为什么人工智能巨头需要重新制定供应链战略

Why AI HyperscalersNeed to Recode TheirSupply ChainStrategy The AI industry has spent the past few years engagedin a high-stakes infrastructure arms race—onedefinedby billion-dollar bets, rapid technologicalleaps, and an unrelenting demand for computationalpower. The companies that can build the fastest, mostpowerful AI systems have long believed they willemerge as theindustry’s dominant players. And untilrecently, thatbelief seemed indisputable. But as 2025 unfolds, the assumptions underpinningthis strategy are beginning to crack. Two newimperatives are reshaping the industry’s approach to AIinfrastructure: •Agility—The ability to pivot instantly in responsetonew breakthroughs, regulatory shifts, orunexpected constraints. •Cost Efficiency—Ensuringthat AI inference(the execution of AI models at scale) can remaineconomically viable, rather than a black hole ofoperational costs. Venture capital poured in. Tech giants doubled down.Governments got involved. The $500 billion StargateProject in the U.S, an unprecedented investment in AIinfrastructure, made it clear that the race to scaleAIisn’t just about private competition—it was aboutnational strategy. The fundamental equation seemedsimple: The problem? The infrastructure built to maximizeperformance and speed was never designed to beflexible or cost-efficient. Vertical integration—wherea company owns and controls everything from chipsto data centers—has long been seen as the optimalapproach. But now, as the financial realities of AIinference come into focus, the industry is being forcedto rethink whether a more externalized, modular supplychain might offer a better path forward. •Maximize Raw Performance—The sheercomputational muscle needed to train ever-larger AImodels. •Maximize Deployment Speed—The ability to scalenew models at an industrial pace. Achieving that balance isn’t a simple choice betweenowning infrastructure or outsourcing it. The trade-offswill vary across different components of the supplychain, from compute hardware to energy sourcing. Arigid, one-size-fits-all approach is no longer feasible—what’s needed instead is a portfolio strategy, one thatoptimizes for speed and performance while maintainingenough flexibility and cost control to sustain AI’s long-term growth. To sustain this, an immense, highly specialized, andrigid physical supply chain emerged—what insidersnow call the “Supply Chain of AI”.This infrastructureisn’t just about silicon chips or cloud storage; it’s asprawling ecosystem of data centers, power generation,fiber optic networks, and high-performance coolingsystems, all working in tandem to sustain AI’s insatiableappetite for compute power. WHY AI HYPERSCALERS NEED TO RECODETHEIR SUPPLY CHAIN STRATEGY The Supply Chain of AI is often framed around four coreelements—Talent, Models, Data, and Chips—thebattlegrounds where companies compete fordominance.But beneath these high-profile componentslies anotherset of six hidden elements that are just as critical and, inmany cases, the biggest bottlenecks slowing AIdeployment. These infrastructure layers—data center construction,power generation, compute hardware, infrastructureequipment, real estate, and telecom networks—arefacing mounting challenges. U.S. data centers areprojected to consume up to 9.1% of the nation’selectricity by 2030,1Microsoft is reportedlyreassessing its datacenter plans,2and gridconnection times for new data centers in NorthernVirginia now exceed seven years.3 For a more detailed look into these elements and theroadblocks shaping AI’s future, see our full analysishere. Six HiddenElements Evolving Goalpostsof AI Deployment For the past five years, the race to deploy artificialintelligence has been driven by two fundamentalmetrics: •Raw Performance—Who has the most advancedmodels or chips? •Deployment Speed—How fast can thesetechnologies be integrated into massive datacenters? This relentless focus on power and speed fueled awave of vertical integration, with companies developingproprietary chips, stockpiling electrical infrastructureequipment, and locking in long-term control over theirAI supply chains. The assumption was clear: owningeverything from silicon to server farms would ensuredominance. •The Cost Explosion:AI compute costs are spiralingout of control, and more waves of change are coming.In particular, agentic consumption—where AI agentsautonomously generate more AI workloads, creatingan exponential growth in resource demand, is veryunpredictable.Microsoft,OpenAI, and Google arealready strugglingwith forecasting these infrastructurecosts, aseach successive AI model iterationdemandsexponentially more compute power.6 But as 2025 unfolds, the equation is changing. Theexplosive growth of AI applications, rising operationalcosts, and mounting revenue pressures are forcinghyperscalers to rethink their priorities. •The Inference Bottleneck:ChatGPT may soonsurge past 1 billion active users, driving demand forAIinference—processing real-tim