OpenAI left, DeepSeek right

This post is part of our special coverage Syria Protests 2011

The DeepSeek V4 preview was officially released on 24 April 2026。

The Pro version of the 1.6 trillion parameters and the Flash version of the 28.4 billion parameters, which is a large model of national production, have smashed the core sales points to the market, with millions of contexts, and become a free-of-charge frame for all official services. Over the same period of time, OpenAI, on the other side of the ocean, also brought out GPT-55, which is much bigger and Agent is much more functional, but much more expensive。

THE TRANSLATION OF "MILLIONS OF CONTEXT" INTO BIG WHITE WORDS MEANS THAT AI IS NO LONGER A "GOLDFISH" THAT CAN ONLY REMEMBER YOUR EARLIER WORDS, BUT RATHER A "SUPER BRAIN" THAT CAN SWALLOW THREE BODIES IN ONE BREATH, READ A TWO-HOUR FILM IN A SECOND AND HELP YOU PICK OUT THE WRONG WORD。

TO GIVE THE MOST DIRECT EXAMPLE, YOU CAN THROW ALL THE COMPANY'S CONTRACTS, MAILS, FINANCIALS, A BRAIN TO V4 FOR THE LAST THREE YEARS, SO THAT IT CAN HELP YOU FIND THE DEFAULT CLAUSE THAT WAS HIDDEN IN THE ANNEX TO PAGE 47. IN THE PAST, THE MATTER REQUIRED A TEAM OF LAWYERS; NOW IT WAS FREE。

GPT-55.5 Enter this superb brain code price, standard version per million token for $5 and output for $30; and GPT-5.5 Pro, which is for high-level missions, sells a daily price of $30 and $180 per million。

But according to the DeepSeek official pricing, the V4-Flash input per million yuan token's cache is only 0.2 yuan to output 2 yuan; even the V4-Pro for the shoulder top-level closed-source model is 1 dollar for the cache and 12 yuan for the cache outstanding, while the output price is only 24 yuan

IT'S ALWAYS THOUGHT THAT THE COMPETITION BETWEEN CHINA AND THE UNITED STATES IS A MODELING EXERCISE, AND IN FACT, IT'S A LONG-STANDING BUSINESS MODEL。

OpenAI, once the dragon-slayer who shouted "for all mankind", is now selling expensive and refined stores; DeepSeek is turning AI into a hydroelectric coal with near-free calculations。

When OpenAI became a smart foreman, DeepSeek, why turn top AI into free tap water at no cost? What kind of dark current is behind this shift in pricing

The cold wind in Ulaanbaatar

The game of the big model, in a 20-degree machine room in Inner Mongolia。

Just before the release of V4, DeepSeek had an unexpected job in his job announcement: The senior delivery manager for the data centre and the senior transport engineer, with the highest monthly salary of 30,000,14, are based in Ulaanbaatar, Mongolia。

This is a light-asset company that used to be labelled "very simple, pure, only algorithmic". Over the past two years, their most proud label has been "four or two dozen kilos" with less than $6 million in training costs, and the DeepSeek-R1 that plunged the US stock AI board。

HOWEVER, THE ENORMOUS COMPUTING NEEDS OF V4, COUPLED WITH THE INCREASINGLY TIGHT UNITED STATES EMBARGO, HAVE COMPLETELY SHATTERED THIS LIGHT ASSET'S HYMN。

In 2025, the U.S. Department of Commerce further tightened its export controls on the Chinese AI chip, and Yvette H100, H800 has been cut off and even downgraded H20 has been brought into the control list. This means that DeepSeek's future will expand in numbers and will have to turn to China for ecological purposes. In the V4 release note, the government made it clear that the new model had been ‘Wawatung plus’ and revealed that the price of Pro would be reduced significantly after the 950 hypernominal batch was put on the market in the second half of the year。

This shift, which could not be accomplished with several rows of adaptive layers in code, would require a complete infrastructure of national production at the physical level, starting at zero。

