OpenAI向左，DeepSeek向右

Original Author: Sleepy.md

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

This domestic large model, featuring a Pro version with 1.6 trillion parameters and a Flash version with 284 billion parameters, threw its core selling point onto the market: a million-token context window becomes a free standard feature for all official services. Almost simultaneously, OpenAI across the ocean released GPT-5.5, which boasts greater computational power and richer agent capabilities, but comes at a much higher price.

Translating "million-token context" into plain language means AI is no longer a "goldfish" that can only remember your last few sentences, 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 along the way.

To give the most direct example, you can throw all your company's contracts, emails, and financial reports from the past three years into V4 and have it find the breach of contract clause hidden in the appendix on page 47. In the past, this required a team of lawyers; 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 for output; its GPT-5.5 Pro version, aimed at high-end tasks, sells at a staggering price of $30 per million input tokens and $180 for output.

But according to DeepSeek's official pricing, V4-Flash costs only 0.2 RMB per million tokens for cached input and 2 RMB for output. Even for the V4-Pro, comparable to top-tier closed-source models, the cached input price is 1 RMB, non-cached input is 12 RMB, and the output price is only 24 RMB.

People always think the US-China AI competition is a race of model capabilities. In reality, it has long become a divergence of business models.

OpenAI, once the dragon-slaying youth shouting "benefit all humanity," is now selling expensive, finely decorated commercial apartments. Meanwhile, DeepSeek is turning AI into water, electricity, and gas with nearly free computing power.

When OpenAI became a shrewd contractor, why is DeepSeek turning top-tier AI into free running water at seemingly any cost? What undercurrents lie hidden behind this shift in pricing power?

The Cold Wind of Ulanqab

The decisive battle for large language models takes place in server rooms facing minus 20 degrees Celsius in Inner Mongolia.

Not long before the V4 release, DeepSeek's job postings included a surprising position: Senior Delivery Manager and Senior Operations Engineer for Data Centers, with a monthly salary of up to 30,000 RMB, 14-month pay, stationed in Ulanqab, Inner Mongolia.

This was a company that once championed the "minimalist, pure, algorithms-only" asset-light model. Over the past two years, their proudest label was "using four ounces to move a thousand pounds," creating DeepSeek-R1 with a training cost of less than $6 million, causing the US AI stock market to plummet.

But the massive computational demands of V4, combined with increasingly tight US chip restrictions, completely shattered this pastoral idyll of asset-light operations.

In 2025, the US Commerce Department further tightened export controls on AI chips to China. Nvidia's H100 and H800 were already cut off, and even the downgraded H20 was added to the control list. This means DeepSeek's future computational expansion must fully pivot towards the Huawei Ascend ecosystem. In the V4 release notes, the company explicitly stated the new model was "supported by Huawei Ascend" and hinted that after the mass availability of Ascend 950 supernodes in the second half of the year, the Pro price would decrease significantly.

This pivot isn't just a matter of changing a few lines of adaptation code. It requires building a complete set of domestic computing infrastructure from scratch at the physical level.

V4's trillion-parameter scale (with pre-training data reaching 33 trillion tokens), coupled with the massive computational requirements of the million-token context, requires thousands upon thousands of Ascend chips, server rooms to house them, power grids to supply these rooms, and an operations team to keep the machines running in the minus 20-degree wind.

Liang Wenfeng moved his methodology from the world of bits to the world of atoms. Computing power must ultimately take root in steel, concrete, and power lines.

On one side, you have AI elites in plaid shirts coding in Silicon Valley, sipping pour-over coffee. On the other side, you have operations personnel bundled in military coats guarding server rooms deep in the Inner Mongolian grasslands. This contrast forms the backdrop of today's Chinese AI resistance against computing power blockades. The cold wind of Ulanqab has become China's strongest physical weapon.

Transitioning from a pure algorithm company to an "asset-heavy" player building its own server rooms means DeepSeek has bid farewell to the guerrilla warfare era of "small force, big miracle" and officially donned the armor of heavy infantry. The cost of this transition is immense: building server rooms, buying chips, laying network cables – every item is a money pit. More importantly, this asset-heavy model means operational costs will rise exponentially, while DeepSeek's commercial revenue remains extremely limited. This pricing strategy is essentially trading losses for an ecosystem, and using free services for leverage within the infrastructure landscape.

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

The $20 Billion Compromise

In April, news emerged that DeepSeek was launching its first external financing round, aiming for a valuation of 300 billion RMB (approximately $44 billion), planning a capital increase of 50 billion RMB, with 30 billion sought from external investors. Rumors swirled that Tencent and Alibaba were vying for a stake.

