GTHT has released a research report maintaining an "Overweight" rating on the computer sector. The National Development and Reform Commission is increasing investment in "AI+" infrastructure, while major domestic and international players are initiating a wave of price hikes for computing power and large models, validating that controlling scarce high-end computing resources and possessing pricing power have become the new critical battleground in the industry. DeepSeek has initiated its first funding round to advance the migration of its V4 hardware and update its underlying Mega MoE architecture, signaling that AI competition is entering a phase focused on ecosystem independence and hardware-software restructuring. Separately, Anthropic has launched its flagship Claude Opus 4.7 model, accelerating the evolution towards a full-stack AI application ecosystem and leveraging its asymmetric model advantages to exert significant pressure on third-party application layer tools. The key points from GTHT are as follows.
The National Development and Reform Commission is boosting investment in "Artificial Intelligence+" infrastructure, and a computing power price surge is sweeping through major players both in China and abroad, indicating the AI industry is entering a new cycle focused on foundational development and a re-evaluation of computing power's value. This week, the Commission clarified that this year it will prioritize a series of actions to expand effective investment in areas such as "AI+" infrastructure and will comprehensively implement "soft infrastructure development" in central government investment projects, utilizing the national venture capital guidance fund to mobilize social capital for supporting emerging industries. Concurrently, Alibaba Cloud issued three price increase notices within just four days, raising prices for its Bailian large model unit by 2%-5%, increasing fees for its DDoS high-defense service by 50%, and transitioning DataWorks API from unlimited free calls to a pay-as-you-go model. Overseas giant Anthropic also changed its Claude enterprise subscription from a maximum of $200 to a model based on "actual computing consumption plus a $20 fixed fee," leading to costs doubling or even tripling for heavy users. This is compounded by a nearly 40% surge in contract prices for H100 rentals from previous lows, exacerbating the global shortage of computing power. These two trends point to the same conclusion: the business model for computing power is transitioning from simply selling raw capacity to selling tokens. Controlling scarce, high-end computing power and having the ability to command premium prices are now the core strategic resources and new competitive frontiers that AI companies are vying for.
DeepSeek has initiated its first funding round to advance V4 hardware migration, with its underlying Mega MoE pushing the limits of computing power, indicating that AI competition is moving into a period defined by ecosystem independence and hardware-software co-redesign. This week, DeepSeek, breaking its previous "no financing" rule, is seeking $300 million in funding at a valuation of approximately $10 billion. The primary goal is to facilitate the complete migration of its trillion-parameter, natively multimodal V4 model away from Nvidia's architecture and onto Huawei's CANN framework. Simultaneously, its underlying infrastructure, DeepGEMM, quietly updated its Mega MoE, integrating fragmented computational processes into a single Mega-kernel to achieve computation and communication synchronization and employing FP8×FP4 precision to maximize efficiency. These two major moves indicate that the battle in advanced AI is extending to the foundational layer. Breaking free from hardware dependencies and achieving extreme optimization of infrastructure have become the new decisive factors for success.
Anthropic has released its flagship Claude Opus 4.7 model and is accelerating its evolution towards a full-stack AI application ecosystem. This week, Anthropic launched the Opus 4.7 version, which achieved the highest scores globally in tests for advanced software engineering and financial analysis agents, features a visual recognition capability with a side length of 2576 pixels, and includes an upgraded tokenizer. Furthermore, the upcoming Claude Code update will introduce a unified interface across code repositories, signaling the company's shift from being a wholesale provider of底层 API computing power towards becoming a provider of high-margin solutions. By leveraging the asymmetric advantage of its model capabilities, it has successfully exerted deep competitive pressure on third-party application layer tools like Cursor.
Risk warnings include technology development falling short of expectations and company business expansion proceeding slower than anticipated.
Comments