Author: Sleepy, Rhythm

On April 24, 2026, DeepSeek V4 preview version was officially released.

This domestically developed large-scale model, featuring a Pro version with 1.6 trillion parameters and a Flash version with 284 billion parameters, focused its core selling point on the market: millions of contextual parameters became a free standard feature in all official services. Around the same time, OpenAI across the ocean also released GPT-5.5, boasting even greater computing power and richer agent functionality, but at a significantly higher price.

In layman's terms, "millions of contexts" means that AI is no longer a "goldfish" that can only remember the first few sentences you say, but has become a "super brain" that can swallow three volumes of "The Three-Body Problem" in one go, understand a two-hour movie in a second, and even help you pick out typos.

To give a straightforward example, you can dump all your company's contracts, emails, and financial statements from the past three years onto V4 and let it find the breach of contract clause hidden in the appendix on page 47. In the past, this would have required a legal team; now, it's free.

GPT-5.5 puts a price tag on this super brain: the standard version costs $5 per million input tokens and $30 per million output tokens; while the GPT-5.5 Pro version, designed for advanced tasks, sells for an even higher price of $30 per million input tokens and $180 per million output tokens.

However, according to DeepSeek's official pricing, V4-Flash's input with a cache hit costs only 0.2 RMB per million tokens, while its output costs 2 RMB. Even V4-Pro, which rivals top-tier closed-source models, costs 1 RMB for a cache hit and 12 RMB for a cache miss, while its output is only 24 RMB.

People often think that the AI ​​competition between China and the US is a race of model capabilities, but in reality, it has long since become a divergence in business models.

OpenAI was once the dragon-slaying hero who shouted "to benefit all mankind," but now it's selling expensive, fully-furnished apartments; while DeepSeek is using almost free computing power to turn AI into utilities like water, electricity, and gas.

When OpenAI becomes a shrewd contractor, why is DeepSeek willing to turn top-tier AI into free tap water regardless of cost? What undercurrents are hidden behind this shift in pricing power?

The cold wind in Ulanqab

The decisive game of the large-scale model took place in a computer room in Inner Mongolia at a temperature of -20 degrees Celsius.

Shortly before the release of V4, DeepSeek added an unexpected position to its job postings: Senior Data Center Delivery Manager and Senior Operations Engineer, with a maximum monthly salary of 30,000 yuan, 14 months' salary, and on-site location in Ulanqab, Inner Mongolia.

This is a company that once touted itself as "minimalist, pure, and focused solely on algorithms" and was known for its asset-light approach. Over the past two years, their proudest label has been "achieving great results with minimal effort," using less than $6 million in training costs to produce DeepSeek-R1, which caused a sharp drop in the US AI stock market.

However, V4's massive computing power requirements, coupled with the increasingly stringent computing power blockade imposed by the United States, completely shattered this idyllic, asset-light lifestyle.

In 2025, the U.S. Department of Commerce further tightened export controls on AI chips to China, cutting off the supply of Nvidia's H100 and H800 chips, and even adding the downgraded H20 to the control list. This means that DeepSeek's future computing power expansion must fully shift to the Huawei Ascend ecosystem. In the V4 release notes, the official statement explicitly stated that the new model was "supported by Huawei Ascend," and revealed that after the mass production and launch of the Ascend 950 supernode in the second half of the year, the price of the Pro version will be significantly reduced.

This shift cannot be accomplished by simply changing a few lines of code in the adaptation layer; it requires building a complete domestic computing power infrastructure from scratch at the physical level.

The trillion-parameter scale of V4 (with pre-training data reaching 33 trillion tokens), coupled with the massive computational requirements of millions of contexts, means that you need tens of thousands of Ascend chips, data centers capable of accommodating these chips, power grids to supply these data centers, and operations and maintenance teams to keep these machines running in sub-zero temperatures of -20 degrees Celsius.

Liang Wenfeng has taken his methodology from the world of bits to the world of atoms. Computing power, in the end, must take root in reinforced concrete and power transmission lines.

