPT-4o to become the largest-usage general model globally.

DeepSeek’s pitch is cheap.

Taking output Tokens as an example, GPT-4o is priced at $2.5 per million Tokens, Claude 3.5 Sonnet is $3, while DeepSeek V4 Flash costs only $0.07—35 times less. Even when benchmarked against OpenAI’s GPT-4o mini ($0.15 per million Tokens), DeepSeek’s price is still less than half of it.

This low price is not a promotional tactic—it is DeepSeek’s core strategy.

Liang Wenfeng has a view: the essence of large models is infrastructure, and the core of infrastructure is cheapness. When the price is low enough, demand grows exponentially. As demand grows, it dilutes fixed costs, creating a positive feedback loop.

DeepSeek’s “price-for-market” strategy is not just talk.

Data disclosed by Vercel, a U.S. AI cloud platform company, shows that in April 2026, DeepSeek’s share of Token usage on the platform was still below 1%; after the V4 series was released in May, this proportion quickly rose to 17%, ranking third among all model vendors on the platform, even surpassing OpenAI. Almost all of the incremental usage came from V4-Flash and V4-Pro.

Rising from under 1% to 17% within a month indicates that developers truly choose models again because of performance and price.

Media reports say that as of April this year, DeepSeek’s latest annual recurring revenue is about $400 million to $500 million, mainly derived from API Token calls. With the release of the V4 series in May, this revenue will inevitably surge further.

CATL for the AI track

If DeepSeek successfully lists on the STAR Market, it is likely to become the most important company in the AI sector.

It would solve three problems in AI at the same time: having revenue (annualized $10k to $500 million), having growth (global #1 in usage volume), and having room for imagination (an infrastructure-level platform).

Once DeepSeek proves that Chinese companies can achieve the #1 position worldwide in a global tech race, capital will systematically reprice the entire industry chain—just like CATL back then.

In 2018, CATL went public. It took 3 years; its market cap rose from 78.6 billion yuan to a peak of 1.5 trillion yuan, nearly a 20-fold increase. More importantly, it wasn’t just one company going up—upstream lithium mines, cathode materials, anode materials, separators, electrolytes; midstream battery equipment and structural parts; downstream charging piles and energy storage—companies across the entire industry chain rose by several times to even dozens of times.

For cathode materials, there’s Rongbai Technology; for anode materials, Putailai; for separators, Enjie Shares; for electrolytes, Tianqi Materials… Pick any one, and they all saw gains of several times to more than ten times.

DeepSeek listing on the STAR Market may play the same role.

DeepSeek’s industry-chain logic is the same. For every $10 billion it raises, about $7 billion is spent on compute power—buying chips, building data centers, leasing bandwidth, and doing liquid cooling. This money doesn’t disappear out of thin air; it becomes revenue and profits for upstream companies.

Specifically across the industry chain, several directions are clear and deterministic:

AI chips: DeepSeek uses a large amount of domestically produced chips. Its success equals the strongest endorsement for domestic chips—previously it was “domestic chips can be used,” and now it is “domestic chips can support the world’s largest large-model usage.” The value of this endorsement is more effective than any advertisement.

Optical modules: AI training and inference require a large number of high-speed optical modules. In 2026, the global optical module market is expected to grow at a rate of over 75%, and domestic suppliers hold more than 70% of the global share. As DeepSeek builds data centers, it directly drives demand for optical modules.

Optical module products

Liquid cooling and IDC: The power consumption of AI servers is 5 to 10 times that of ordinary servers, and liquid cooling is shifting from “optional” to “mandatory.” In 2026, the liquid cooling market’s growth rate exceeds 150%. IDC rack rental prices in first-tier cities increased by 20% to 30% in the first half of the year, and they are still rising.

Advanced packaging: Requirements for packaging of AI chips are getting higher. The value share of advanced packaging has risen from 10% for traditional chips to 20% to 30%.

However, DeepSeek still needs to prove that it can continuously reduce intelligent production costs the same way CATL reduced battery costs; that it can continuously generate orders for chips, servers, data centers, and Agent companies the same way CATL drove the materials and equipment industries; and that while lowering prices, it can truly keep the massive Token usage on its own income statement and profit and loss statement.

If these three things can be achieved, DeepSeek could become the “CATL” of China’s AI sector.