GuShiio.com Wen| Digital Intelligence Frontline, author| Zhao Yanqiu, Editor| Niu Hui
Recently, there was an article claiming that DeepSeek had stopped the construction of smart computing centers in many places, but many industry insiders disagreed. After the Spring Festival,Many companies, especially listed companies, are connecting to DeepSeek。The next few months will be an application verification period.
Industry feedback, this year’s domestic smart computing market will present several characteristics:
1. Large smart computing orders in the training market will be reduced, and project decisions in large smart computing centers will be on the sidelines of half a year to a year.
2. In the reasoning market, large models with high availability and low prices have caused the demand of ordinary enterprises to explode. Cloud Dachang’s reasoning services have grown. The privatization deployment of 1 to 20 units within the enterprise will break out.
3. Computing power demand has shifted from centralized to fragmented. The overall smart computing market is still expected to grow, and the construction of smart computing centers will not stop.
4. Reasoning reduces the requirements for computing power, which benefits localized computing power. Some of the previously vacant domestic smart computing centers are expected to be put into operation.
5. Intelligent computing centers using liquid cooling technology are less urgent.
6. Nvidia’s share price is attributed to GPT5, and domestic market demand is still very tight.
Smart computing orders will wait and see for more than half a year
“Big smart computing orders will drop sharply。& rdquo; A server veteran commented on the digital intelligence front line.
In the training market, although major Internet companies will not hand over the big model market to DeepSeek, except for the top few companies, most other companies will focus on application development. In application development, it depends on the degree of secondary development. If you only make fine adjustments to the large model, you won’t use large-scale computing power.
DeepSeek breaks previous myths, reduces computing power consumption, and opens up a new idea for the industry. ldquo; This is definitely a good thing. quo; A smart calculator told Digital Frontline that“It is really unreasonable to follow the method of unlimited expansion of computing power and the whole country providing electricity to it. The industry is not supposed to be like that。& rdquo;
There is a consensus in the industry that being tricked by OpenAI into building a cluster of 100,000 cards or 500,000 cards is no longer so urgent.
In addition, leaders ‘decisions will be affected by the decision of super-large smart computing centers that were blindly or hesitatingly invested in before, such as projects that wanted to invest more than one billion yuan or billions to build.
“For large-scale smart computing projects, half a year to a year this year should be a wait-and-see period.& rdquo; Analysis from a senior person in smart computing. Companies with large model applications estimate that the next few months will be the stage of application verification for DeepSeek.
However, large orders will still exist.& ldquo;DeepSeek is not the end. The competition between multimodal models and world models is just the beginning.” quo; A person from a big model company said that“These trainings are at least the beginning of Wanka。& rdquo;
In the reasoning market, large public cloud companies are urgent
“From the perspective of the reasoning market, in the past year, the implementation of large models has not yet truly entered a stage of explosive growth in reasoning and applications in various industries. rdquo; Some smart computing people said that DeepSeek’s low cost may accelerate smart applications in various industries into an explosive growth period.
In the reasoning market, one is a public cloud and the other is a private cloud.
The public cloud belongs to big manufacturers. Two years after ChatGPT was launched, big manufacturers are in urgent need. In the United States, once the DeepSeek R1 open source model was launched, Microsoft and Amazon AWS were even faster than China cloud companies. Officials announced that they would provide relevant services. They have long been looking forward to the arrival of the reasoning market.
Microsoft CEO Nadella even posted in the early morning after the release of DeepSeek R1: Jevons ‘Paradox has come true again! As AI becomes more efficient and easier to obtain, its use will rocket and become a commodity we can’t use enough. rdquo;
Picture transferred from Z Potencials public account
According to analyst Ben Thompson, the core economic reason why Microsoft and OpenAI were drifting away was that Microsoft was interested in providing reasoning services, but was reluctant to pay for a $100 billion data center because these models were commercialized. They may have been distilled, copied, and made cheap before they were commercialized.
At present, domestic public cloud enterprises and operators ‘cloud computing platforms are connected to DeepSeek. Some people told Digital Intelligence Frontline that a large number of users have poured into early adopters, and some cloud platforms have been unable to support them.
In addition to large public cloud companies, the industry speculatesDeepSeek itself may provide similar services。Years ago, MongoDB was an open source database, and eventually it became a SaaS service itself. ldquo; Big factories will not give you a penny if they take these open source models. The American public cloud has long used a large amount of open source software to make money, and this story will repeat itself in China. rdquo; Analysis by a senior cloud computing person.
