OpenAI向左，DeepSeek向右

OpenAI Goes Left, DeepSeek Goes Right

By Dongcha Beating

Source:

Reprint: Mars Finance

On April 24, 2026, DeepSeek V4 preview officially launched.

This homegrown model comes in a Pro version with 16 trillion parameters and a Flash version with 284 billion parameters. It threw its most core selling points straight at the market—million-context—turning it into a free standard feature across all official services. Almost at the same time, across the ocean, OpenAI also rolled out GPT-5.5. It has more massive computing power and richer Agent capabilities, but it also comes with a much higher price.

“Million-context,” translated into plain terms, means AI is no longer a “goldfish” that can only remember the first few things you say; it becomes a “superbrain” 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.

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

GPT-5.5 puts this kind of superbrain up for sale at a set price. The standard version costs $5 for every million input tokens and $30 for output; while the GPT-5.5 Pro version, aimed at higher-end tasks, sells at an eye-watering $30 per million input tokens and $180 for output.

But according to DeepSeek’s official pricing, for V4-Flash, inputs with cache hits cost only 0.2 RMB per million tokens, and output costs 2 RMB. Even V4-Pro—comparable to top closed-source models—cache-hit inputs cost 1 RMB, cache-miss inputs cost 12 RMB, and output is priced at only 24 RMB.

People always assume that the China–US AI competition is a race of model capability. In reality, it has long become a split in business models.

OpenAI used to be the dragon-slaying youth shouting “benefiting all humankind,” but now it’s selling expensive, hardcover luxury apartment-style products; meanwhile DeepSeek is turning near-free computing power into water, electricity, and coal.

When OpenAI has become a shrewd contractor, why does DeepSeek turn top-tier AI into free tap water at almost no cost? What kind of hidden undercurrent lies behind this shift in pricing power?

Cold Winds in Ulanqab

The decisive battle for large models happens in a data center in Inner Mongolia at below -20°C.

Just before V4 was released, a surprising position appeared in DeepSeek’s job postings: Senior Data Center Delivery Manager and Senior Operations Engineer, with a maximum monthly salary of 30,000 RMB and 14 months’ pay, stationed in Ulanqab, Inner Mongolia.

This was once a light-asset company that marketed itself as “minimalist, pure, and only doing algorithms.” Over the past two years, their proudest label was “doing more with less,” producing DeepSeek-R1 with training costs under $6 million and causing the US-listed AI sector to plunge.

But V4’s massive computing power requirements—combined with the tightening of US computing power restrictions—shattered that light-asset pastoral fantasy.

In 2025, the US Department of Commerce further tightened export controls on AI chips to China. NVIDIA’s H100 and H800 were already cut off, and even the down-graded H20 was pulled into the controlled list. This means DeepSeek’s future expansion of computing power must fully pivot to Huawei Ascend’s ecosystem. In V4’s release notes, the official statement explicitly says the new model is “powered by Huawei Ascend,” and it also reveals that after the Ascend 950 super-node is batch-launched in the second half of the year, the Pro price will be significantly reduced.

This pivot can’t be completed by changing a few lines of code in an adaptation layer. It requires building, from scratch, a complete domestically-built computing infrastructure at the physical level.

V4’s trillion-parameter scale (pretraining data reaches 33 trillion tokens), plus the massive compute demand of million-context windows, means you need thousands of Ascend chips. You need data centers capable of housing those chips. You need the power grids to supply those data centers. And you need operations teams to keep the machines running without downtime amid cold winds below -20°C.

Liang Wenfeng took his methodology from the bit world to the atomic world. In the end, computing power must take root in reinforced concrete and transmission lines.

On one side are AI elites in plaid shirts in Silicon Valley, coding and sipping pour-over coffee. On the other side are operations personnel wrapped in military-style coats going deep into Inner Mongolia’s grasslands to guard data centers. This contrast forms the underlying tone of today’s China’s resistance to computing power lockouts. The cold winds of Ulanqab have become China’s strongest physical external “power-up” for AI.

Moving from a pure algorithm company to a “heavy-asset” player that builds its own data centers means DeepSeek bids farewell to the guerrilla era of “small efforts, big miracles” and formally dons the armor of heavy infantry. The cost of this transformation is enormous: building data centers, buying chips, pulling cables—each item is a bottomless pit. More importantly, this heavy-asset model means operating costs will rise exponentially, while DeepSeek’s commercialization revenue remains extremely limited. This pricing strategy, at its core, is using losses to build an ecosystem and using free infrastructure to secure a say in the market.

A former hard-nosed fighter who once refused all major players and subsidized AI with his own money through quantitative trading—how long can he hold out in this bottomless pit?

