OpenAI向左，DeepSeek向右

Author: Sleepy.md

On April 24, 2026, DeepSeek V4 Preview Edition was officially released.

This domestically developed large model, with 1.6 trillion parameters in the Pro version and 284 billion parameters in the Flash version, has focused its core selling points on the market, with a million-context window becoming the standard free feature for all official services. Almost simultaneously, across the ocean, OpenAI also launched GPT-5.5, which has even greater computing power and more extensive Agent capabilities, but at a much higher price.

“Million-context” translated into plain language means AI is no longer a “goldfish” that can only remember your last few sentences, but a “superbrain” capable of swallowing three volumes of “The Three-Body Problem” at once, understanding a two-hour movie in a second, and even helping you spot typos.

For example, you can throw all your company’s contracts, emails, and financial reports from the past three years into V4, and it will help you find the breach clause hidden in the attachment on page 47. In the past, this task required a team of lawyers; now, it’s free.

GPT-5.5 prices this kind of superbrain openly: the standard version charges $5 per million input tokens and $30 for output; while the high-end GPT-5.5 Pro version is sold at a staggering $30 per million input tokens and $180 for output.

However, according to DeepSeek’s official pricing, cached inputs for V4-Flash cost only 0.2 RMB per million tokens, with outputs costing 2 RMB; even the V4-Pro, comparable to top-tier closed-source models, has cached input costs of 1 RMB, non-cached input costs of 12 RMB, and output prices of 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 become a divergence of business models.

OpenAI once was the dragon-slaying youth shouting “benefiting all humanity,” but now it sells expensive, premium apartments; meanwhile, DeepSeek is turning near-free computing power into water, electricity, and coal.

When OpenAI becomes a shrewd contractor, why does DeepSeek spend at almost no cost to turn top-tier AI into free tap water? What hidden currents are behind this shift in pricing power?

Ulanqab Cold Wind

The decisive battle of large models takes place in a data center in Inner Mongolia, where the temperature drops below -20°C.

Just before the release of V4, DeepSeek’s job postings unexpectedly added a position: Senior Data Center Delivery Manager and Senior Operations Engineer, with a maximum monthly salary of 30k RMB, 14 months’ pay, stationed in Ulanqab, Inner Mongolia.

This was once a lean company that championed “minimalism, purity, and doing only algorithms.” Over the past two years, their proudest label was “using a few pounds to move a thousand pounds,” with less than $6 million in training costs, creating DeepSeek-R1, which caused a plunge in the US stock AI sector.

But the enormous computing demands of V4, combined with the tightening US export restrictions on AI chips, shattered this light-asset idyll.

In 2025, the US Department of Commerce further tightened export controls on AI chips to China; NVIDIA’s H100 and H800 are already cut off, and even the downgraded H20 has been added to the control list. This means DeepSeek’s future expansion of computing power must fully shift to Huawei’s Ascend ecosystem. The V4 release notes explicitly state that the new model is “powered by Huawei Ascend,” and it was also revealed that after the Ascend 950 super-node batch launch in the second half of the year, the Pro version’s price will be significantly reduced.

This shift isn’t something that can be achieved by changing a few lines of code or adaptation layers; it requires starting from scratch, building a complete domestic computing infrastructure at the physical level.

V4’s trillion-parameter scale (with pretraining data reaching 33 trillion tokens), plus the massive computational needs of a million-context window, means you need thousands of Ascend chips, data centers capable of housing these chips, power grids supplying these data centers, and a maintenance team to keep these machines running smoothly in -20°C cold.

Liang Wenfeng has taken his methodology from the bit world into the atomic world. Computing power ultimately takes root in reinforced concrete and power lines.

On one side are AI elites in Silicon Valley, coding in plaid shirts, sipping pour-over coffee; on the other side are maintenance personnel wrapped in military coats, guarding data centers deep in the Inner Mongolian grasslands. This disparity forms the underlying tone of China’s resistance to the US’s compute power blockade. The cold wind of Ulanqab has become China’s strongest physical external support for AI.

Transforming from a pure algorithm company into a “heavy asset” player building its own data centers means DeepSeek has bid farewell to the guerrilla era of “small efforts, big miracles,” and has donned the armor of heavy infantry. This transformation comes at a huge cost—building data centers, buying chips, laying cables—each a bottomless pit. More importantly, this heavy-asset model causes operational costs to rise exponentially, while DeepSeek’s commercial revenue remains extremely limited. This pricing strategy is essentially sacrificing profit to build an ecosystem, and offering free infrastructure to gain discourse power.

How long can a tough guy who once refused all giants and relied on quantitative trading to subsidize AI with his own money hold out in this bottomless pit?

$20 Billion Compromise

In April, news broke that DeepSeek was initiating its first external funding round, with a target valuation of up to 300 billion RMB (about $44 billion), planning to raise 50 billion RMB, including 30 billion RMB from outside investors. Rumors of Tencent and Alibaba competing to invest are rampant.

