Altman pulls out GPT-5, then look at DeepSeek and Liang Wenfeng

Written by

Clara Bennett

Updated on:July-17th-2025

People seem to have forgotten about GPT-5 . Yes, we also wrote an article last year saying that GPT-5 has been shocked . However, in our outlook for 2025 , we believe that the progress of the basic model will still ultimately determine whether AGI can be achieved . If there is no GPT-4.5 or 5 in 2025 , or a large model of a new paradigm emerges, from the national treasure DeepSeek to the various reasoning models in Silicon Valley, they are all quantitative changes at the level of GPT-4o and the o1 paradigm. The only difference between the two is whether it is open source and the cost-effectiveness.

In the past two days, Altman talked about GPT-5 at the Berlin University of Technology stop on his global tour . He asked the students present: "Who else dares to say that they are smarter than GPT-5 ?"

He continued: “I don’t think I’m going to be smarter than GPT-5 . I don’t feel bad about that because I think it just means we’re going to be able to do incredible things with it. You know, we want to get more science done. We want to enable researchers to do things they couldn’t do before.”

A week ago, Altman said in a Reddit interview that there was no timetable for GPT-5 . But he has talked about GPT-5 publicly several times recently . He believes that the changes from 2025 to 2027 will be greater than from 2023 to 2025. This still requires huge investment and computing power.

OpenAI 's new round of financing is linked to the Stargate project . Masayoshi Son's SoftBank will invest $ 40 billion, bringing OpenAI's post-investment valuation to $ 300 billion. SoftBank is likely to replace Microsoft as OpenAI's largest shareholder. Then OpenAI and SoftBank will each invest $ 19 billion in Stargate . Stargate needs to invest $ 100 billion this year and $ 300 billion in the next three years . The annual investment exceeds that of any technology giant.

In the recently released financial reports, Microsoft, Google, Amazon and Meta have significantly increased their capital expenditure expectations for the next fiscal year, totaling $ 320 billion, far exceeding last year's $ 250 billion. Amazon is about $ 100 billion, Microsoft is about $ 80 billion, Google is about $ 75 billion, and Meta is about $ 65 billion. In other words, in the next four years, OpenAI plus the above and other technology giants' investments in AI will reach $ 2 trillion.

Don’t forget Ilya Sutzkever, former chief scientist of OpenAI, one of the most important AI scientists today, who still has the potential to change the paradigm of large models. SSI (Safe Superinterlligence Inc.), founded by him, is launching a new round of financing with a valuation of $20 billion. After declaring the scaling law of pre-training dead at the end of last year , he believes that “AI can only defeat humans if it becomes unpredictable.”

There is also Anthropic, which is expected to be oversubscribed, raising more than $2 billion from venture capital institutions such as MGX from the Middle East, with a valuation of more than $60 billion. The company's CEO has recently commented frequently on DeepSeek , believing that many companies can now train excellent inference models, and as the basic model continues to develop upward along the expansion curve, this temporary window will close; as for whether DeepSeek can continue to catch up, it depends on how many more chips it can obtain than now.

One thing is certain, OpenAI is still raising huge amounts of money, purchasing hundreds of thousands of the latest NVIDIA GPU cards, and building computing clusters that are unprecedentedly complex in communication and engineering to prepare for training the next generation of basic models. Super computing clusters still play a decisive role in training the most basic cutting-edge models of general intelligence.

If the United States relies on its capital, computing power, and top talent density to break through the most cutting-edge AI technology, while China takes a cost-effective, rapid popularization and application path to form an industry, the two will complement each other and be more conducive to the development of AI . Coupled with geopolitical factors, a competition between the two routes has emerged.

And the market has begun to seriously evaluate these two routes. DeepSeek seems to be making investors re-evaluate the capital markets in China and the United States. Since February , China's large and small cloud service-related listed companies have been sought after by funds, while American giants have been more strictly scrutinized: Are American companies trapped in the GPU Rich Curse, not making efficient use of computing resources, and lagging behind research and infrastructure in application?

China is running into the era of artificial intelligence. Almost every company is trying to get involved with DeepSeek. More than 16 domestic AI chip companies have adapted or launched DeepSeek model services, and there are more than 22 cloud computing and intelligent computing infrastructure companies, including the three state-owned operators, technology giants, as well as small and medium-sized cloud and computing service providers.

