AI 基礎設施轉型：推理時代的到來

前沿模型能力與基礎設施重構

Anthropic 的 Claude Mythos Preview 在 2026 年 4 月發布，標誌著前沿模型能力發生了質的飛躍。這不僅僅是模型性能的提升，更揭示了 AI 基礎設施正在經歷從「訓練為主」到「推理為主」的結構性轉變。

核心能力指標

Mythos Preview 在多項基準測試中展現了遠超前代模型的性能：

漏洞發現能力

• CyberGym 漏洞複現：Mythos Preview 83.1% vs Opus 4.6 66.6%
• SWE-bench Verified：93.9% vs Opus 4.6 80.8%
• SWE-bench Multilingual：87.3% vs Opus 4.6 77.8%

代理編碼與推理

• SWE-bench Pro：77.8% vs Opus 4.6 53.4%
• Terminal-Bench 2.0：82.0% vs Opus 4.6 65.4%
• GPQA Diamond：94.6% vs Opus 4.6 91.3%

關鍵成果

• 在 40 多組織的防禦性安全工作中部署 Mythos Preview
• 發現並協助修復 OpenBSD、FFmpeg、Linux Kernel 等關鍵系統中的漏洞
• 發現多個維護多年未發現的高危零日漏洞

這些指標不僅展示了技術能力，更揭示了基礎設施需求的重構：從「訓練為主」的計算模式，轉向「推理為主」的 24/7 連續運行模式。

基礎設施計算範式的根本轉變

訓練 vs 推理的基礎設施差異

NVIDIA 的 Vera Rubin 平台技術博客揭示了訓練和推理在基礎設施需求上的根本差異：

訓練工作負載

• 同步全對全通信階段
• 兆瓦級電力峰值
• 大規模 GPU 功率擺動
• 無緩解措施會導致電力網壓力、違反電網約束或強制運營商擴建基礎設施

推理工作負載

• 銳利的突發需求峰值
• 連續 24/7 運行
• 每個用戶查詢、推理步驟、API 調用都是推理負載

「AI 現在正嵌入到客戶服務、編程工具等產品中，推理需求實現 24/7 運行。這完全改變了基礎設施計算範式。」——NVIDIA GTC 2026

電力約束的硬性門檻

計算瓶頸

• AI 數據中心正面臨物理限制
• 單一晶片升級無法完全解決
• 需要全新的基礎設計方法

基礎設施投資模式

• 訓練：週期性事件，一次性大規模投資
• 推理：連續事件，持續性基礎設施投資

IBM 在 2026 年預測：「2026 將是前沿與高效模型類別之間的決定性一年」，這種轉變意味著基礎設施投資模式將從週期性訓練投資，轉向持續推理基礎設施建設。

技術架構演進：GPU/CPU 協同設計

Vera Rubin 架構的設計重點

NVIDIA Vera Rubin 平台專為 agentic AI 和推理時代設計：

核心目標

• 消除通信和內存移動的關鍵瓶頸
• 超級提升推理性能
• 每瓦更多 token，每 token 更低成本

性能指標

• 相比 Blackwell 架構：每瓦性能提升，每 token 成本降低
• 網絡存儲：每秒 token 數提升 5 倍
• TCO（總體擁有成本）：性能提升 5 倍
• 電力效率：提升 5 倍

部署實踐

• AWS、Google Cloud、Microsoft、OCI 在 2026 年部署 Vera Rubin 實例
• Microsoft 部署 NVIDIA Vera Rubin NVL72 機架規模系統
• CoreWeave、Lambda、Nebius、Nscale 等雲合作伙伴