03
Competition for the privatization deployment market
In addition to public clouds, another reasoning market is private clouds or private deployments. The industry has reached a consensus that the market should explode.
After the Spring Festival, a smart computer person visited a customer. ldquo; All listed companies are accessing DeepSeek.& rdquo; He told Digital Intelligence Frontline that companies are making their own local deployments and integrating them with their own businesses, whether AI digital people, AI games, AI singers, AI programmers, or AI sales.
“It’s like a bunch of people were trying their best to use big computers, and suddenly IBM launched a personal computer. rdquo; Another smart calculator said. In the past, large computers were very expensive, and only a country could buy a few. IBM introduced PCs, which became something that every household could use.
Enterprises will build their own small smart computing centers and deploy 1 to 10 servers (within 100 cards), or 10 to 20 servers (100 cards scale)。& ldquo; This is also the direction that all IT business owners are pushing to customers this year.& rdquo; Analysis from senior server personnel. In the next few months, each company will verify its application and show its own abilities.
“Why privatization deployment? For example, can a law firm throw contracts with clients into the public cloud? This will leak the leak, so there will be many business opportunities for privatization deployment. rdquo; Cloud computing veterans said.
butEveryone uses the big open source model, and the hardware is also homogeneousTherefore, this year, technical service providers will make their own differences, with the core being to improve reasoning speed and cost performance.
Software and hardware integration will become critical. Cooperation between chip companies, server companies, solution companies, and private cloud service providers will become active.
In terms of business models, the industry predicted last year that under the large-scale refining model, the concentration of computing power demand will become higher and higher. This year, various strategies are being adjusted to serve more medium-sized and even small customers, the granularity will become smaller, and channel cooperation will become critical.
The industry is cautiously optimistic about the smart computing market
If there are fewer large smart computing orders and more fragmented orders for privatization deployments in 2025, will China’s smart computing market as a whole rise or fall?
Many people remain optimistic while being cautious.& ldquo; If a small business buys a machine this year for two to three hundred thousand yuan, and a larger business buys two to three million yuan,10,000 companies will cost 30 billion yuan。Instead, I think this data will increase year by year。& rdquo; said the above-mentioned senior cloud computing person.
“DeepSeek is equivalent to the watt moment。After Watt improved the steam engine, it achieved a stable power output and allowed steam engines to enter various industries. rdquo; The above-mentioned server person said that the big model is a steam engine. After being improved, it can be used in all walks of life.” rdquo;
He found that the stocks of some professional application companies, such as medical companies, have begun to soar, and DeepSeek has made the application market more imaginative. ldquo; In the past, the only person making money would be Nvidia.& rdquo;
“Is the scale of a large order large, or is the scale of a number of small orders combined large?”Senior cloud computing professionals further analyzed that the significance of DeepSeek is that a large model with high availability and cheap is like a personal computer, which will cause the demand of ordinary enterprises to explode.” rdquo;
“When smart applications become popular, it will promote the large-scale expansion of computing infrastructure. rdquo; Intelligent computing people say that this means that even if the computing card is cheap, the overall amount is not small.
Nvidia’s share price is based on GPT5
On the day DeepSeek released R1, Nvidia’s share price plunged, with a one-day drop of 16.97%, and its market value evaporated by about US$590 billion.
“Everything used to be the best for Nvidia, but now there is a small crack and everyone is questioning its prospects.” quo; A chip veteran told Digital Intelligence Frontline that“Nvidia now has a gross profit margin of 90%, a little too high. Its share price has long needed adjustment, and he has earned all the money in the entire industry. It is unreasonable and unsustainable. rdquo;
A phenomenon worthy of attention is that although the major U.S. companies, Microsoft, Google, Meta and xAI, are not cutting capital expenditures (Capex), these companies are all making their own chips.
“Why did Broadcom’s share price rebound faster than Nvidia this time?Broadcom helps companies such as Microsoft, Meta, and Google make ASIC chips (dedicated chips). rdquo; The above-mentioned person said. ASIC chips have obvious advantages in terms of performance, power consumption and cost, and these companies will reduce their reliance on Nvidia in the future.
“Industry won’t let a person lie down for too long, you can’t always win by lying down.。& rdquo; The above-mentioned intelligent computing person said that although Nvidia has not laid down in recent years, it has essentially not launched anything as revolutionary as CUDA for more than a decade. rdquo;CUDA improves industrial efficiency by providing rich development tools, library functions and ecological support from the community.