Application-layer companies have begun to embrace DeepSeek in addition to their own self-developed models ; companies that have been watching the excitement of AI also require each system and department to develop an operational and quantifiable application plan based on their own business needs and include it in the assessment of cadres. Many securities companies have completed the local deployment of the R1 model, covering business lines such as research, compliance, marketing and coding.

Goldman Sachs and UBS have both launched a " China AI Package " to introduce Chinese chip, infrastructure, and application-related companies to their clients. Deutsche Bank even called DeepSeek " China's Sputnik moment " and will expand the scope of promotion to all high-value-added fields with intellectual property rights.

Investment institutions are much more demanding of American companies. DeepSeek has become a whip for analysts to whip the seven giants of the US stock market, and it is the most frequently asked keyword in the earnings conference. But they all believe that the emergence of DeepSeek is a positive, lowering the threshold of AI will accelerate application and innovation, and reasoning will increase the value of companies under the " Jevons paradox " of computing power . But the market is still skeptical about this.

Economist Paul Krugman ca n't help but ask, is the US overreacting like it did in 1999 ? He thinks it's even worse than last time, because this time the market is abnormally expecting disruptive technology to continue to consolidate the market structure of these monopolies. However, Krugman also famously predicted that the Internet's impact on the economy will not exceed that of the fax machine. Gary Marcus, a big short of generative AI , said that you should listen to what companies don't say, not what they say. " If Google has made a lot of money from artificial intelligence, it will definitely mention it in the conference call with investors. "

American companies still follow the innovation path of " violent aesthetics " in the field of large models . Before the next generation of basic models are truly launched, no one dares to slack off. This has led to an increasing concentration of funds in a few powerful leading players. Only they can purchase chips with the most advanced processes in bulk, handle data centers with ultra-large-scale clusters, power stations, power grids and supporting energy storage, and deal with intellectual property lawsuits from data producers.

In just two years, the electricity consumption of a single intelligent computing cluster in the United States has increased tenfold . Dylan Patel of Semianalysis estimated last year that AI would account for 10% of the electricity consumption in the United States, which was considered a fantasy, but the feedback he got from OpenAI and Anthropic was that this was far from enough. Almost every giant has abandoned its carbon neutrality commitment , and xAI and Meta have embraced natural gas power generation. " Screw the sustainable development goals, " he quoted industry insiders as saying, " because this competition is too important to lose. "

Before DeepSeek came out, chip companies, cloud giants and energy-related companies on the US stock market repeatedly reinforced this narrative logic. They are all tied to the same rope. What is lacking, what will rise. First it was Nvidia, then data centers, and then it was the turn of power facilities and nuclear power companies.

However, the market has always been concerned about the "600 billion US dollars " issue. How can the huge infrastructure investment be recouped in terms of application returns? If the large model is implemented later, the monetization cycle will be longer and the financial pressure will be greater. The " Stargate Project " has also raised a " 500 billion US dollars" question mark, which is almost self-proven based on the assumption that AGI will be achieved in the next four years .

Before DeepSeek came out, the narrative logic of the Chinese market also revolved around scarcity. Advanced process chips are the biggest shortcoming. Cambrian's market value growth in the past two years has even exceeded that of Nvidia. In addition, Chinese investors also frequently hype up-stream equipment, consumables and foundry in the chip industry.

However, the market's concern about this narrative logic lies in the fact that an ecological closed loop cannot be established: the foreign cutting-edge models are encapsulated, and the API can be discontinued at any time; it relies on advanced process chips and advanced equipment, and export controls are constantly upgraded; domestic chips, domestic models and domestic applications cannot be adapted one by one.

This seems to be the case. The question of whether there is a bubble in the intelligent computing data center has become a footnote to the hidden concerns of this narrative logic. For a long time, the market was confused about whether China was short of computing power or had an excess of computing power. Both existed almost at the same time.

On the one hand, the US government believes that its chip export controls have caused China's large-scale model enterprises to be short of computing power, and they have to secretly rent intelligent computing power from Southeast Asia and other countries, and even "physically smuggle" advanced process chips. On the other hand, in this round of large-scale model wave, China's intelligent computing data center projects have been launched in large numbers, and the total scale approved by various regions far exceeds the national plan . The media has also repeatedly exposed that China's data center computing power is idle, and there are too many intelligent computing centers, even the large models in the "100 Model War" are not enough.

There are many analyses in the industry, and the reasons are summarized as follows: First, China's infrastructure construction often " moderately advances investment " , and by 2024 , the large model training market has gradually converged to at least a few companies, and reasoning has not been fully implemented; second, the early market was chaotic, many participants lacked technology and experience, and the location and scale of data centers did not match the computing power requirements of the " brute force " training of the next generation of large models; third, domestic chips are " usable " but not " easy to use", and Nvidia's H20 is a bit useless.

According to conservative statistics obtained by Caixin from supply chain interviews, in 2024 , there will be about 700,000 NVIDIA H20 series chips and 300,000 domestic mainstream AI chips in China ; SemiAnalysis estimates that the former will exceed 1 million and the latter will exceed 550,000 .

DeepSeek seems to be activating them. Dai Guohao, co-founder of Wuwen Xinqiong, believes that DeepSeek's innovation in underlying optimization and collaborative optimization has finally filled the closed loop of " model - system - chip " domestic substitution. Wuwen Xinqiong is a computing infrastructure company that was born in early 2023. It provides heterogeneous computing solutions for the current situation of "M models " and "N chips" in the market, and is very sensitive to system architecture innovation. The DeepSeek V3 model paper spent the most space to introduce the system architecture.

Filling the closed loop of domestic substitution means filling the gap in the market valuation of Chinese companies: reasoning has already arrived, and it does not only rely on ultra-large-scale computing clusters, domestic chips will be equally useful; and open source allows any chip, cloud infrastructure and application to be actively adapted and called. On the contrary, Cambrian has not yet cooperated with DeepSeek 's public promotion, and its stock price has lagged behind recently; however, the Nanjing Intelligent Computing Center and the Southeast Zhejiang Intelligent Computing Center, which are supported by Cambrian, have begun to support related models.

The United States is still planning to increase chip export controls on China. Nvidia originally planned to export 2 million H20 chips to China this year , but according to SemiAnalysis , all these orders have been canceled, and there is no news about the next-generation B20 orders. Recently, the channels for a group of Chinese chip design companies to TSMC for chip manufacturing have been further blocked. The optimization of DeepSeek has changed domestic chips from " usable " to " easy to use " , but the pace of continuous model upgrades will be very fast. Domestic chips and their ecosystems must make full use of this time window for adaptation and commercialization.

The big model competition between China and the United States seems to have entered a familiar track: the United States leads the cutting-edge technology from 0 to 1 , and China follows closely to innovate applications from 1 to N with cost-effectiveness . However, DeepSeek is catching up too closely this time, which makes innovators confused for a while: since the technology developed with huge investment may be reproduced at one-tenth of the price in a few months, and the whole world can use it for free, is the huge investment in innovation worth it, or will it become a target? Ultraman is reflecting and will make adjustments.

In this sense, DeepSeek and a group of AI technology companies have allowed China to regain a familiar and effective feeling in an emerging field. It is no longer just a passive follower, but has its own route.

At the same time, the launch of large models and intelligent bodies such as the free version of o3 mini , Gemini 2 Flash Thinking , and Deep Research have begun to crush DeepSeek again in terms of performance and even price , not to mention the upcoming launch of open source Llama 4 , which aims to establish a true open source standard worldwide. DeepSeek itself has also exposed prominent problems such as "hallucinations".

Next, can the next-generation basic model, GPT-5 , once again achieve an order of magnitude of transcendence with the support of brute force computing power? Can the cost-effective system architecture optimization successfully replicate the competition in the next-generation model? And can domestic chips keep up with the computing power requirements of large-scale model evolution and application in terms of computing power performance innovation and large-scale mass production in a commercial environment?

If the United States has already established an ecosystem and industrial chain axis such as OpenAI-Microsoft-Nvidia-TSMC-ASML, then it will depend on whether DeepSeek and Liang Wenfeng have enough motivation and resources to create an ecosystem and industrial chain.