In the past 24 hours (January 14 to January 15, 2026), the main line of global AI competition has further shifted from "building models" to "spelling computing power, data, and landable agents". Overseas actions around computing power supply, chip trade and platform security governance are frequent; In China, it has continuously launched new products in the direction of consumer-grade AI services, industrial agents, localized multi-modality and data bases.
1. OpenAI and Cerebras have reached a computing power cooperation of more than 10 billion US dollars
OpenAI signed a multi-year agreement with Cerebras to purchase up to about 750 megawatts of computing power, focusing on online services for inference and "inference models". This move sends a signal: inference-side computing power is becoming a new bottleneck, and infrastructure diversification will accelerate. For the industry, dedicated acceleration chips and cloud service providers are expected to get more orders from major customers.
2. Microsoft launched a "community first" data center plan, promising not to let residents pay for AI electricity
Microsoft has released a series of commitments for U.S. datacenters, including support for data centers to cover full power and new grid costs, and disclosure of water use and recharge progress by region. As the expansion of AI data centers has led to controversies over electricity prices and resources, large manufacturers have begun to trade more specific cost and transparency commitments for landing space. This trend may lead to the formation of new infrastructure rules in which energy consumption can be measured and costs can be shared.
3. The United States imposes a 25% tariff on some high-end AI chips and related equipment
The United States announced tariffs on high-end chips that meet certain performance thresholds and equipment containing them, and explained that there is room for exemptions for some data centers and other uses. In the short term, it will increase the uncertainty of cross-border supply chains, and may also prompt enterprises to accelerate regional procurement and product allocation adjustments. In the medium and long term, trade tools are more directly involved in the AI computing power race.
4. Alibaba Qianwen App has been upgraded: move shopping, payment, and travel "services" into the dialog box
Alibaba has made important upgrades to the Qianwen App, opening up capabilities such as e-commerce, payment and local life, supporting orders and payments in chat, and launching a test function for task-based assistants. The key to the transition of consumer-grade AI from "chatting" to "doing things" lies in whether it can stably call real services and conduct closed-loop transactions. For ecological platforms, this is also an accelerated sprint to compete for the "super entrance".
5. The new version of Baidu Wenxin Model 5.0 has risen in the ranking of the public arena list
The latest version has entered the top of the world in the open text list, and the math ability ranking is also in a leading position, and it was revealed that more progress will be announced in the near future offline event. Behind the popularity of the list, the focus of competition is shifting from a single release to continuous iteration and verifiable evaluation. For enterprise users, it is more important to improve stability, cost and tool chain support.
6. Midea released the industrial intelligent agent matrix, emphasizing that "entering the workshop" brings real cost reduction and efficiency increase
Midea's digital subsidiary released more than 40 industrial agents and launched a new version of the AIGC platform, disclosing the savings targets and phased data brought by AI applications. The focus of the manufacturing industry is no longer "how many agents", but whether it can be embedded in key processes such as production, supply chain and quality to form a closed loop. The threshold for the scale of industrial agents is still data quality, talent and scenario selection.
7. Zhipu cooperates with Huawei to train domestic multi-modal mapping models, focusing on low-cost and commercialized
The two sides cooperated to train and launch an image generation model, emphasizing the full-link localization technology stack and lower call costs, and providing lightweight deployment ideas. The significance of this kind of cooperation is to verify the support ability of domestic computing power and framework for complex multimodal training. For small and medium-sized enterprises, "commercializability, low threshold" will directly affect the speed of adoption.
8. AI data base continues to increase: storage vendors release data solutions for training and inference
Domestic storage manufacturers have released data solutions for AI scenarios, focusing on the cost pressure brought about by training, reading, writing, cross-domain data flow, and long inference context. An industry consensus is forming: if computing power improvement cannot keep up with data supply, GPU utilization will be dragged down by "data waiting". Engineering capabilities around parallel file systems, global namespaces, and hierarchical storage will receive more attention.
9. xAI tightens Grok's image editing capabilities, and platform security governance continues to be upgraded
xAI announced restrictions on Grok's specific editing methods for real images for all users and strengthened content protection against the backdrop of rising regulatory pressure. Generative AI's "image editing" is becoming a high-risk area for compliance, and the platform needs stronger rules, interception and auditing capabilities. For product teams, security capabilities will change from "checking before go-live" to "continuous operation after go-live".
10. The tight supply of storage and HBM is driven by AI, and the cost of consumer electronics may be affected by spillover
Industry analysis pointed out that the demand for high-bandwidth storage in AI data centers is squeezing the supply in other fields, and some manufacturers have locked in future production capacity in advance. If the price of key storage continues to be high, the cost of terminals such as mobile phones and computers may passively rise, which will affect the pace of shipments. For AI companies, optimizing inference memory usage and improving data efficiency will be more "valuable".
Frequently Asked Questions (Q&A)
Q: What are the core industry signals in this issue?
A: Competition is expanding from "model capabilities" to "intelligent agents with computing power supply, data efficiency and closed-loop capabilities", and the importance of infrastructure and ecological integration has increased significantly.
Q: What is the difference between domestic and foreign focus?
A: China emphasizes the application closed-loop and industrial scenario implementation (service assistants, industrial agents, localized multi-modality and data bases), while foreign countries focus more on computing power supply, public cost allocation and platform compliance governance.
Q: How will tariffs and computing power constraints affect corporate decision-making?
A: Enterprises will pay more attention to supply chain diversification and regional deployment, and increase investment in model compression, caching and data pipeline optimization to reduce sensitivity to single hardware and cross-border uncertainties.
Q: Where are the opportunities for developers and startup teams?
A: Opportunities are concentrated in three categories: task-based agents that can call real services, data and storage infrastructure that improves training and inference efficiency, and compliance and security toolchains for image and content generation.