On one side are AI elites in plaid shirts coding and sipping hand-drip coffee in Silicon Valley; on the other are maintenance personnel wrapped in military overcoats guarding data centers deep in the Inner Mongolian grasslands. This contrast forms the backdrop for China's AI resistance to computing power blockades today. The cold winds of Ulanqab have become China's strongest physical advantage against AI.

The transformation from a pure algorithm company to a "heavy-asset" player with its own data centers signifies that DeepSeek has bid farewell to its guerrilla warfare era of "small-scale miracles" and officially donned the armor of heavy infantry. This transformation comes at a huge cost: building data centers, buying chips, and laying network cables—each a bottomless pit. More importantly, this heavy-asset model means operating costs will rise exponentially, while DeepSeek's commercial revenue remains extremely limited. This pricing strategy is essentially trading losses for an ecosystem, and offering free services for greater control over infrastructure.

How long can this tough guy, who once rejected all the giants and subsidized AI with his own money through quantitative trading, hold on in the face of this bottomless pit?

$20 billion compromise

In April, DeepSeek announced its first external funding round, targeting a valuation of 300 billion RMB (approximately 44 billion USD), with plans to raise 50 billion RMB, including 30 billion RMB from external sources. Rumors swirled that Tencent and Alibaba were vying to enter the market.

Many people assume that DeepSeek's funding is due to the high cost of building data centers. However, the core driving force behind DeepSeek's financing, besides purchasing graphics cards, is its "pure technological ideals," which proved vulnerable in the face of the giants' talent-grabbing machine.

During the critical sprint of V4 development, major domestic companies launched a frenzied targeted poaching campaign against DeepSeek. From the second half of 2025 to the present, at least five core R&D members of DeepSeek have confirmed their departure. Wang Bingxuan, the core author of the first-generation model, went to Tencent; Luo Fuli, a core contributor to V3, was poached by Lei Jun to Xiaomi with a multi-million yuan annual salary; and Guo Daya, the core author of R1, joined ByteDance's Seed team.

This is the most naked way a market economy operates: when your competitors have unlimited resources while you insist on operating with your own capital, the talent market becomes your most vulnerable weakness. You can ask geniuses to take pay cuts and work overtime for the ideal of changing the world, but when a large company slams a check for millions in cash and stock options on the table, promising unlimited computing power, the pricing power of idealism is no longer in your hands.

Liang Wenfeng's predicament is actually a dilemma faced by every entrepreneur trying to build a "slow company" in China. In a market where large companies can buy anyone with money, the "no financing, no commercialization, just focus on technology" approach is extremely luxurious. The price is that you must accept that your team could be wiped out by competitors at any time with money.

This 300 billion valuation financing is not Liang Wenfeng's compromise with capital, but rather a redemption war he launched against major companies to preserve the V4 R&D team. He must sit at the capital table and use the same amount of real money to give those who remain a sufficient reason to stay.

The potential entry of Tencent and Alibaba means that DeepSeek is no longer the solitary, purely technological idealist it once was. It has become a company with external shareholders and commercial pressures. The cost of this transformation is that the "research freedom unaffected by external pressure" that Liang Wenfeng was once most proud of will inevitably be diluted.

But he had no choice.

When idealism is forced to don the armor of capital, where does the confidence to keep this massive machine running and the Ulanqab computer room roaring day and night come from?

Another kind of "miracle happens with great effort"

The answer isn't in the algorithm, it's in the power grid.

Silicon Valley's biggest anxiety right now isn't a lack of chips, but a lack of electricity. Musk is building a massive data center in Memphis, Tennessee, OpenAI is even discussing investing in nuclear power plants, and Microsoft has announced the restart of its Three Mile Island nuclear power plant in Pennsylvania to power its AI data centers. Computing power ultimately comes down to electricity—a stark and undeniable fact of physics.