A Compromise of $20 Billion

In April, DeepSeek leaked news that it would start its first round of external fundraising. The target valuation is as high as 300 billion RMB (about $44 billion). It plans to increase capital by 50 billion RMB, including 30 billion RMB raised from outside investors. Rumors that Tencent and Alibaba are competing to join have been flying.

Many think this is because building data centers is too expensive. But in fact, the core driving force behind DeepSeek’s fundraising—besides buying GPUs—is “pure technological ideals.” In the face of the talent bloodbath among the giants, it doesn’t stand a chance.

During the critical sprint phase of V4 development, domestic tech giants launched a frenzy of targeted poaching. From the second half of 2025 to the present, at least 5 key R&D members at DeepSeek have confirmed they have left. The lead author of the first-generation model, Wang Bingxuan, went to Tencent; V3’s core contributor Luo Fuli was lured away by Lei Jun with a 10-million-yuan annual salary to Xiaomi; and R1’s core author Guo Daya joined ByteDance’s Seed team.

This is the most naked way market economics operates: when your competitor holds unlimited ammunition and you insist on running on your own funds, the talent market becomes your most vulnerable soft underbelly. You can ask geniuses to accept pay cuts and work overtime for the ideal of changing the world. But when big firms slap a check with tens of millions in cash and options on the table and promise unlimited computing resources, the pricing power of idealism is no longer in your hands.

Liang Wenfeng’s dilemma is, in fact, the dilemma that every entrepreneur trying to do a “slow company” in China will inevitably face. In a market where large firms can buy anyone with money, the path of “no fundraising, no commercialization, just technology” is extremely extravagant. The price you pay is that you must accept the reality that your team could be cleared out by the other side at any time with money.

This 300 billion RMB valuation fundraising isn’t Liang Wenfeng’s capitulation to capital; it’s a redemption war he launched to preserve the V4 R&D lineup—waging a “buy-back” war against big firms. He has to sit at the capital table and use the same real money to give those who stay enough reasons to keep staying.

Tencent and Alibaba potentially entering the picture would mean DeepSeek is no longer that lonely, purely idealistic technical thinker. It becomes a company with external shareholders and commercial pressure. The cost of this shift is that the “research freedom” Liang Wenfeng once took pride in—freedom from 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 keep the Ulanqab data center roaring day and night—actually come from?

Another Kind of “Great Effort, Miracles”

The answer isn’t in algorithms. It’s in the power grid.

Silicon Valley’s current anxiety isn’t that chips aren’t enough; it’s that electricity isn’t enough. Musk is building super data centers in Memphis, Tennessee. OpenAI has even begun discussing investing in nuclear power plants. Microsoft announced restarting the Three Mile Island nuclear power plant in Pennsylvania to power AI data centers. The end of computing power is electricity—that’s an extremely cold physical reality.

In the US, the electricity consumption of a large AI data center is comparable to that of a medium-sized city’s daily usage. But the US power grid is an old network built in the 1950s, expanding slowly, regionally fragmented, and fundamentally unable to keep up with the pace of AI-era computing expansion.

And supporting China’s AI race to catch up with the US isn’t only those algorithm geniuses earning tens of millions in salary. It’s also the silent, powerful ultra-high-voltage transmission lines.

The reason the Ulanqab data center can rise from the ground lies in Inner Mongolia’s abundant green electricity and China’s world-leading grid dispatching capability. Public data shows Ulanqab’s green electricity installed capacity reaches 19.402 million kW, accounting for about 65.9%. Locally priced low-cost green electricity is about 50% cheaper than in eastern regions. In addition, with an average annual temperature of only 4.3°C, the natural cooling period is close to 10 months, enabling equipment energy savings of 20% to 30%.

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

Here’s a fascinating and brutal historical parallel. In 1986, the US used the US–Japan Semiconductor Agreement to bring Japan’s semiconductor industry to its knees, forcing Japan to open up its market and accept price controls. Japan’s global market share in semiconductors fell from 40% in 1986 to 15% in 2011. Japan couldn’t recover even after three decades.

Today, the US is trying to lock down China’s AI with the same logic—blocking chips, restricting computing power, and cutting off the technology supply chain. But China’s counter-move is completely different from Japan’s. Japan’s failure was that its semiconductor industry was highly dependent on US technology licensing and market access; once cut off, it lost the ability to survive independently. China’s AI counterattack starts by rebuilding from the most fundamental physical infrastructure—building chips itself, building data centers itself, pulling power grids itself, and open-sourcing models itself.

It’s a route that’s extremely heavy, extremely expensive, but also extremely hard to “strangle.” When Silicon Valley builds magnificent Babel towers in the cloud, China digs trenches in the soil.

If the cloud-based computing arms race is a brutal, heavy-asset consumption war, is there another way to escape cloud dominance besides building data centers in Inner Mongolia and pulling cables?

