2025年第三季度发布报告：高级计算

INDUSTRY RESEARCH Launch Report:Advanced Computing PE and VC trends and investment strategies Contents Executive summary3Introduction6Market scenario8Market risks9Market growth areas10Recommendations to stakeholders12Overview and scope14 Institutional Research Group Analysis Rudy TorrijosDirector, Industry Researchrudy.torrijos@pitchbook.com DataOscar AllawayData Analystpbinstitutionalresearch@pitchbook.com Publishing Report designed byJenna O’MalleyandMegan Woodard Published on December 10, 2025 Executivesummary This report launches our coverage of the advanced computingvertical. Advanced computing comprises all end marketsthat enable artificial intelligence (AI) to scale. This includes We see hyperscaler capital expenditures as the single definingmetric for investors and vendors to gauge the growth andsustainability of the advanced computing ecosystem. Webelieve the current expected hyperscaler capital expenditure EXECUTIVE SUMMARY (training and inference). Without this achievement, end-user productivity from generative AI (GenAI) models will be •Elasticity of intelligence:Demand for processors, projectedto generate $2.8 trillion in cumulative revenue, is effectively in cumulative datacenter chip revenue through 2030,effectively decoupling the industry from traditional •Power constraints:Power availability has replaced siliconas the primary constraint on growth, driving $1.2 trillion in The primary beneficiaries of hyperscaler investments withinthe advanced computing sector are graphics processors,network processors, server microprocessors, AI accelerators, •Architectural efficiency:The industry-wide transition toMixture-of-experts (MoE) architectures has drastically •Thermal obsolescence:Legacy air-cooled datacenters(limited to around 30 kW per rack) are now technologically •Fidelity risk:Language model fidelity (LMF) is the singlemost critical performance metric, as a failure to achieve Key takeaways To assist investors and stakeholders in navigating thistransformational shift, we offer these key takeaways that will •The 60% threshold:Surpassing the 60% score on theARC-AGI-2 benchmark (average human intelligence) is the •Supply chain fragility:Geopolitical concentration remainsan existential risk, as the entire advanced computing stack •Capital expenditure magnitude:Hyperscaler capitalexpenditures are projected to reach a cumulative $6.4 trillion •Semiconductor supercycle:The semiconductor sectorhas entered a secular supercycle driving $3.7 trillion EXECUTIVE SUMMARY •Not dot-com:Unlike the speculative dot-com era, currenthyperscaler capital expenditures are grounded in broad- •VC pivot:Alpha generation has migrated from generalfoundation models to the “reasoning layer” and physical AI, •The rise of neoclouds:Specialized “neoclouds” haveemerged as a critical market segment, capitalizing on the •Enterprise mandate:Enterprises should abandon buildingfoundation models in favor of fine-tuning and retrieval- •The importance of 2 nm:The transition to 2-nm processnodes is a physical requirement to meet power-efficiency •PE strategy:Value creation for PE is shifting from passivedatacenter real estate to the active consolidation of critical requiring reinforced floors and specialized liquid-cooling loops.AI workloads rely on high-performance GPUs and accelerators using all-flash arrays. Data throughput speed is now as criticalas total storage capacity for training workloads. Datacenter infrastructure The datacenter infrastructure market includes all property,plants, and equipment needed to build out AI computing to We see the networking hardware market as the fastest growthmarket within datacenter infrastructure. Optical interconnects Power equipment OEMs are critical for on-site generation,power availability, and redundancy. Natural gas and dieselgenerators remain standard for primary and backup power.Cooling and thermal management equipment is no longer a The AI build-out represents a generational capital deploymentcycle of unprecedented scale. Given the magnitude of capital Within the operators & service providers segment, we trackthe AI cloud service providers (CSPs) themselves. CSPs can bedivided into generalist hyperscalers and specialized neoclouds Within the facilities & power infrastructure segment, we trackpower-generation innovation and new construction of facilities.Site selection is now focused on secure power access over Within the IT hardware & components segment, we trackdatacenter ICs: processors, accelerators, HBM, and networkingICs. Semiconductor companies do not sell directly tohyperscalers but rather work with ODMs. Datacenters require Hyperscalers leverage massive scale while neocloudscompete on availability of GPUs. We forecast that the Big Fivehyperscalers (Microsoft, Amazon, Google, Meta, and Oracle) For reference, traditional datacenters support 10 to 15 kW perrack. Advanced AI clusters require 50 to 100 kW per rack, Foundation model comp