In the United States, the electricity consumption of a large AI data center is equivalent to the daily electricity consumption of a medium-sized city. However, the US power grid is an old network built in the 1950s, which is slow to expand, regionally fragmented, and simply cannot keep up with the speed of computing power expansion in the AI ​​era.

What supports China's AI development in catching up with the United States is not only the algorithm geniuses who earn millions of dollars a year, but also the unsung heroes of ultra-high voltage power transmission lines.

The data center in Ulanqab was able to rise from the ground thanks to Inner Mongolia's abundant green electricity and China's world-leading power grid dispatching capabilities. Public data shows that Ulanqab's installed green electricity capacity reaches 19.402 million kilowatts, accounting for approximately 65.9% of its total electricity output. Locally available green electricity is about 50% cheaper than in eastern regions. In addition, the average annual temperature is only 4.3℃, with a natural cooling period of nearly 10 months, allowing equipment to save 20% to 30% on energy.

When DeepSeek V4 is running, what truly powers it is China's massive and extremely cheap power infrastructure. This is another dimension of "miracles through sheer force".

Here is a fascinating yet brutal historical contrast. In 1986, the United States crippled Japan's semiconductor industry with the U.S.-Japan Semiconductor Agreement, forcing Japan to open its market and accept price controls. Japan's global semiconductor market share plummeted from 40% in 1986 to 15% in 2011. It took Japan thirty years to recover.

Today, the US is attempting to use the same logic to stifle China's AI, blocking chips, limiting computing power, and cutting off technology supply chains. However, China's counterattack is entirely different from Japan's. Japan's failure stemmed from its semiconductor industry's heavy reliance on US technology licensing and market access; once cut off, it lost its ability to survive independently. China's AI counterattack, on the other hand, begins with rebuilding the most fundamental physical infrastructure: manufacturing its own chips, building its own data centers, establishing its own power grid, and open-sourcing its own models.

This is an extremely cumbersome, extremely expensive, but also extremely difficult route to "strangle." While Silicon Valley was building magnificent Tower of Babel in the clouds, China was digging trenches in the mud.

If the battle for computing power in the cloud is an extremely brutal war of attrition involving heavy assets, besides building server rooms and laying cables in Inner Mongolia, is there another way for us to escape cloud hegemony?

Escape from the Clouds

As Silicon Valley giants build ever larger data centers, even planning computing power clusters worth hundreds of billions of dollars like OpenAI, China's counterattack has quietly shifted underground.

The ultimate weapon against the US's blocking of computing power is not actually creating a chip more powerful than the H100, but rather stuffing a large model into everyone's mobile phone.

Since we can't compete with heavy firepower in cloud data centers, let's shift the battlefield back to 1.4 billion smartphones and edge devices. This is a typical guerrilla warfare tactic, and one that's extremely difficult to block. You can ban the export of high-end GPUs, but you can't confiscate the smartphone in every Chinese person's pocket.

In 2026, amidst the computing power anxiety triggered by DeepSeek, Chinese mobile phone manufacturers Xiaomi, OPPO, and vivo embarked on a frantic "device-side migration." They were no longer satisfied with simply using the mobile phone as a display that called cloud APIs, but instead, through extreme model distillation and compression, they squeezed a miniaturized super brain into a domestically produced mobile phone costing only a few thousand yuan.

The core of this technical approach is "distillation." Simply put, it involves using a super-large model (the teacher) to train a small model (the student), allowing the small model to learn the teacher's "thinking methods" rather than simply memorizing all the teacher's "knowledge." Through extreme distillation and quantization compression, a large model that originally required hundreds of GPUs to run is compressed to only 1.2GB to 2.5GB, allowing it to run smoothly on a single mobile chip.

Mobile AI applications like MNN Chat already allow users to run DeepSeek R1 distillation models locally on their phones. The significance of this edge AI is that you don't need a constant 5G connection, nor do you need to pay a $100 monthly subscription fee to Silicon Valley giants. The large model is in your pocket, can run even offline, and doesn't cost a penny in cloud computing power.

