2025年初人工智能格局报告：推理模型、主权AI及代理型AI的崛起

The AI Landscape of Early 2025: Reasoning, Sovereignty,And the Rise of Agentic AI Jeongkyu Shin, Lablup Inc. Table of Contents Chapter 1The New Frontier: Reasoning Modelsand the Test-Time Compute Revolution2 Chapter 2China's Rise: DeepSeek Shock andChina's AI Trajectory4 Chapter 3The AI Infrastructure War:Power, Cooling, and Geopolitics6 Chapter 4The Silicon Engine: Accelerators, NPUs,and the Journey Toward Independence7 Chapter 5The Rise of Sovereign AI:Age of Competition, National Strategies9 Chapter 6Innovators: Startup Approaches andBig Tech Power Dynamics11 Chapter 7The H1 2025 Frontier Model Landscape13 Chapter 8The Proliferation of Coding AI:From Assistant to Autonomous Teammate14 The AI Landscape of Early 2025:Reasoning, Sovereignty,And the Rise of Agentic AI Executive Summary As we reach mid-2025, the artificial intelligence (AI) ecosystem is undergoing a fundamentalparadigmshift amid intensifying geopolitical competition.This analysis examines the majortransformations from the second half of 2024 through July 2025, aiming to provide guidance fornavigating this rapidly evolving landscape. The most significant change in the first half of this year has been the shift in how AI models improvetheirperformance—from the capital-intensive'Train-Time Compute'approach to'Test-TimeCompute,' which invests more computation during inference to induce deeper reasoning. Thistransformation has catalyzed the emergence of 'Reasoning Models' such as OpenAI's o1 and China'sDeepSeek R1, creating economic ripple effects that lower the barriers to entry for frontier modeldevelopment. This became the technical backdrop for the 'DeepSeek Shock' of January 2025, whenDeepSeek's claims of achieving frontier performance at remarkably low cost sent shockwavesthrough global markets. This event elevated the U.S.-China technology competition to a newdimension, demonstrating that Chinese AI capabilities had reached world-class levels in softwareand algorithmic efficiency despite hardware sanctions. This competition is spreading into an infrastructure war. Data centers are transitioning to liquidcooling systems to handle power densities exceeding 250kW per rack, with nuclear energy beingmobilized to secure stable power supplies. This signifies the emergence of energy sovereignty as anew geopolitical variable directly linked to AI sovereignty. In the AI accelerator market, NVIDIA isconsolidating its dominance through integrated 'AI Factory' platforms with its Blackwell Ultra andRubin roadmap, while AMD challenges with cost, availability, and openness as its weapons throughthe MI400 series and open ROCm software stack. From a geopolitical perspective, 'Sovereign AI' has emerged as a core national security agenda.Major nations including the United States, United Kingdom, France, and Japan have announcedmassive investment plans to control their own AI models, infrastructure, and data, while South Koreaaims to leap into the global AI top three with a 100 trillion won investment plan. Meanwhile, the first half of 2025 saw market differentiation with the release of frontier modelsspecialized in specific domains, including Claude 4, Gemini 2.5 Pro, and Llama 4. Competition forleadership is intensifying particularly in high-value areas such as coding, scientific reasoning, andmultimodal capabilities. In the coding AI sector, autonomous 'AI teammates' like GitHub Copilot Agentand Devin are fundamentally changing developers' roles, though a 'productivity paradox' has beenobserved where actual skilled developers' productivity decreases despite impressive benchmarkperformance, presenting new challenges. In conclusion, the 2025 AI ecosystem stands at an inflection point where technological innovation israpidly reshaping economic and geopolitical landscapes. The universalization of reasoning models,the rise of sovereign AI, the intensification of infrastructure wars, and the proliferation of agentic AIwill serve as both threats and opportunities for all enterprises and nations. Chapter 1 The New Frontier: Reasoning Models and the Test-Time ComputeRevolution 1.1. Paradigm Shift: From Train-Time to Test-Time Compute Until mid-2024, performance improvements in large language models (LLMs) primarily came throughwhat's known as 'Train-Time Compute'—scaling three key elements: model size, dataset size, andtraining computation.[¹] While highly effective, this approach became increasingly untenable aspre-training model scales grew exponentially, with training costs reaching billions of dollars. This cost barrier triggered a fundamental paradigm shift in AI development. Research labs likeOpenAI and DeepMind confirmed that the scaling laws previously applied to training also apply to theinference stage.[¹] This means models can "think" more deeply and solve complex problems through'Test-Time Compute'—investing more computational resources during the inference process whenmodels answer questions. This discovery prompted AI developers