The trillions of parameters of V4 (the pre-training data are as high as 33 trillion Token), combined with the huge computing demand of millions of context, means that you need thousands of trip chips, rooms that can house them, electricity grids that need to power the rooms, and teams that are inoperable in a cold of 20 degrees below zero。

Liang Wenbing has taken the methodology from the Bit world to the Atom world. Calculator, eventually, takes root in steel and steel cement and power lines。

ON THE ONE HAND, THE AI ELITE, WHO KNOCKED ON THE CODE IN SILICON VALLEY AND DRANK COFFEE WITH HIS HANDS, AND ON THE OTHER HAND, THE TRANSPORTERS WHO WENT TO THE ENGINE ROOM DEEP IN THE NEPALI SAVANNAH IN THEIR MILITARY COATS. THIS DIFFERENCE FORMS THE BACKGROUND OF TODAY'S CHINA'S AI RESISTANCE TO THE CALCULUS BLOCKADE. THE COLD WIND IN ULANCHABU BECAME CHINA'S STRONGEST PHYSICAL HANGOUT。

The transition from a pure algorithm to a "heavy asset" player in a self-built machine house meant that DeepSeek had left the age of the "magic" guerrilla warfare, officially wearing heavy infantry armor. The costs of this transformation are enormous, with machine rooms, chips, wiring, each of which is a bottomless hole. More importantly, this heavy asset model implies an exponential rise in operating costs, while DeepSeek ' s commercialization revenue remains extremely limited. This pricing strategy is essentially based on the exchange of losses for ecology and free access to infrastructure。

HOW MUCH LONGER CAN A TOUGH GUY WHO ONCE REFUSED ALL THE GIANTS TO PAY FOR AI ON A QUANTITATIVE DEAL

$20 billion compromise

In April, DeepSeek launched its first external financing, with a target estimate of RMB 30 billion (approximately US$ 44 billion) and a planned increase of $50 billion, of which $30 billion was raised from external sources. There have been reports that Ali has been involved in a fight to gain access。

A lot of people think it's because building a room is too expensive. In practice, however, the core driving force of DeepSeek's financing, in addition to purchasing graphic cards, is the "pure technological ideal" that is vulnerable to a meat-scrowding machine for giants。

During the critical sprinting period of V4, large domestic plants opened a crazy, directed dig for DeepSeek. From the second half of 2025 to date, DeepSeek has confirmed the departure of at least 5 core R & D members. Wang Bing, the first generation model core author, went to the plaza, and the V3 core contributor, Roofley, was dug up to millennia by the Rey Army, while the R1 core author, Guo Dayazi, joined the byte beat Seed team。

This is the most naked way in which a market economy operates, and the talent market is your weakest weakness when your competitors hold unlimited ammunition and you insist on operating with your own funds. You can ask geniuses to work overtime in order to change the ideal world, but when a large factory puts on a desk a check with millions of cash and options, and promises unlimited computing resources, idealistic pricing is not in your hands。

The plight of Liang Wensai is the one that every entrepreneur trying to become a slow company in China faces. In a market where a large factory can buy anyone with money, the route of "no financing, no commercialization, no technology" is extremely expensive. The price is that you have to accept that your team can be cleared with money at any time。

THE FINANCING OF THIS $30 BILLION VALUATION WAS NOT A COMPROMISE ON CAPITAL BY LIANG WENFENG, BUT A FORECLOSURE WAR HE LAUNCHED AT LARGE FACTORIES TO PRESERVE THE V4 FORMATION. HE MUST SIT AT THE CAPITAL TABLE AND USE THE SAME REAL SILVER AND SILVER TO GIVE THE PEOPLE WHO STAYED ENOUGH REASON TO STAY。

It means DeepSeek is no longer the lonely, purely technical idealist. It became a company with external shareholders and commercialized pressure. The price of this transformation is the kind of "freedom of research free from external pressure" that Liang Wenfeng was most proud of, and will inevitably be diluted。

But he had no choice。

Where does idealism come from when it is forced to wear capital armour to support the continuation of this huge machine and to sustain the lull of the Ulanchabu room

Another "magic force."