電力與成本的硬性約束

約束場景

• 電力：實時推理必須在電力約束下運行
• 成本：需要控制推理成本
• 部署：需要實際可部署的架構

解決方案方向

• GPU/CPU 協同設計處理 agentic 工作負載
• 優化通信和內存訪問模式
• 適配 24/7 持續推理需求

商業模式與投資邏輯的變化

從「模型性能競賽」到「推理效率競賽」

前沿模型定位

• 訓練：週期性、高風險、高回報
• 推理：持續性、高可用性、運營優化

投資邏輯變化

• 從「訓練一次，服務長期」轉向「持續推理，優化運營」
• 基礎設施投資從「訓練中心」轉向「推理中心」

企業級部署策略

生產級 AI Agent 部署

• 目標：到 2026 年底運行 100+ AI Agent
• 每位員工配備 Agent 支援
• 端到端供應鏈的統一數據和治理基礎

供應鏈 AI Agent

• 自主系統跨供應鏈運作，無需人類觸發
• 持續優化供應鏈
• 動態個性化客戶體驗

量化影響

• 領先企業可實現 4 倍影響力，一半時間
• MIT 和 McKinsey 研究：統一數據和治理基礎可實現 4 倍影響力，一半時間

地緣政治與治理的戰略影響

前沿模型訓練與部署的競爭

監管環境差異

• 歐盟：權利為基礎的框架（EU AI Act）
• 中國：國家中心模式
• 美國：聯邦 AI 治理框架

戰略考量

• 選擇訓練地點 = 選擇監管環境 = 選擇部署模式
• 前沿模型可能被視為「關鍵基礎設施」而非「通用工具」

2026 的關鍵決策點

決策 1：訓練為主還是推理為主？

• 訓練：週期性、高風險、高回報
• 推理：持續性、高可用性、運營優化

決策 2：前沿模型如何監管？

• 歐盟 AI Act：分級監管
• 美國：聯邦監管
• 亞洲：國家級監管
• 關鍵問題：前沿模型是關鍵基礎設施還是通用工具？

決策 3：誰來制定規則？

• 聯合國全球對話：合作還是對立？
• 單邊監管還是多邊協調？
• 技術標準 vs 法律法規

硬性門檻與技術邊界

電力約束的硬性門檻

不可逾越的物理限制

• AI 數據中心正面臨電力物理限制
• 晶片升級無法單一解決問題
• 需要全新的基礎設計方法

投資約束

• 訓練投資：週期性、可擴張
• 推理投資：持續性、運營成本控制

技術架構的硬性邊界

協議與標準

• 通訊協議：需要優化全對全通信
• 內存協議：需要減少內存移動瓶頸
• 系統協議：需要適配 24/7 推理需求

部署邊界

• 雲計算：需要支持 24/7 推理
• 邊緣計算：需要低延遲推理
• 離線部署：需要自主推理能力

商業模式重塑：推理效率即競爭力

從「模型性能」到「推理效率」

競爭維度變化

• 從「訓練一個更好的模型」轉向「運營更好的推理系統」

商業模式重構

• 訓練成本：週期性、可預測
• 推理成本：持續性、可優化
• 基礎設施：連續投資、運營優化

企業級 AI Agent 商業化

AI Agent 類型

• 客戶服務 Agent：24/7 自動響應
• 編碼 Agent：持續代碼優化
• 供應鏈 Agent：自主運營

商業化路徑

• 訓練 → 推理 → 運營優化
• 從「模型性能競賽」到「推理效率競賽」

結論：基礎設施計算範式的不可逆轉變

2026 年標誌著 AI 基礎設施從「訓練時代」到「推理時代」的轉折點：

1. 能力層面：前沿模型能力已經跨越門檻，可達到甚至超越人類專家水平
2. 基礎設施層面：推理負載的連續性要求改變了基礎設計邏輯
3. 商業模式層面：從週期性訓練投資轉向持續推理運營
4. 地緣政治層面：訓練地點、監管環境、部署模式成為戰略選擇

這種轉變不僅是技術升級，更是基礎設計哲學的根本改變。企業需要從「訓練為主」的思維模式，轉向「推理為主」的運營思維模式。投資者需要從「訓練週期」的估值邏輯，轉向「推理運營」的估值邏輯。

硬性結論：AI 基礎設施的計算範式轉變不可逆轉，這將重新定義前沿 AI 的定價模式、投資邏輯和競爭維度。

#AI Infrastructure Transformation: The Coming of the Age of Inference

Cutting edge model capabilities and infrastructure reconstruction

The April 2026 release of Anthropic’s Claude Mythos Preview marks a quantum leap in cutting-edge model capabilities. This is not only an improvement in model performance, but also reveals that AI infrastructure is undergoing a structural transformation from “training-based” to “inference-based”.