The above-mentioned people believe that judging from Nvidia’s actions in recent years, its cards are getting bigger and more integrated. In fact, it is still following the path of inertia.& ldquo; In the CPU market at that time, Intel beat AMD, but now AMD’s market value is three times that of Intel. You can’t always win lying down in a field, you have to stand up.& rdquo;
At present, DeepSeek has also been introduced into the NVIDIA CUDA ecosystem. ldquo; It is very smart. If the big ones don’t work, I will do the small ones. If the demand is reduced, I will do the side.” rdquo; At this year’s CES, Huang Renxun has begun to focus on introducing mid-to-low-end 5090 graphics cards to compete for the reasoning market.
“Nvidia’s stock depends on the effect of GPT5 after it comes out.& rdquo; Some cloud computing professionals told Digital Intelligence Frontline that if the effect of GPT5 is very poor, that is, it is not much better than it is now, Nvidia’s share price will definitely collapse. If GPT5 is amazing after it comes out, everyone thinks that Dali’s miracle still exists and Nvidia’s stock will stabilize.
However, judging from the actions of major Silicon Valley companies, with the explosion of the reasoning market, Nvidia will face very severe competition in the future.
How can domestic computing power seize the opportunity
In the reasoning market, computing power is blooming, and this opportunity is good for domestic computing power.
Currently,NvidiaIn addition to high-end chips, 4090 and 5090 chips have also joined the market competition, and some markets have begun to increase prices;domestic chipsWe are stepping up efforts to adapt DeepSeek to improve cost performance; in the global market,Three American reasoning chips Groq, Cerebras, and SambaNova, both claim performance several times that of the H100, and they already support the DeepSeek-R1-Distill Llama-70B model and improve reasoning speed.
How can domestic computing power seize this opportunity? The industry believes that the key is to combine it with software to improve reasoning efficiency.
On the other hand, if you look at overseas companies and developers, they are voting with their feet to use DeepSeek. In this case,Is it possible for domestic computing power to go out to sea with large models?”The above-mentioned smart computing person said, don’t just focus on the domestic market.& rdquo;
“In the last wave of applications, frankly speaking, the cost performance of domestic computing power was not high.& rdquo; The above-mentioned person further said that now that algorithm models have achieved lower computing power costs through innovation, domestic computing power should also be reduced in price. He believes that this is also the application side that forces computing power.
The large model will also force the chip to be redesigned. DeepSeek’s paper or related reports have mentioned the requirements for chip redesign. For example, DeepSeek’s technical architecture and optimization strategies provide new ideas for the design of SoC chips, are also promoting the development of edge computing chips, and further promoting customized AI chips.
“The technology iteration itself has moved from universal to dedicated, and then to universal…& rdquo; Some intelligent computing people say that when it is general, technology develops faster in an open way across the entire AI industry chain, while when it is dedicated, it is pursuing efficiency, cost, etc. in normalized scenarios.
Competition will become more intense in the future
After the launch of DeepSeek, competition in the industry intensified.
“DeepSeek will stimulate a large number of innovative small and medium-sized enterprises to try to develop and deploy some smart applications in the short term. In some scenarios, it is possible to spawn popular smart applications. The rapid commercialization of these popular models will bring about a lot of consumption in the reasoning market. rdquo; Some people analyzed it.
But the industry admitted that AI has been hyped into a very expensive thing by American companies and is ready to be harvested. Now DeepSeek has jumped out and achieved extremely low costs, and a large group of people are rushing forward and trying. But the AI business application level is still in its early stages,There are too few talents in various industries who understand business and AI technologyIf there is no popular AI application in the short term, everyone can only try it and experimentally toss it out.
At the same time, can subsequent versions of DeepSeek continue to maintain the same evolution speed as overseas models? While DeepSeek is making progress, competing companies are also studying.
When Musk recently commented on DeepSeek, while admitting that DeepSeek is great, he also said that soon you will see that companies like xAI will release more advanced models than them. When these companies quickly integrate DeepSeek technology and support from leading computing power, evolution will accelerate.
Competition will become more intense in the future, but DeepSeek’s innovation is stimulating and driving the rapid prosperity of domestic industries. A person engaged in smart computing found that now a primary school student can use DeepSeek to create Weixin Mini Programs, and he also made a Tetris game himself.” ldquo; If everyone can access it and run it, there will be a way out. It’s just time.& rdquo;