Many thought this was simply because building server rooms was too expensive. But in reality, the core driver for DeepSeek's financing, besides buying graphics cards, was that "pure technological idealism" proved fragile in the face of the talent wars waged by tech giants.

During the critical sprint phase of V4 development, domestic tech giants launched an aggressive poaching campaign against DeepSeek. From the second half of 2025 to the present, at least five core R&D members confirmed their departure. Wang Bingxuan, a core author of the first-generation model, went to Tencent. Luo Fuli, a core contributor to V3, was poached by Lei Jun to Xiaomi for an annual salary of 10 million RMB. Guo Daya, a core author of R1, joined ByteDance's Seed team.

This is the most naked operation of a market economy. When your competitors wield unlimited ammunition and you insist on operating with your own funds, the talent market becomes your weakest flank. You can ask geniuses to work overtime for less pay for the ideal of changing the world, but when a giant slaps a check with tens of millions in cash and stock options on the table, promising unlimited computational resources, the pricing power of idealism is no longer in your hands.

Liang Wenfeng's dilemma is the same one every entrepreneur trying to build a "slow company" in China faces. In a market where giants can buy anyone with money, the path of "no financing, no commercialization, just technology" is extremely luxurious. Its price is accepting that your team could be poached away by competitors at any moment.

This 300 billion RMB valuation financing is not Liang's compromise with capital, but a rescue mission he launched against the giants to secure the V4 development team. He had to sit at the capital's table, using the same real money to give those who stayed sufficient reason to continue staying.

The potential involvement of Tencent and Alibaba means DeepSeek is no longer that lonely, pure technological idealist. It has become a company with external shareholders and the pressure of commercialization. The cost of this change is the inevitable dilution of that "research freedom free from external pressure" that Liang was most proud of.

But he had no choice.

When idealism is forced to don the armor of capital, where does the confidence come from to keep this massive machine running, to keep the server rooms in Ulanqab humming day and night?

Another Kind of "Power Creates Miracles"

The answer is not in the algorithms, but in the power grid.

What Silicon Valley is most anxious about now isn't a chip shortage, but an electricity shortage. Musk is frantically building super data centers in Memphis, Tennessee. OpenAI has even started discussing investing in nuclear power plants. Microsoft has announced the restart of the Three Mile Island nuclear plant in Pennsylvania to power AI data centers. The end of computing power is electricity – a stark, cold physical reality.

In the US, a large AI data center's electricity consumption equals that of a medium-sized city. And the US power grid, built in the 1950s, is an aging network with slow expansion and regional fragmentation, unable to keep up with the AI era's demand for computing power expansion.

What supports China's AI in catching up with the US isn't just algorithm geniuses with multi-million dollar salaries, but also the unheralded ultra-high voltage (UHV) transmission lines.

The data center in Ulanqab could be built thanks to Inner Mongolia's abundant green electricity and China's world-leading grid dispatch capabilities. Public data shows Ulanqab has a green electricity installed capacity of 19,402 MW, accounting for about 65.9%. Local low-cost green electricity is roughly 50% cheaper than in eastern regions. Combined with an average annual temperature of only 4.3°C, providing nearly 10 months of natural free cooling, equipment energy consumption is reduced by 20% to 30%.

When DeepSeek V4 runs, its true lifeblood is China's vast and incredibly cheap electricity infrastructure. This is another dimension of "power creates miracles."

There's an incredibly interesting and harsh historical parallel here. In 1986, the US used the US-Japan Semiconductor Agreement to cripple Japan's semiconductor industry, forcing Japan to open its markets and accept price controls. Japan's global semiconductor market share plummeted from 40% in 1986 to 15% in 2011. Japan hasn't recovered in thirty years.

Today, the US is trying to lock down China's AI with the same logic, blocking chips, restricting computing power, cutting technology supply chains. But China's counterattack path is completely different from Japan's. Japan's failure lay in its semiconductor industry's heavy reliance on US technology licensing and market access. Once cut off, it lost the ability to survive independently. China's AI counterattack, however, starts from rebuilding the most fundamental physical infrastructure: making its own chips, building its own server rooms, laying its own power grids, and open-sourcing its own models.

This is an extremely heavy, extremely costly, but also extremely difficult path to "strangle." While Silicon Valley builds ornate towers of Babel in the cloud, China digs trenches in the dirt.

If the cloud computing power struggle is a brutal, asset-heavy war of attrition, is there another way to escape the hegemony of the cloud besides building server rooms in Inner Mongolia and laying power lines?

Escaping the Cloud

While Silicon Valley giants build ever-larger data centers, some like OpenAI planning trillion-dollar computing clusters, China's counterattack line has quietly shifted underground.