Many believe this is because building data centers is too expensive. But in fact, the core driver of DeepSeek’s fundraising isn’t just buying GPUs; it’s also driven by “pure technological ideals”—which are no match for the talent war among giants.

During the critical sprint of V4 development, major domestic companies launched aggressive targeted headhunting of DeepSeek’s key researchers. Since late 2025, at least five core members have left. The primary author of the first-generation model, Wang Bingxuan, went to Tencent; core contributor Luo Fuli was recruited by Lei Jun with a 10-million-yuan annual salary to Xiaomi; and core author Guo Daya joined ByteDance’s Seed team.

This is the most naked form of market economy operation: when your competitors hold unlimited ammunition, and you insist on using your own funds to sustain operations, talent becomes your most vulnerable weakness. You can ask geniuses to accept lower salaries and work overtime for the sake of changing the world, but when big firms slap a check with tens of millions of cash and stock options on the table, promising unlimited compute resources, the pricing power of idealism is no longer in your hands.

Liang Wenfeng’s dilemma is actually a common predicament faced by every startup trying to build a “slow company” in China. In a market where big firms can buy anyone with money, the route of “no fundraising, no commercialization, just focusing on technology” is extremely luxurious. Its cost is that you must accept that your team could be cleared out at any moment by your competitors with cash.

This 300 billion valuation fundraising isn’t Liang Wenfeng’s capitulation to capital; it’s a human rescue war to preserve the V4 R&D team. He must sit at the capital’s table, using the same real money to give the remaining team members enough reason to stay.

The possible entry of Tencent and Alibaba means DeepSeek is no longer the lonely, purely idealistic tech startup. It has become a company with external shareholders and commercial pressures. The cost of this transformation is that the “research freedom” Liang Wenfeng once prided himself on—free from external interference—will inevitably be diluted.

But he had no choice.

When idealism is forced to wear the armor of capital, where does the confidence to keep this massive machine running, and to sustain the Ulanqab data center’s day-and-night hum, come from?

Another Kind of “Big Power, Big Miracle”

The answer isn’t in algorithms, but in the power grid.

Silicon Valley’s current anxiety isn’t about a lack of chips, but about insufficient electricity. Musk is building massive data centers in Memphis, Tennessee; OpenAI is even discussing investing in nuclear power plants; Microsoft announced restarting the Three Mile Island nuclear plant in Pennsylvania to power AI data centers. Power, the limit of compute, is a cold, fundamental physical law.

In the US, the electricity consumption of a large AI data center is equivalent to that of a medium-sized city’s daily usage. Yet, the US power grid is an aging network built in the 1950s—slow to expand, regionally fragmented, unable to keep pace with the rapid expansion of AI’s compute demands.

Supporting China’s AI race against the US are not only those algorithm geniuses earning tens of millions annually but also the silent ultra-high-voltage transmission lines.

The reason the Ulanqab data center can rise from the ground is due to Inner Mongolia’s abundant green electricity and China’s top-ranked power grid dispatching capabilities. Public data shows Ulanqab’s green energy capacity reaches 19.4 GW, accounting for about 65.9%, with local low-cost green electricity about 50% cheaper than in eastern regions. Plus, with an average annual temperature of only 4.3°C and nearly 10 months of natural cooling, energy savings of 20% to 30% are achievable.

When DeepSeek V4 runs, what truly feeds it is China’s vast and extremely cheap electrical infrastructure. This is another dimension of “big power, big miracles.”

Here’s a stark and fascinating historical contrast: In 1986, the US used the “US-Japan Semiconductor Agreement” to crush Japan’s semiconductor industry, forcing Japan to open markets and accept price controls. Japan’s share of the global semiconductor market fell from 40% in 1986 to 15% in 2011. Japan took thirty years to recover.

Today, the US tries to use similar logic to lock down China’s AI—blocking chips, restricting compute power, cutting off supply chains. But China’s counterattack path is entirely different. Japan’s failure was due to its semiconductor industry’s heavy reliance on US technology licensing and market access; once cut off, it lost its independence. China’s AI counterattack starts from the most fundamental physical infrastructure—making chips domestically, building its own data centers, laying its own power grids, and open-sourcing models.

This is a heavy, costly, yet extremely resilient route—difficult to “strangle.” While Silicon Valley builds magnificent Babel towers in the cloud, China digs trenches in the soil.

If cloud compute wars are a brutal, heavy-asset consumption battle, beyond building data centers in Inner Mongolia and laying cables, is there another way to escape cloud dominance?

Escape from the Cloud

As Silicon Valley giants build ever larger data centers, even planning trillion-dollar-scale compute clusters like OpenAI, China’s counterattack has quietly shifted underground.

The ultimate weapon against US compute blockade isn’t creating chips stronger than H100, but embedding large models into every person’s smartphone.