The answer is not in algorithms, in grids。

Silicon Valley's most anxious now is not the lack of chips, but the lack of electricity. Mask built a super data centre in Memphis, Tennessee, and OpenAI even started talking about investing in a nuclear power plant, and Microsoft announced the reopening of the Trinity Nuclear Power Plant in Pennsylvania to power the AI data centre. At the end of the calculation is electricity, which is an extremely cold physical common sense。

IN THE UNITED STATES, A LARGE AI DATA CENTRE USES ELECTRICITY EQUIVALENT TO DAILY ELECTRICITY IN A MEDIUM CITY. AND THE UNITED STATES GRID, AN OLD NETWORK BUILT IN THE 1950S, IS SLOW TO EXPAND, IS FRAGMENTED, AND DOES NOT KEEP PACE WITH THE PACE OF EXPANSION IN THE AI ERA。

AND CHINA AI, WHICH IS SUPPORTING AMERICA, IS NOT ONLY THE ALGORITHMIC GENIUSES WITH MILLIONS OF YEARS' PAY, BUT ALSO THE UNHEARD OF HIGH-VOLTAGE TRANSMISSION LINES。

The data centre in Ulanchabu was made possible by the abundance of green electricity in Inner Mongolia and the power grid deployment capability of the world's first in China. Public data show that the capacity of the Ulanchab green power plant is 1940.2 kilowatts, representing about 65.9 per cent of the total, and that local green power is about 50 per cent cheaper than the eastern region. Plus an average annual temperature of only 4.3°C, the natural cooling period is close to 10 months, saving equipment from 20% to 30%。

When DeepSeek V4 runs, the real blood transfusion is China's huge and extremely cheap power infrastructure. It's another dimension of the "magic miracle."。

Here is an extremely interesting and cruel historical contrast. In 1986, the U.S. used the US-Japan Semiconductor Agreement to crush Japan ' s semiconductor industry, forcing Japan to open its markets and accept price regulation, and the global market share of Japanese semiconductors fell from 40 per cent in 1986 to 15 per cent in 2011. It took Japan 30 years to get here。

TODAY, THE UNITED STATES IS TRYING TO USE THE SAME LOGIC TO LOCK CHINA ' S AI, BLOCK THE CHIP, LIMIT THE CALCULUS, AND CUT THE TECHNOLOGY SUPPLY CHAIN. BUT CHINA'S COUNTERATTACK PATH IS COMPLETELY DIFFERENT FROM JAPAN'S. JAPAN ' S FAILURE THAT YEAR WAS DUE TO THE FACT THAT ITS SEMICONDUCTOR INDUSTRY WAS HIGHLY DEPENDENT ON UNITED STATES TECHNOLOGY AUTHORIZATION AND MARKET ACCESS, WHICH, ONCE CUT OFF, LOST ITS ABILITY TO SURVIVE INDEPENDENTLY. CHINA AI'S COUNTER-ATTACK BEGAN WITH THE LOWEST PHYSICAL INFRASTRUCTURE, BUILDING ITS OWN CHIPS, BUILDING ITS OWN MACHINE HOUSE, PULLING ITS OWN GRID, AND OPERATING ITS OWN OPEN SOURCE MODEL。

This is a very heavy and expensive route, but it is also extremely difficult to "strung." When Silicon Valley built the gorgeous barbitas in the clouds, China dug trenches in the dirt。

If cloud computing is an extremely heavy asset-consumption war, do we have another way to escape cloud hegemonic domination, except to repair the engine room in Inner Mongolia and pull the wire

Escape the clouds

When Silicon Valley giants were building data centres increasingly, even like OpenAI, planning a multi-billion-dollar calculator cluster, China’s counter-attack line was quietly transferred underground。

THE ULTIMATE WEAPON AGAINST THE U.S. EMBARGO IS NOT A CHIP STRONGER THAN H100, BUT RATHER A LARGE MODEL INTO EVERYONE'S CELL PHONE。

