OpenAI向左，DeepSeek向右

Text | Sleepy.md

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

This homegrown large model—its Pro version has 1.6 trillion parameters and its Flash version has 284 billion parameters—has thrown its core selling points directly at the market: a million-context window has become the free default feature across all official services. At almost the same time, across the ocean, OpenAI also rolled out GPT-5.5. It has even more compute power and richer Agent capabilities, but its price is much higher.

Translating “million-context” into plain, everyday language means AI is no longer a “goldfish” that can only remember your first few sentences; instead, 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 spot typos along the way.

Here’s the most straightforward example: you can dump every contract, email, and financial report from your company from the past three years into V4 at once, 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 openly prices this kind of “superbrain” in plain terms. 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, is even sold at a staggering $30 for every million input tokens and $180 for output.

But according to DeepSeek’s official pricing, for V4-Flash, inputs with cache hits are only 0.2 RMB per million tokens, and outputs are 2 RMB; even for V4-Pro, which is on par with top-tier closed-source models, cached-hit inputs cost 1 RMB, uncached inputs cost 12 RMB, and output prices are only 24 RMB.

Everyone thinks the AI competition between China and the US is a race of model capability. In reality, it has long turned into a divergence of business models.

OpenAI used to be that dragon-slaying youth shouting “to benefit all humankind,” but now it’s selling expensive, turnkey luxury apartment units; meanwhile, DeepSeek is using near-free compute power to turn AI into water, electricity, and gas.

When OpenAI turns into a shrewd contractor, why does DeepSeek pour top-tier AI into the world at near-zero cost as free tap water? Behind this shift in pricing power, what kind of undercurrent is hidden?

The Cold Wind of Ulanqab

The decisive round for large models takes place in a data center in Inner Mongolia at below 20 degrees Celsius.

Not long before the release of V4, a DeepSeek job posting added an unexpected role: Senior Data Center Delivery Manager and Senior Operations Engineer. The monthly salary can reach 30,000, with 14 payments per year, and the position is stationed at Ulanqab, Inner Mongolia.

This is a light-asset company that once marketed itself as “minimalist, pure, only doing algorithms.” Over the past two years, their most proud tagline has been “a light touch that moves mountains.” With training costs of less than 6 million USD, they produced DeepSeek-R1, which made the US stock AI sector plunge.

But the enormous compute requirements of V4, combined with increasingly tight US compute-power restrictions, shattered that light-asset pastoral dream completely.

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 pulled into the controlled list. This means DeepSeek’s future compute expansion must fully pivot to Huawei Ascend’s ecosystem. In V4’s release documentation, the official stated clearly that the new model is “empowered by Huawei Ascend,” and it also revealed that after the batch rollout of Ascend 950 supernodes 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, an entire domestic compute infrastructure at the physical layer.

With V4’s trillion-parameter scale (pretraining data reaches 33 trillion tokens), plus the massive compute demands of a million-context window, it means you need tens of thousands of Ascend chips, data centers capable of housing those chips, power grids to supply electricity to those data centers, and operations teams that can keep these machines running without downtime in sub-zero 20-degree cold winds.

Liang Wenfeng took his methodology from the world of bits into the world of atoms. In the end, compute power has to take root in steel, concrete, and power transmission lines.

On one side are AI elites in plaid shirts coding in Silicon Valley and sipping pour-over coffee. On the other are operations staff wrapped in padded military coats going deep into Inner Mongolia’s grasslands to guard the data centers. This contrast forms the background color of China’s AI resistance to compute-power restrictions. The cold wind of Ulanqab has become China’s strongest physical external support for AI.

Transitioning from a pure algorithm company to a “heavy-asset” player that builds its own data centers means DeepSeek has said goodbye to the guerrilla era of “small power achieves extraordinary results,” and has officially put on the armor of heavy infantry. The cost of this transformation is enormous. Repairing data centers, buying chips, laying cables—every item is a bottomless pit. More importantly, this heavy-asset model means operating costs will rise exponentially, while DeepSeek’s commercialization revenues remain extremely limited. This pricing strategy, in essence, is exchanging losses for ecosystem-building and exchanging free access for the right to shape infrastructure and discourse power.

A tough guy who once refused all big players and used his own money from quantitative trading to subsidize AI—how long can he last in this bottomless pit?

A Compromise of $20 Billion

In April, DeepSeek reported that it had started its first round of external financing, with a target valuation as high as 300 billion RMB (about 44 billion USD). It plans to increase capital by 50 billion, including 30 billion from outside fundraising. Rumors that Tencent and Alibaba are competing to enter the deal have been swirling.

