AI 2030：计算、存储和网络领域更强大、更持久

AI 2030: Stronger for longer for compute,memory, networking Price Objective Change 13 May 2026 AI data center TAM now~$1.7Tn by CY30, +4 5% CAGR In-line with our view of strongerAI infra capex outlook(seeQ1cloud capextakeaways),weupdate our CY30 AI data center systems TAM outlook to ~$1.7Tn AI DC vs. ~$1.4Tnprior.Despite ongoing concerns around capex sustainability and the ability to fundthem, we believe: 1)2026 will continue to be a year of accelerating AI salesand ROIs(OpenAI, Anthropic upcoming IPOs likely to provide improving prospects), while 2) 2027could see improving tokenomics/efficiency as new architecture compute/memorysystems ramp. While the current supply chain bottlenecks could limit shipment ofleading-edge components, we believe key AI compute, networking,and memory vendorsthathave planned properly should continue to deliver andoutperform. EquityUnited StatesSemiconductors Vivek AryaResearch AnalystBofASvivek.arya@bofa.com Duksan JangResearch AnalystBofASduksan.jang@bofa.com Michael ManiResearch AnalystBofASmichael.mani@bofa.com Diversification of compute/memory additive to AI TAM Within our new~$1.7Tn TAM,we raise our AI accelerator outlook to~$1.2Tn from~$1.0Tn on increased hyperscaler custom ASIC shipments (such as Google TPU, AWSTrainium), data center server CPU outlook to ~$110bn (AI CPUs $88bn) from ~$80bn,and AI networking outlook to ~$316bn from ~$240bn on ongoing optics/switch ramp.Importantly, we flagthe ongoingdiversification of compute/memory componentsas AIworkload tails elongate, butweviewthisas additive to overallTAM, not replacingexistingworkloads/components. For instance, we expect CPUs (in their new standaloneracks) to work in tandem with existing GPU-CPU compute racks, and SRAM-based ultra-low-latency memory racks to coexist with HBM-based GPU racks. Liam PharrResearch AnalystBofASliam.pharr@bofa.com Exhibit1: 2030 AI TAMestimatechanges,toppicks,recentPOchangesToppicks,new/recentPOchanges Memorystill attractive, MU $950 PO on sum-of-parts Despite recent runsin memory stocks,we expectmemory demand to continueoutgrowing supplydriven by AI,memory pricing to generally hold up (supply/demandsufficiency ratio not expected to rise above 110%),andmemory (MU) earnings to remainrelatively stable through CY28. Given capital, packaging, and power limitations, we viewmemory supply elasticity as now structurally lower, and thusexpectmemory vendors(demand) to outperform equipment vendors (supply) over the medium term. Accordingly,we raise MU estimatesand PO to $950 from $500 on a much stronger mid-term pricingoutlook, now based on sum-of-parts that values: 1) sustainable AI/HBM business at~$240/sh (27x CY27 PE, in-line with AI compute peers), and 2) traditional cyclicalDRAM/NAND business at ~$710/sh (3.1x CY27 P/B, in-line with historical upcycle). Compute:NVDA, AVGO, AMD, MRVLMemory:MUNetworking:CRDO, MTSI Top picks: NVDA, AVGO, MU, AMD, MRVL 1) NVDA:top sector pick, raise est./PO to $320 from $300. Upcoming catalystsincludeearnings, Computex tradeshow (possible new CPU launch), Vera Rubin launch, and 2Hreturn of cash, perour recentreport.2) AVGO:recent frame contracts with Google andMeta provide FY27 certainty with potential upside to consensus $110bn AIsales, thoughwedon’t expect another raise on thenext call.3) MU:raise est./PO per above.4) AMD:CPU strength, July analyst day event, potential for additional GW wins without warrantrequirements.5) MRVL:raise est./PO to $200 from $125 oncontinued 800G/1.6Topticsramps, broad custom silicon demand strength vs. mgmt’s conservative MSFTprogram ramp outlook in CY27 (only ~$600mn contribution vs. $2bn supply). See glossary on page 22-23 BofA Securities does and seeks to do business with issuers covered in its researchreports. As a result, investors should be aware that the firm may have a conflict ofinterest that could affect the objectivity of this report. Investors should consider thisreport as only a single factor in making their investment decision.Refer to important disclosures on page29to32.Analyst Certification on page27.PriceObjective Basis/Risk on page24.12973153 Contents AI 2030 TAM Outlook Update Of the overall global IT spend of ~$6.3Tn today, we believedata center systemstorepresent $772bn across both AI and non-AI. By CY30, we expect TAM to grow toward~$2.0Tn, growing at +32% CAGR CY25-30 and outpacing overall IT spend at +9% CAGR. AI Data Center Systems:~$1.7Tn by CY30 ForAI data center systemsspecifically, wenowsee TAM growing to~$1.7Tn by CY30from $264bn in CY25, with AI servers representing~75% of TAM at $1.3Tn, followed bynetworking at~20% of TAM at $316bn, and storage at ~5% of TAM at $81bn. -Within AI servers, we seeAI accelerators to represent ~$1.2TnTAM by CY30 fromjust $120bn in CY24 and $197bn in CY25, or the vast majority of server spend. -We expectHBMto remain ~14-20% of overall accelerator spend,reaching $168bnTAM by CY30,though likely declines toward the lower-end of the range as memorytier