Just three working days after the holiday, the AI industry has seen several new developments. We have selected the three most important and relevant ones to share our insights.
Tech giants are exploring new pathways, next year's AI budgets are setting directions, and domestic open-source models face new thresholds for going global. Let's start with the two distinct approaches of Google and OpenAI.
01. Tech Giants Explore New Pathways, but the Problem of Insufficient AI Demand Remains Unsolved
On October 7, OpenAI integrated charts, calculators, and forms into ChatGPT using Intelligent UI. A day later, Google connected the Gemini Agent to Gmail, Docs, Sheets, and Calendar. OpenAI continues to strengthen the ChatGPT entry point, while Google embeds agents into its existing product ecosystem.
Behind these strategies lies their foundational resources. Google has Workspace, cloud services, and enterprise data; OpenAI has over 1.2 billion weekly active users. A similar divergence exists domestically: Alibaba relies on Alibaba Cloud, DingTalk, and enterprise customers, while ByteDance is betting on Doubao as a personal entry point.
We are more optimistic about the approaches of OpenAI and ByteDance. Insufficient demand remains the primary challenge for AI implementation. Enterprises are willing to pay for labor savings and process shortening, and coding can also be profitable, but the long-term value of mere efficiency improvement is limited. Compared to killer applications like e-commerce and short videos during the mobile internet era, the real major opportunity for AI still awaits the emergence of new consumer products.
Google and Alibaba are taking the enterprise efficiency route, while OpenAI and ByteDance are taking the personal entry point route. The former relies on existing enterprise user business needs, making it easier to generate revenue first, but this route has a clear ceiling; the latter relies on the high-frequency usage of massive individual users, providing a better chance to test and discover new consumer scenarios.
02. Focus of Next Year's AI Budgets is Set; AI Vendors Should Plan Ahead
In 2027, the AI budgets of the government and central state-owned enterprises will focus on pilot test bases. In October, central documents listed pilot test bases and high-value application scenarios as key directions. At this time, the government and central state-owned enterprises are formulating next year's plans and budgets. The policy clarifies the direction before the budgets are finalized, making pilot test bases a priority for project reserves and funding arrangements.
The policy focus has shifted from promoting AI to driving AI into industries and achieving large-scale implementation. In 2024, "AI+" was proposed; in 2025, targets for the penetration rate of smart terminals and agents were set; in March 2026, commercial and large-scale applications were required; and by October, pilot test bases and high-value scenarios were further specified. Policy directives are increasingly pointing towards specific business outcomes.
Table 1: Changes in the Policy Focus of "AI+" from 2024 to 2026
Pilot test bases target exactly the pain points of AI implementation in government and enterprises. Many projects have been carried out over the past two years, but most have stopped at peripheral scenarios that are easy to showcase. Core businesses are left untouched due to the fear of uncontrollable risks. Local governments, leading enterprises, and research institutions jointly build pilot test bases, running business processes with real data and integrating them into production systems only after verifying stability. Multiple enterprises sharing the platform can also reduce redundant investments.
For AI vendors, the construction, operation, and core business verification of pilot test platforms represent key opportunities to secure government and enterprise budgets in 2027. The Ministry of Finance has initiated the budget preparation for central departments, while local governments and large enterprises are also finalizing next year's investment plans in the fourth quarter. Central documents clarify the direction before budgets are finalized, and local project reserves will follow up as a priority. Vendors should also arrange their business priorities for next year accordingly.
03. Domestic Open-Source Models Going Global Must Factor in Security Costs
This AI attack in South Korea will lead overseas customers to conduct stricter procurement reviews of domestic open-source models. On October 7, the attack records published by CrowdStrike included DeepSeek and Zhipu GLM. Attackers invoked these models through the open-source penetration testing agent ARTEX. South Korean media reported that data breaches occurred in at least seven financial institutions, affecting approximately 68,000 customers.
Open-source models are at a greater disadvantage in this regard. Closed-source services can record calls, ban accounts, and cut off interfaces upon detecting anomalies; however, once an open-source model is downloaded, it is difficult for the original developer to know who deploys it where, nor can they control its modified usage. Models can be open, but the risk of abuse does not disappear as a result.
The European Union will begin regulatory enforcement on general-purpose AI models from August 2026. Open-source models cannot be exempt from all compliance obligations, and the South Korean incident provides overseas financial, government, and enterprise customers with further reasons for stricter reviews. Vendors need to supplement security evaluation and abuse monitoring materials to explain how incidents will be handled. An extra layer of review and a slower step in integration mean additional costs.
Domestic models have entered the overseas developer market through open source. In the past, the competition was mainly about performance and price; going forward, they must also compete on how security risks are managed. Good performance and low prices remain advantages, but if they fail security reviews, customers will be reluctant to pay. Open source remains an important means of going global, but vendors cannot only calculate the cost of model invocation.