Many people think this is because building data centers is too expensive. But in fact, aside from buying GPUs, the core driving force behind DeepSeek’s fundraising is “pure technological ideals,” which are defenseless in the talent meat grinder run by tech giants.

During the critical sprint period of V4 development, domestic big companies launched a frenzy of targeted poaching. From the second half of 2025 to now, at least 5 core R&D members of DeepSeek have confirmed resignations. The first-generation model’s core author Wang Bingxuan went to Tencent. V3’s core contributor Luo Fuli was lured away by Lei Jun with a ten-million-yearly-salary offer to Xiaomi. And R1’s core author Guo Daya joined ByteDance’s Seed team.

This is the most naked operation of a market economy: when your competitors hold unlimited ammunition and you insist on using your own funds to keep running, the talent market becomes your most vulnerable weak link. You can ask geniuses to accept pay cuts and overtime out of ideals to change the world, but when big companies put a check worth ten million in cash and options on the table and promise unlimited compute resources, the pricing power of idealism is no longer in your hands.

Liang Wenfeng’s predicament is, in fact, a dilemma that every entrepreneur trying to run a “slow company” in China will eventually face. In a market where big companies can buy anyone with money, the path of “no financing, no commercialization, only technology” is extremely extravagant. The cost is that you must accept the possibility that your team could be cleared out at any moment by the other side using money.

This 300 billion RMB valuation financing isn’t Liang Wenfeng’s capitulation to capital. It’s a ransom/hostage war he launched to preserve the V4 R&D lineup. He has to sit at the capital table, using the same real cash to give those who stay enough reasons to keep staying.

If Tencent and Alibaba possibly enter, it means DeepSeek is no longer that lonely, purely idealistic technology player. It becomes a company with external shareholders and commercialization pressure. The cost of this shift is that the “research freedom” that Liang Wenfeng was once most proud of—freedom from external pressure—will inevitably be diluted.

But he had no choice.

When idealism is forced to put on capital’s armor, where does the confidence to keep this huge machine running—and to keep the Ulanqab data center roaring day and night—ultimately come from?

Another Kind of “Big Power, Big Miracle”

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

What Silicon Valley is most anxious about right now isn’t that there aren’t enough chips, but that there isn’t enough electricity. Musk is crazily 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 limit of compute power is electricity—this is an extremely cold physical reality.

In the US, the electricity usage of a large AI data center is equivalent to the daily electricity consumption of a mid-sized city. But the US power grid is an old network built in the 1950s—slow to expand, regionally fragmented, and fundamentally unable to keep up with the pace of compute expansion in the AI era.

What supports China’s AI to catch up with the US isn’t only those algorithm geniuses earning ten-million-a-year salaries; it’s also those high-voltage transmission lines that stay out of the spotlight.

The reason data centers in Ulanqab can rise from the ground is Inner Mongolia’s abundant green electricity and China’s world-leading power grid dispatching capability. Public data shows Ulanqab’s green power installed capacity is 19.402 million kW, accounting for about 65.9%. Local low-priced green electricity is about 50% cheaper than in the eastern regions. Combined with an average annual temperature of only 4.3°C and a natural cooling period close to 10 months, it can cut equipment energy use by 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 “big power, big miracles.”

Here is an extremely interesting—and brutal—historical comparison. In 1986, the US used the US-Japan Semiconductor Agreement to knock Japan’s semiconductor industry flat, forcing Japan to open its market and accept price controls. Japan’s global market share in semiconductors fell from 40% in 1986 to 15% in 2011. Japan took thirty years and still didn’t recover.

Today, the US is trying to lock down China’s AI with the same logic—blocking chips, limiting compute power, and cutting off the technology supply chain. But China’s counterattack path is completely different from Japan’s. Japan’s failure back then was that its semiconductor industry was highly dependent on US technology licensing and market access. Once that was cut off, it lost the ability to survive independently. China’s AI counterattack starts by rebuilding from the very lowest-level physical infrastructure: making chips, building data centers, building power grids, and open-sourcing models.

It’s an extremely bulky, extremely costly, yet also extremely hard to “strangle” approach. While Silicon Valley builds magnificent Babel towers in the cloud, China digs trenches in the soil.

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

Escape from the Cloud

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

The ultimate weapon against US compute-power restrictions isn’t to make chips stronger than H100—it’s to put large models into everyone’s smartphones.