Escaping the Cloud

When Silicon Valley giants keep building larger and larger data centers, even planning trillion-dollar-level compute clusters like OpenAI, China’s counterattack line has quietly shifted underground.

The ultimate weapon against US computing power lockouts isn’t creating chips stronger than the H100, but stuffing large models into everyone’s smartphones.

Since we can’t win with overwhelming firepower in cloud data centers, we bring the battlefield back to 1.4 billion smart phones and edge devices. This is a typical guerrilla strategy—and one that’s extremely hard to lock down. You can ban the export of high-end GPUs, but you can’t confiscate every phone in every Chinese person’s pocket.

In 2026, amid the computing power anxiety triggered by DeepSeek, China’s phone makers—Xiaomi, OPPO, and vivo—began a frantic “on-device shift.” They were no longer satisfied with using phones merely as displays for calling cloud APIs; through extreme model distillation and compression, they directly cram a miniaturized superbrain into domestic phones that cost just a few thousand yuan.

The core of this technical route is “distillation.” In simple terms, it uses a super large model (“teacher”) to train a smaller model (“student”), so that the student learns the teacher’s “thinking approach” rather than memorizing all the teacher’s “knowledge.” After extreme distillation and quantized compression, a model that originally required hundreds of GPUs to run is compressed to only 1.2GB to 2.5GB, and can run smoothly on a single phone chip.

Mobile AI applications like MNN Chat already allow users to run the DeepSeek R1 distilled model locally on their phones. The significance of on-device AI is that you don’t need to stay connected to 5G all the time, and you don’t need to pay $100 monthly subscription fees to Silicon Valley giants. The large model lives in your pocket; it can run even without internet, and you don’t spend a single cent on cloud compute.

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

Of course, on-device AI isn’t perfect. Limited by the compute and memory of phone chips, the capability ceiling of on-device models is far below that of giant cloud models. It can help you write an email, translate a passage, or summarize an article. But if you want it to derive a complex mathematical theorem or analyze a few hundred pages of legal contracts, it will still fall short.

But that’s already enough. For most ordinary people, the AI they need was never that kind of superbrain that can derive mathematical theorems; they need a “personal assistant” that helps handle everyday hassles.

When large models become extremely cheap—so cheap they can even fit into your pocket—how will they change the corners of the world forgotten by Silicon Valley?

Digital Equality for the Global South

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

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

While Silicon Valley giants discuss how to use AI to rule the world, Ugandan farmers and poor students in Southeast Asia, for the first time, have stepped into the digital age thanks to DeepSeek’s open source.

GPT-5.5 serves people who can afford to pay, and its corpus is almost entirely English. If you ask it a question in Swahili or Javanese, it not only answers clumsily, but the tokens it consumes are several times that of when you ask in English. Silicon Valley giants have voluntarily abandoned these marginal markets due to “low commercial return.”

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

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

In Malaysia, tech companies fine-tuned open-source foundations into AI models that comply with Islamic legal standards. They support Malay and Indonesian, and also ensure outputs meet religious and cultural standards for Muslim markets. From Indonesia’s digital identity systems to Kenya’s Swahili medical Q&A, Chinese technology is penetrating the social grassroots structures of these countries.

Data released in early 2026 by OpenRouter, the world’s largest AI model API aggregation platform, shows that for the first time, Chinese AI models’ token consumption on the platform surpassed that of US competitors. In a certain statistical week, the top 10 popular global models consumed 87 trillion tokens in total, with Chinese models accounting for about 61%.

Open source breaks the US monopoly on AI discourse, allowing resource-scarce developing countries to leap over the digital divide. This isn’t some grand narrative of US–China rivalry. It’s the real “rural encirclement of cities” in the AI era.

China’s AI open-source strategy is objectively turning into an extremely effective form of “soft power” export. When Silicon Valley’s giants build high walls in the cloud to become digital landlords of the new era, those “tech refugees” who can’t afford rent finally find their own spark in open source and on-device “soil.”

Tap Water

Technology should never be a lofty, overpriced luxury.

Silicon Valley built exquisitely designed luxury apartments—strict access control, open only to VIPs. But we laid a tap-water pipeline to thousands of households.

The pipeline’s starting point is in data centers in Inner Mongolia at -20°C, amid the roar of ultra-high-voltage transmission lines, in a war of 300 billion RMB valuations. Every segment is heavy, expensive, and full of forced compromises. Liang Wenfeng once wanted to build a purely technical company, but reality pushed him to build data centers, raise funds, and fight with big firms for talent. He had no choice, because he chose a harder path: not turning AI into a luxury, but turning it into tap water.

And the endpoint of this pipeline is on domestic smartphones costing just a few thousand yuan, in the rough fingers of Ugandan farmers, and in the lives of ordinary people who long to cross the digital divide.

No matter how high the walls of computing power are built, they can’t stop tap water from flowing to the lowlands.