SINCE WE CAN'T GET HEAVY FIRE IN THE CLOUD ROOM, WE'RE PULLING BACK 1.4 BILLION SMARTPHONES AND PERIPHERALS. IT'S A TYPICAL GUERRILLA WARFARE METHOD, AND IT'S AN EXTREMELY DIFFICULT METHOD TO BLOCK, SO YOU CAN BAN THE EXPORT OF HIGH-END GPUS, BUT YOU CAN'T CONFISCATE CELL PHONES FROM EVERY CHINESE POPULATION BAG。

In 2026, along with DeepSeek's calculus anxiety, Chinese mobile phone manufacturers Mi, OPPO, and vivo began a crazy "side shift." They're no longer content with just using a mobile phone as a monitor for the cloud-based API, but instead, using extreme model distillation and compression, they put a reduced version of the superb brain into thousands of dollars of domestic phone。

AT THE HEART OF THIS TECHNOLOGY IS DISTILLATION. IN SHORT, IT'S USING A SUPER-LARGE MODEL (TEACHER) TO TRAIN A SMALL MODEL (STUDENT) TO LEARN THE TEACHER'S "THINKING STYLE" RATHER THAN REMEMBER ALL THE "KNOWLEDGE" OF A HARD-BACK TEACHER. AFTER EXTREME DISTILLATION AND QUANTITATIVE COMPRESSION, A LARGE MODEL THAT WOULD HAVE REQUIRED HUNDREDS OF GPUS TO RUN, WAS COMPRESSED TO A SCALE OF 1.2 TO 2.5 GB, AND COULD RUN ON A CELL PHONE CHIP。

Mobile-end AI applications like MNN Chat already allow users to locally run DeepSeek R1 distillation models on their mobile phones. The point of this end-side AI is that you don't need to be connected to a 5G signal and pay 100 dollars a month for subscriptions to Silicon Valley giants. The big model is in your pocket, and you can run without a penny for cloud computing。

Now that I can't fix the super boiler room, I'll give each house a little stove。

OF COURSE, THE END-SIDE AI IS NOT PERFECT. THE UPPER LIMIT OF THE CAPACITY OF THE END-SIDE MODEL, WHICH IS LIMITED TO THE COMPUTING POWER AND MEMORY OF THE MOBILE CHIP, IS FAR LESS THAN THAT OF THE SUPERMODEL OF THE CLOUD. IT CAN HELP YOU WRITE AN E-MAIL, TRANSLATE A TEXT, SUMMARIZE AN ARTICLE, BUT IF YOU WANT IT TO PUSH A COMPLEX MATHEMATICAL THEOREM FOR YOU, OR ANALYZE A HUNDREDS OF PAGES OF LEGAL CONTRACTS, IT'S NOT GOING TO WORK。

BUT THAT'S ENOUGH. BECAUSE FOR THE VAST MAJORITY OF ORDINARY PEOPLE, THE AI THEY NEED IS NEVER THE SUPER-BRAIN THAT CAN DICTATE THE THEOREM OF MATHEMATICS, BUT RATHER A "BODY HELPER" WHO CAN HELP THEM WITH THEIR DAILY CHORES。

How will a large model change those corners that have been forgotten by Silicon Valley when it becomes extremely cheap and can even be packed in imported bags

Digital parity in the global South

IF YOU'RE SITTING IN MANHATTAN'S PANORAMIC GLASS OFFICE, YOU PROBABLY THINK GPT-5.5 IS WORTH $100 BECAUSE IT'S GONNA HELP YOU WRITE A PERFECT MERGER AND BUY IN A SECOND。

But if you're standing in a corn field in East Africa, Uganda, facing crops that are dying of climate anomalies, no one can afford $100 for subscriptions, because Uganda has less than $150 per person per month。

The giants of Silicon Valley are discussing how to rule the world with AI, while farmers in Uganda and poor students in South-East Asia are entering the digital age for the first time because of DeepSeek ' s opening source。