Core competency indicators

Mythos Preview has demonstrated far superior performance to previous generation models in multiple benchmark tests:

Vulnerability discovery capability

• CyberGym vulnerability recurrence: Mythos Preview 83.1% vs Opus 4.6 66.6%
• SWE-bench Verified: 93.9% vs Opus 4.6 80.8%
• SWE-bench Multilingual: 87.3% vs Opus 4.6 77.8%

Agent Coding and Reasoning

• SWE-bench Pro: 77.8% vs Opus 4.6 53.4%
• Terminal-Bench 2.0: 82.0% vs Opus 4.6 65.4%
• GPQA Diamond: 94.6% vs Opus 4.6 91.3%

Key results

• Deploy Mythos Preview in defensive security efforts at 40+ organizations
• Discover and help fix vulnerabilities in critical systems such as OpenBSD, FFmpeg, Linux Kernel and more
• Discovered multiple high-risk zero-day vulnerabilities that had not been discovered for many years

These indicators not only demonstrate technical capabilities, but also reveal the reconstruction of infrastructure requirements: from a “training-based” computing model to a “reasoning-based” 24/7 continuous operation model.

A fundamental shift in the infrastructure computing paradigm

Infrastructure differences for training vs inference

NVIDIA’s Vera Rubin platform technology blog reveals the fundamental difference in infrastructure requirements for training and inference:

Training Workload

• Synchronized all-to-all communication phase
• Megawatt power peaks
• Massive GPU power swings
• No mitigation measures could stress the power grid, violate grid constraints or force operators to expand infrastructure

Inference Workload

• Sharp sudden demand peaks
• Continuous 24/7 operation
• Every user query, inference step, and API call is an inference load

“AI is now being embedded into products such as customer service and programming tools, with inference requirements running 24/7. This completely changes the infrastructure computing paradigm.” - NVIDIA GTC 2026

Hard threshold for power constraints

Computing Bottleneck

• AI data centers are facing physical limitations
• Single chip upgrade cannot completely solve the problem
• Requires a new fundamental design approach

Infrastructure Investment Model

• Training: periodic events, one-time large-scale investments
• Reasoning: continuous events, continuous infrastructure investment

IBM predicted in 2026: “2026 will be a decisive year between cutting-edge and efficient model categories.” This shift means that the infrastructure investment model will shift from cyclical training investment to continuous inference infrastructure construction.

Technology architecture evolution: GPU/CPU co-design

Design Focus of Vera Rubin Architecture

The NVIDIA Vera Rubin platform is designed for the era of agentic AI and inference:

Core Goal

• Eliminate critical bottlenecks in communication and memory movement -Super improve reasoning performance
• More tokens per watt, lower cost per token

Performance Index

• Compared with Blackwell architecture: performance per watt is improved and cost per token is reduced
• Network storage: The number of tokens per second increases by 5 times
• TCO (Total Cost of Ownership): 5x performance improvement
• Electrical efficiency: 5 times improved

Deployment Practice

• AWS, Google Cloud, Microsoft, OCI deploy Vera Rubin instances in 2026
• Microsoft deploys NVIDIA Vera Rubin NVL72 rack-scale system
• Cloud partners such as CoreWeave, Lambda, Nebius, Nscale and more

Hard constraints on power and cost

Constraint Scenario

• Electricity: Real-time inference must operate within electricity constraints
• Cost: Need to control reasoning costs
• Deployment: Requires actual deployable architecture

Solution Direction

• GPU/CPU co-design to handle agentic workloads
• Optimize communication and memory access patterns
• Adapt to 24/7 continuous reasoning needs

Changes in business models and investment logic

From “model performance competition” to “inference efficiency competition”

Front-edge model positioning

• Training: cyclical, high risk, high reward
• Reasoning: continuity, high availability, operational optimization