Since I can't afford to build a super boiler room for centralized heating, I'll just give each household a small stove.

Of course, edge AI is not perfect. Limited by the computing power and memory of mobile phone chips, the capabilities of edge models are far less than those of the massive models in the cloud. It can help you write an email, translate a text, or summarize an article, but if you want it to help you derive a complex mathematical theorem or analyze a legal contract of hundreds of pages, it will still fall short.

But that's enough. Because for the vast majority of ordinary people, what they need is not a super brain that can derive mathematical theorems, but a "personal assistant" that can help them handle daily chores.

When large models become so cheap they can fit in your pocket, how will they change those forgotten corners of Silicon Valley?

Digital equality in the Global South

If you were sitting in a panoramic glass office in Manhattan, you would most likely think that the price increase of GPT-5.5 to $100 is worthwhile, because it can help you write a perfect M&A financial report in a second.

But if you stand in a cornfield in Uganda, East Africa, facing crops withered and yellow due to abnormal climate, no one can afford the $100 subscription fee, because the average monthly income in Uganda is less than $150.

While Silicon Valley giants are discussing how to rule the world with AI, farmers in Uganda and poor students in Southeast Asia are stepping into the digital age for the first time thanks to DeepSeek's open-source nature.

GPT-5.5 caters to those who can afford it, and its corpus is almost entirely in English. If you ask it a question in Swahili or Javanese, it not only provides a halting response but also consumes several times more tokens than in English. Silicon Valley giants have deliberately abandoned these peripheral markets due to "low returns on investment."

China’s open-source model has become the digital infrastructure of the global South.

In Uganda, the local NGO Sunbird AI has expanded its support for local languages ​​from 6 to 31 using its Sunflower system, which is based on the Chinese open-source model Qwen. This system is now deployed within the Ugandan government's agricultural extension system, sending planting advice to farmers in Swahili.

In Malaysia, tech companies have fine-tuned AI models using open-source platforms to comply with Sharia law, supporting not only Malay and Indonesian languages ​​but also ensuring that output content conforms to the religious and cultural standards of the Muslim market. From Indonesia's digital identity system to Kenya's Swahili-language medical Q&A platform, Chinese technology is penetrating the underlying social structures of these countries.

Data released by OpenRouter, the world's largest AI model API aggregation platform, in early 2026 showed that Chinese AI models consumed more tokens on the platform than their American competitors for the first time. In one statistical week, the top 10 most popular models globally consumed a total of 8.7 trillion tokens, with Chinese models accounting for approximately 61%.

Open source has broken the US monopoly on AI discourse, enabling resource-poor developing countries to bridge the digital divide. This is not some grand narrative of a US-China rivalry; it is the true "rural-to-urban" strategy of the AI ​​era.

China's AI open-source strategy is objectively becoming an extremely effective form of "soft power" output. As Silicon Valley giants build high walls in the cloud, attempting to become the digital landlords of the new era, those "tech refugees" who cannot afford the rent have finally found their own spark in the soil of open source and edge computing.

tap water

Technology should never be an unattainable luxury.

Silicon Valley has built exquisitely beautiful apartments with tight security, open only to VIPs. But we've built a water pipe that reaches every household.

This water pipe begins in a server room in Inner Mongolia at -20 degrees Celsius, amidst the roar of ultra-high-voltage power lines, and in a battle for a valuation of 300 billion. Every segment of it is heavy, expensive, and filled with coercion and compromise. Liang Wenfeng once wanted to create a purely technology company, but reality forced him to build server rooms, raise funds, and compete with large companies for talent. He had no choice, because he chose a more difficult path: not to make AI a luxury, but to make it a tap water supply.

And the end of this water pipe is a domestically produced mobile phone worth several thousand yuan, in the rough fingers of Ugandan farmers, and in the lives of every ordinary person who longs to bridge the digital divide.

No matter how high the wall of computing power is built, it cannot stop the tap water flowing downhill.