GPT-5.5 serves those who can afford it, and its language library is almost exclusively English. If you ask it a question in Kiswahili or Javan, it's not only twitching, it's consuming Token several times as much as English. Silicon Valley giants have voluntarily abandoned these marginal markets because of "low commercial returns"。

China ' s open source model has become the digital infrastructure of the global South。

In Uganda, the local NGO Sunbird AI, using the SunFlower system, which was fine-tuned on the basis of the Chinese open source model Qwen, expanded the number of local languages supported from six to 31. The system is now being deployed in the agricultural extension system of the Government of Uganda to send farming advice in Kiswahili to farmers。

IN MALAYSIA, TECHNOLOGY COMPANIES USED THE OPEN SOURCE BASE TO FINE-TUNE THE ISLAMIC SHARIAH-COMPLIANT AI MODEL, NOT ONLY SUPPORTING MALAY AND INDONESIAN, BUT ALSO ENSURING THAT THE OUTPUT CONFORMED TO RELIGIOUS AND CULTURAL STANDARDS IN THE MUSLIM MARKET. FROM THE DIGITAL IDENTITY SYSTEM IN INDONESIA TO THE MEDICAL QUESTIONS AND ANSWERS IN KISWAHILI IN KENYA, CHINESE TECHNOLOGY IS PENETRATING THE SOCIAL FABRIC OF THESE COUNTRIES。

The data released by OpenRouter at the beginning of 2026 showed that China ' s AI model for the first time consumed more Token on the platform than US competitors. In one statistical week, the top 10 global hot-gate models consumed 8.7 trillion Token, of which China accounted for about 61 per cent。

THE OPEN SOURCE HAD BROKEN THE UNITED STATES MONOPOLY ON THE VOICE OF AI AND HAD ALLOWED RESOURCE-POOR DEVELOPING COUNTRIES TO CROSS THE DIGITAL DIVIDE. THIS IS NOT A BIG STORY ABOUT CHINA AND AMERICA, IT'S THE REAL "RURAL SIEGE" OF THE AI ERA。

CHINA’S AI OPEN SOURCE STRATEGY IS OBJECTIVELY BECOMING AN EXTREMELY EFFECTIVE “SOFT POWER” EXPORT. WHEN THE GIANTS OF SILICON VALLEY BUILT WALLS IN THE CLOUDS AND TRIED TO BECOME DIGITAL OWNERS OF THE NEW ERA, THOSE “TECHNICAL REFUGEES” WHO COULD NOT AFFORD TO PAY RENT FINALLY FOUND THEIR OWN KIND OF FIRE IN OPEN SOURCES AND IN THE MUD ON THE END。

Water supply

Technology should never be a luxury of high heights。

SILICON VALLEY BUILT AN EXTREMELY FINE STOREHOUSE, CLOSED DOORS, OPEN ONLY TO VIP. BUT WE HAVE FIXED A PIPE TO THOUSANDS OF FAMILIES。

THE STARTING POINT FOR THIS PIPE, IN THE 20-DEGREE MACHINE ROOM IN INNER MONGOLIA, IN THE SENSATION OF THE ULTRA-HIGH VOLTAGE TRANSMISSION LINE, IN THE 30 BILLION-VALUED WAR. EACH OF ITS SEGMENTS IS HEAVY, EXPENSIVE AND FULL OF FORCED COMPROMISE. LIANG WENSAI HAD WANTED TO BE A PURELY TECHNICAL COMPANY, BUT THE REALITY FORCED HIM TO BUILD A MACHINE HOUSE, FINANCE, AND ROB PEOPLE. HE HAD NO CHOICE, BECAUSE HE HAD CHOSEN A MORE DIFFICULT PATH, NOT TO MAKE AI A LUXURY, BUT TO MAKE IT A TAP WATER。

And the end of this pipe, on a thousands-dollar national phone, between the rough fingers of Ugandan farmers, in the lives of every ordinary person eager to cross the digital divide。

The height of the calculator wall cannot stop the flow of running water to the low。