Investment logic changes

• Shift from “train once, long-term service” to “continuous reasoning, optimized operation”
• Infrastructure investment shifts from “training centers” to “inference centers”

Enterprise-level deployment strategy

Production Level AI Agent Deployment

• Goal: Run 100+ AI Agents by the end of 2026 -Each employee is equipped with Agent support
• A unified data and governance foundation for the end-to-end supply chain

Supply Chain AI Agent

• Autonomous systems operate across supply chains without human triggering
• Continuously optimize the supply chain
• Dynamically personalize customer experience

Quantified impact

• Leading companies achieve 4x impact, in half the time
• MIT and McKinsey study: Unifying data and governance foundations achieves 4x impact, in half the time

Strategic Impact of Geopolitics and Governance

Competition in cutting-edge model training and deployment

Regulatory Environment Differences

• EU: Rights-based framework (EU AI Act)
• China: state-centric model
• United States: Federal AI Governance Framework

Strategic Considerations

• Choose training location = choose regulatory environment = choose deployment mode
• Cutting edge models may be viewed as “critical infrastructure” rather than “general purpose tools”

Key decision points in 2026

**Decision 1: Training or inference? **

• Training: cyclical, high risk, high reward
• Reasoning: continuity, high availability, operational optimization

**Decision 2: How will cutting-edge models be regulated? **

• EU AI Act: hierarchical supervision
• United States: Federal Regulation
• Asia: National level regulation
• Key question: Are cutting-edge models critical infrastructure or general purpose tools?

**Decision 3: Who sets the rules? **

• UN Global Dialogue: Cooperation or Confrontation?
• Unilateral regulation or multilateral coordination?
• Technical standards vs laws and regulations

Hard threshold and technical boundary

Hard threshold for power constraints

Insurmountable Physical Limitations

• AI data centers are facing physical limitations of power
• Chip upgrade cannot solve the problem alone
• Requires a new fundamental design approach

Investment Constraints

• Training investment: cyclical, scalable
• Reasoning investment: sustainability, operating cost control

Hard boundaries of technical architecture

Protocols and Standards

• Communication protocol: All-to-all communication needs to be optimized
• Memory protocol: Need to reduce memory movement bottlenecks
• System protocol: needs to adapt to 24/7 reasoning requirements

Deployment Boundary

• Cloud computing: Need to support 24/7 inference
• Edge computing: requires low-latency inference
• Offline deployment: requires autonomous reasoning capabilities

Business model reshaping: reasoning efficiency is competitiveness

From “model performance” to “inference efficiency”

Changes in competition dimensions

• Shift from “training a better model” to “operating a better inference system”

Business Model Reconstruction

• Training costs: cyclical and predictable
• Reasoning cost: sustainable and optimizable
• Infrastructure: continuous investment, operation optimization

Enterprise-level AI Agent commercialization

AI Agent Type

• Customer Service Agent: 24/7 automatic response
• Coding Agent: continuous code optimization
• Supply chain agent: autonomous operation

Commercialization Path

• Training → Inference → Operation Optimization
• From “model performance competition” to “inference efficiency competition”

Conclusion: An irreversible shift in the infrastructure computing paradigm

2026 marks a turning point in AI infrastructure from the “training era” to the “inference era”:

1. Capability level: The capabilities of cutting-edge models have crossed the threshold and can reach or even surpass the level of human experts.
2. Infrastructure level: The continuity requirements of inference workloads change the basic design logic
3. Business model level: From periodic training investment to continuous inference operation
4. Geopolitical level: Training location, regulatory environment, and deployment model become strategic choices

This transformation is not only a technological upgrade, but also a fundamental change in the basic design philosophy. Enterprises need to shift from a “training-based” thinking model to a “reasoning-based” operational thinking model. Investors need to shift from the valuation logic of “training cycle” to the valuation logic of “reasoning operations”.

Hard conclusion: The computing paradigm shift in AI infrastructure is irreversible, which will redefine the pricing model, investment logic and competitive dimensions of cutting-edge AI.