The ultimate weapon against the US computing blockade isn't necessarily making a chip more powerful than the H100, but stuffing the large model into everyone's phone.

Since we can't win against heavy firepower in the cloud server rooms, we pull the battlefield back to 1.4 billion smartphones and edge devices. This is a classic guerrilla tactic, and one extremely difficult to blockade. You can ban the export of high-end GPUs, but you can't confiscate the phone in every Chinese person's pocket.

In 2026, driven by the computing anxiety triggered by DeepSeek, Chinese phone makers like Xiaomi, OPPO, and vivo started a frenzied "migration to the edge." They are no longer satisfied with just using the phone as a display calling cloud APIs. Through extreme model distillation and compression, they are forcibly squeezing a shrunk-down super brain into Chinese phones costing a few thousand RMB.

The core of this technical path is "distillation." Simply put, it uses a super-large model (teacher) to train a smaller model (student), teaching the student the teacher's "way of thinking" rather than memorizing all the teacher's "knowledge." Through extreme distillation and quantization compression, a large model originally requiring hundreds of GPUs is compressed to just 1.2GB to 2.5GB, running smoothly on a single mobile phone chip.

Mobile AI applications like MNN Chat already allow users to run the Distilled DeepSeek R1 model locally on their phones. The significance of this edge AI is that you don't need to be constantly connected to a 5G signal or pay a $100 monthly subscription fee to Silicon Valley giants. The large model is in your pocket, works offline, and doesn't cost a penny for cloud computing power.

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

Of course, edge AI isn't perfect. Limited by the computing power and memory of phone chips, the upper limit of edge model capabilities is far lower than that of cloud-based super-models. It can help you write an email, translate a paragraph, summarize an article. But if you want it to derive a complex mathematical theorem or analyze a hundred-page legal contract, it will still fall short.

But that's enough. Because for the vast majority of ordinary people, the AI they need has never been the super brain capable of deriving mathematical theorems, but a "personal assistant" that can handle their daily chores.

When large models become incredibly cheap, even pocket-sized, how will they change the corners forgotten by Silicon Valley?

Digital Equality for the Global South

If you sit in a panoramic glass office in Manhattan, you probably think GPT-5.5's price hike to $100 is worth it, because it can help you draft a perfect M&A financial report in a second.

But if you stand in a cornfield in Uganda, facing crops withered by climate anomalies, no one can afford the $100 subscription fee, because the average monthly income in Uganda is less than $150.

While Silicon Valley giants discuss using AI to rule the world, 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.

GPT-5.5 serves those who can pay, and its training data is almost entirely in English. If you ask it a question in Swahili or Javanese, it not only stumbles through the answer but also consumes several times the tokens compared to English. Silicon Valley giants have actively abandoned these fringe markets due to "low commercial returns."

China's open-source models have become the digital infrastructure for the Global South.

In Uganda, the local NGO Sunbird AI used the Sunflower system, fine-tuned from the Chinese open-source model Qwen, to expand the number of supported local languages from 6 to 31. This system is now deployed within the Ugandan government's agricultural extension system, sending planting advice to farmers in Swahili.

In Malaysia, a tech company used an open-source base model to fine-tune an AI model compliant with Islamic law. It not only supports Malay and Indonesian but also ensures output meets the religious and cultural standards of the Muslim market. From Indonesia's digital identity system to Swahili medical Q&A in Kenya, Chinese technology is penetrating the foundational 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' token consumption on the platform surpassed US competitors for the first time. In a certain statistical week, the world's top 10 hottest models consumed 8.7 trillion tokens, with Chinese models accounting for approximately 61%.

Open source has broken the US monopoly on AI discourse, allowing resource-poor developing countries to cross the digital divide. This isn't some grand narrative of US-China hegemony; this is the real "using the countryside to encircle the cities" of the AI era.

China's open-source AI strategy is objectively becoming an incredibly effective form of "soft power" export. As Silicon Valley giants build high walls in the cloud, trying to become the new digital landlords, the "tech refugees" who can't afford the rent have finally found their spark in the dirt of open source and edge computing.

Tap Water

Technology was never meant to be an unattainable luxury.

Silicon Valley built exquisitely crafted commercial apartments, heavily guarded, open only to VIPs. But we laid a tap water pipe leading to millions of households.

The starting point of this pipe is in a server room facing minus 20 degrees Celsius in Inner Mongolia, amidst the hum of ultra-high voltage power lines, within a war valued at 300 billion RMB. Every segment is heavy, expensive, and filled with pressure and compromise. Liang Wenfeng once wanted to build