Since we can’t outgun the heavy firepower in cloud data centers, we’ll pull the battlefield back to 1.4 billion smartphones and edge devices. This is a classic guerrilla warfare style, and it’s extremely hard to block. You can ban the export of high-end GPUs, but you can’t take away every phone from the pockets of 1.4 billion people in China.

In 2026, with the compute-power anxiety triggered by DeepSeek, Chinese phone manufacturers Xiaomi, OPPO, and vivo began a frantic “end-side shift.” They no longer just treat phones as displays for calling cloud APIs. Instead, through extreme model distillation and compression, they cram a scaled-down version of a superbrain into domestic smartphones that cost just a few thousand yuan.

The core of this technical route is “distillation.” Put simply, it means using a super large model (the teacher) to train a smaller model (the student), so that the student learns the teacher’s “way of thinking,” not just rote-memorizing all the teacher’s “knowledge.” After extreme distillation and quantized compression, a large model that previously required hundreds of GPUs to run can be compressed down to only 1.2GB to 2.5GB, running smoothly on a single phone chip.

Mobile AI apps like MNN Chat can already let users run the DeepSeek R1 distilled model locally on their phones. The significance of on-device AI is that you don’t have to stay constantly connected to the 5G signal, and you don’t have to pay $100 per month in subscriptions to Silicon Valley giants. The large model sits in your pocket, can run without a network connection, and you don’t have to spend a single cent on cloud compute.

Since I can’t install a centralized heating super-boiler room, I’ll give every household a small stove.

Of course, on-device AI isn’t perfect. Limited by the compute power and memory of phone chips, the upper limit of on-device model capability is far lower than that of huge cloud models. It can help you write an email, translate a passage, or summarize an article, but if you want it to help you derive a complex mathematical theorem or analyze a few 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 a superbrain that can derive mathematical theorems. What they need is a “personal assistant” to handle everyday hassles.

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

Digital Equality in the Global South

If you sit in a panoramic glass office in Manhattan, you’ll probably think it’s worth paying $100 for GPT-5.5 because it can help you write a perfect M&A financial report in a second.

But if you stand in a cornfield in East Africa, in Uganda, facing crops that have yellowed because of abnormal weather, no one can afford the $100 subscription fee, because Uganda’s monthly per-capita income is under $150.

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

GPT-5.5 serves people who can afford to pay, and its training corpora are almost entirely English. If you ask it a question in Swahili or Javanese, it not only answers awkwardly, but also consumes several times more tokens than it does in English. Because Silicon Valley giants view the “commercial return” as low, they actively give up on these edge markets.

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

In Uganda, the local NGO Sunbird AI uses Sunflower—a system fine-tuned based on China’s open-source model Qwen—to expand the local supported languages from 6 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 compliant with Islamic law. They not only support Malay and Indonesian, but also ensure that outputs meet the religious and cultural standards of Muslim markets. From Indonesia’s digital identity systems to Kenya’s Swahili medical Q&A, Chinese technology is seeping into these countries’ social grassroots infrastructure.

According to data released by OpenRouter, the world’s largest AI model API aggregation platform, in early 2026, token consumption by Chinese AI models on the platform first surpassed that of its US competitors. In a certain statistical week, the top 10 popular models worldwide consumed 8.7 trillion tokens in total, and Chinese models accounted for about 61%.

Open source breaks the US monopoly on AI discourse, enabling resource-scarce developing countries to cross the digital divide. This isn’t some grand narrative of a China-US rivalry; in the AI era, it’s the real “rural encirclement of cities.”

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

Tap Water

Technology should never be a luxury item reserved for the lofty elite.

Silicon Valley built exquisite luxury apartments—strict access control, open only to VIPs. But we laid a pipeline of tap water leading to every household.

The starting point of this pipeline is in a data center in Inner Mongolia at below 20 degrees Celsius, amid the roar of ultra-high-voltage transmission lines, in a war with a 300 billion valuation. Every segment is heavy, every segment is expensive, and every segment is full of forced compromises. Liang Wenfeng once wanted to build a purely technical company, but reality forced him to build data centers, seek financing, and compete for talent with big firms. He had no choice—he chose a harder path: not making AI a luxury, but making it tap water.

And the endpoint of this pipeline is in a few-thousand-yuan domestic smartphone, in the rough fingers of farmers in Uganda, in the lives of every ordinary person who longs to cross the digital divide.

No matter how high the wall around compute power is built, it can’t stop tap water from flowing to lower places.

Click to learn about Rhythm BlockBeats job openings

Welcome to join the official Rhythm BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia