Guoxin Securities: Reasoning side demand is expected to explode in '26, and the office scene is expected to usher in more AI products

The Zhitong Finance App learned that Guoxin Securities released a research report saying that inference demand is expected to explode in '26. In the primary market, programming scenarios and agent explosions are the main application directions. Judging from the number of users and the revenue valuation level of startups, the current rapidly growing industry is mainly AI programming, AI agents, and AI content creation, focusing on productivity improvement. Since this year, many popular applications have been created, and the office scene is expected to launch more products next year. Furthermore, the bank believes that as model capabilities mature, next year we will see significant growth in the fields of AI phones, AI glasses, and distributors that help big models land in the enterprise.

Guoxin Securities's main views are as follows:

Reviewing the stock price trends of US tech giants over the past three years, AI narratives continue to advance

In '23, OpenAI led the world in starting AI acceleration. Microsoft benefited from OpenAI's exclusive cooperation, and valuation rose significantly. In 24, the market underestimated the room for model advancement, and the narrative turned to the speculative side, believing that the app company was the best. Meta had a social monopoly ecosystem (potential agent entry) and advertising scenarios, and the stock price was the only PE giant other than Nvidia to rise. (In February '24, Nvidia's results estimate that about 40% of data center revenue comes from reasoning.) In the same year, cloud vendors experienced delays in cloud revenue transmission due to sharp increases in capital expenditure but limited supply, and the valuations of the three major CSP fell slightly. The 25-year model gap clearly converged with OpenAI. Google later took the lead, and the ecological advantage was the market chase. In '26, the bank believed that the Scaling Law would continue, model makers would open up a differentiated application market, and demand on the model inference side may enter an inflection point in volume. The model and computing power may be the optimal investment direction.

The four giants Capex grew by more than 50% year on year in '25, and the bank estimates that Capex will continue to achieve a growth rate of more than 30% in '26

The four North American giants Capex continued to upgrade in '25, from an increase of 320-330 billion US dollars at the beginning of the year to close to 400 billion US dollars at the end of the year. Each company's annual Capex investment increased by more than 50% year-on-year. Huge amounts of Capex are investing in data center construction, or facing power bottlenecks. The North American data center capacity was about 25 GW in '24. According to Grid Strategies estimates, 80 GW of demand will be added in the five years up to '29. Considering the decommissioning of coal power and the long construction cycle for supporting transformers, etc., the power gap is expected to become a major contradiction. Therefore, in the process of building data centers, the computing power to energy consumption ratio may become a key consideration factor for giants.

The model architecture continues to evolve, scaling laws continue, and multi-modals+long text provide the foundation for agent outbreaks

23 years ushered in the Scaling Law dividend period, 24 years of multi-modality and inference models emerged, and 25 years of algorithm engineering went hand in hand with Scaling Law. In the long run, achieving AGi still requires breakthroughs in model architecture and scaling to the limit. Looking ahead to 26 years: 1) In terms of architecture, the next-generation model architecture currently needs to address two core pain points: ① the computing volume and memory consumption bottlenecks of the Transformer during the training phase are becoming more and more prominent; ② the model's memory ability is limited during inference, and model parameters cannot keep up with memory changes. Overseas, Google's Titans architecture and Mamba architecture already exist, while domestically, from the perspective of cost efficiency optimization, including Qwen3-Next and DeepSeek V3.2, have achieved significant improvements. 2) In terms of scaling, it is expected that the scaling law will continue in 26 years, including from pre-training to post-training and inference scenarios, while reinforcement learning will become a key breakthrough direction in the future; 3) Multi-modal and long text capabilities are more mature, which provides a technical foundation for the emergence of agents. Currently, the gap between the Chinese and US models is 3-6 months, and computing power and algorithms are the key to catching up.

GM's big model capabilities have yet to be won or lost, and there are differences in the commercialization paths of manufacturers

1) OpenAI: Although short-term model capabilities have been overtaken, the performance of the next generation model is still worth looking forward to. 800 million C-side users are the core barrier, and the company will also boost corporate business next year; 2) Gemini: currently the SOTA benchmark for the big model, thanks to the adherence to the native multi-modal valuation route and the ecosystem of self-developed chips, the consumption of tokens is expected to continue to increase dramatically next year; 3) Anthropic: Adhere to the 2B route, and is expected to achieve better profit margin levels by relying on 2B's pricing advantages. At 350 billion US dollars, the AI programming product ARR launched in early '25 also reached 1 billion US dollars; 4) Grok: Believes in vigorous miracles. Since inference scenarios are limited, training computing power is sufficient. Combined with Tesla's unique data advantages, the next generation of native multi-modal models is worth looking forward to.

The model lowers the threshold for software development, opens the demand ceiling, but players are reshuffling the cards

In '25, tokens were used more to restructure large model companies' internal and recommendation systems, but starting in '26, demand for downstream applications is expected to continue to increase. AI actually opened the ceiling of software demand. According to IDC data, the global SaaS market is expected to reach nearly 1 trillion US dollars in 2029 (a significant increase compared to 580 billion US dollars in 25 years), but the bank believes that players will reshuffle their cards. Industries with data barriers, mainly arranged in vertical segmented scenarios, where software-defined workflows are complicated or require extremely high accuracy, are less at risk of being replaced by larger models, such as healthcare, energy, accounting, and security. At the same time, the bank observed that large model manufacturers have begun to develop more industry needs through cooperation with B-side software service providers, or compete head-on with traditional SaaS vendors.

Risk warning: Macroeconomic fluctuations, downstream demand falling short of expectations, risk of core technology upgrades falling short of expectations, and AI affirmative competition increasing the risk of cloud business profit margins.