Since we can’t outgun the heavy firepower in cloud data centers, let’s bring the battlefield back to 1.4 billion smartphones and edge devices. This is a classic guerrilla tactic, and a highly resistant one—export bans on high-end GPUs can be imposed, but you can’t confiscate every phone in China.

By 2026, driven by the compute anxiety sparked by DeepSeek, Chinese smartphone manufacturers Xiaomi, OPPO, and vivo have launched a frantic “edge-side shift.” They are no longer satisfied with just using phones as displays for cloud API calls; through extreme model distillation and compression, they are squeezing a mini superbrain into a few thousand-yuan domestic phone.

The core of this tech route is “distillation.” Simply put, it involves training a small model (“student”) to mimic the “thinking” of a large model (“teacher”), so the small model learns the teacher’s reasoning rather than memorizing all its “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, capable of running smoothly on a single phone chip.

Apps like MNN Chat for mobile AI already enable users to run DeepSeek R1 distilled models locally on their phones. The significance of this edge AI is that you don’t need a constant 5G connection, nor do you have to pay $100 monthly to Silicon Valley giants. The large model is in your pocket, running offline, without spending a penny on cloud compute.

Since I can’t build a centralized super boiler for heating, I’ll give each household a small stove.

Of course, edge AI isn’t perfect. Limited by the computing power and memory of mobile chips, the capabilities of edge models are far below those of large cloud models. They can help you write emails, translate texts, summarize articles, but if you want them to derive complex mathematical theorems or analyze hundreds of pages of legal contracts, they still fall short.

But that’s enough. For most ordinary people, the AI they need isn’t a superbrain capable of deriving mathematical theorems, but a “personal assistant” that helps handle daily chores.

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

Digital Equality in the Global South

If you sit in a Manhattan glass office with a panoramic view, you might think that GPT-5.5’s price hike to $100 is worth it because it can help you write a perfect M&A report in a second.

But if you stand in a cornfield in Uganda, East Africa, facing yellowing crops due to climate anomalies, no one can afford the $100 subscription fee, because Uganda’s per capita monthly income is less than $150.

Silicon Valley giants discuss how to dominate the world with AI, but farmers in Uganda and poor students in Southeast Asia, thanks to DeepSeek’s open source, are stepping into the digital age for the first time.

GPT-5.5 serves those who can pay, and its training data is almost entirely in English. If you ask it questions in Swahili or Javanese, it answers haltingly, and consumes several times more tokens than in English. Due to “low commercial return,” Silicon Valley giants have voluntarily abandoned these marginal markets.

Meanwhile, China’s open-source models have become the digital infrastructure for the Global South.

In Uganda, local NGO Sunbird AI used a Chinese open-source model, Qwen, fine-tuned into the Sunflower system, expanding local language support from 6 to 31 languages. This system is now deployed in Uganda’s agricultural extension services, sending planting advice to farmers in Swahili.

In Malaysia, tech companies fine-tuned open-source bases into AI models compliant with Islamic law, supporting Malay and Indonesian, and ensuring outputs meet religious and cultural standards for Muslim markets. From Indonesia’s digital ID systems to Kenya’s Swahili medical Q&A, Chinese technology is penetrating the social fabric of these countries.

According to early 2026 data from OpenRouter, the largest AI model API aggregator platform, Chinese AI models’ token consumption on the platform has surpassed that of US competitors for the first time. In a certain week, the top 10 models worldwide consumed 87 trillion tokens, with Chinese models accounting for about 61%.

Open source has broken the US’s monopoly on AI discourse, allowing resource-scarce developing countries to leap over the digital divide. This isn’t a grand narrative of US-China rivalry; it’s the real “rural encircles the city” in the AI era.

China’s open-source AI strategy is objectively becoming an extremely effective form of “soft power” export. As Silicon Valley giants build high walls in the cloud, trying to become the new digital landlords, those “tech refugees” who cannot afford the rent finally find their own spark in open source and edge computing.

Tap Water

Technology should never be a luxury reserved for the elite.

Silicon Valley has built exquisite, gated luxury apartments, open only to VIPs. But we have laid a pipeline to thousands of households.

This pipeline starts in data centers in Inner Mongolia’s -20°C cold, amid the roar of ultra-high-voltage transmission lines, in the midst of the trillion-dollar war. Every segment is heavy, expensive, filled with forced compromises. Liang Wenfeng once wanted to create a purely technical company, but reality pushed him to build data centers, seek funding, and compete with giants for talent. He had no choice—because he chose a harder path: not to make AI a luxury, but to turn it into tap water.

And the endpoint of this pipeline is in a few-thousand-yuan domestic smartphone, in the rough fingers of Ugandan farmers, in the lives of ordinary people eager to cross the digital divide.

No matter how high the walls of compute power are built, they cannot stop the flow of tap water to the lowlands.