AI infrastructure technology supply chain: JPMorgan sees the AI-infrastructure trade at a crossroads but expects capex and supply-chain demand to remain resilient.
The report argues that AI token economics, accelerating model development and available hyperscaler funding should sustain AI-infrastructure spending over the next two years. It favors selected Asian equipment, substrate, packaging and optical beneficiaries while flagging funding, regulation and memory-demand uncertainty.
Summary
The report argues that AI token economics, accelerating model development and available hyperscaler funding should sustain AI-infrastructure spending over the next two years. It favors selected Asian equipment, substrate, packaging and optical beneficiaries while flagging funding, regulation and memory-demand uncertainty.
- The report estimates a hypothetical frontier-model vendor could generate US$20-40bn of annual token revenue per GW, versus US$10bn per GW in 2025.
- Hyperscaler net debt to equity is cited at 13%, which JPMorgan views as leaving room for further capex funding.
- Semiconductor equipment, IC substrates and advanced packaging are identified as key beneficiaries of tightening capacity.
- NPO is preferred near term for AI interconnects, while CPO is viewed as the longer-term architecture.
Report Interpretation
Overview
JPMorgan examines whether the AI-infrastructure rally can continue as investors question compute monetization, AI safety regulation and hyperscaler funding. Its conclusion is constructive: current AI economics and supply-chain conditions still support strong capex growth, with particular opportunities in semiconductor equipment, advanced packaging, substrates and near-term optical interconnects.
Core views
JPMorgan frames the next phase of the AI-infrastructure trade around three questions: whether buying compute remains economically attractive for AI-model and vertical-AI vendors; whether safety and regulatory concerns could slow compute demand; and whether hyperscalers can finance another leg of capex as many move into negative free-cash-flow territory. The report’s answer is broadly positive. It says selling compute remains highly profitable across hardware and cloud ecosystems, while current token economics indicate that buying compute can also be sustainable. It cites reported inference gross margins of 60-80% for some frontier-model vendors and estimates that a hypothetical model vendor could produce US$20-40bn of annual revenue per GW from AI tokens, compared with US$10bn per GW in 2025. It also points to more than 80 vertical-AI companies reportedly reaching over US$100m of ARR rapidly as evidence of downstream value creation. The report argues that open-source models should expand, rather than reduce, hardware demand. It expects open source to push token costs back toward their historical 80-90% cost-decline trend after the compute-shortage and agentic-AI-driven anomaly in 1H26. In a compute-constrained market, JPMorgan expects both proprietary frontier models and open-source models to experience strong token-consumption growth. Cheaper tokens should broaden AI adoption and enable more domain-specific vertical applications, while open-source progress should also keep competitive pressure on frontier model labs to continue training and improving their models. On regulation, JPMorgan treats recent AI-safety concerns as a near-term disruption rather than evidence of a structural training slowdown. It argues that rapid model advancement, including the prospect of recursive self-improvement, implies that AI scaling and technology evolution are continuing. New-model releases could be constrained by alignment restrictions, but the report does not expect the pace of frontier-model training to slow. It believes continued advances could open new addressable markets over the next one to two years, analogous to the expansion in coding and agentic workflows during the prior 18 months. Funding is the report’s principal area of investor concern. JPMorgan notes that many hyperscalers are entering negative free-cash-flow territory and increasingly require external funding for additional capex. Nevertheless, it sees capacity for further spending because large hyperscalers’ net debt-to-equity ratio is still only 13%, public-cloud growth is accelerating and can lift operating cash flow, and both hyperscalers and downstream compute buyers can issue equity. Public-cloud backlog data also show growth through 2Q26: Google Cloud backlog reached US$514bn, AWS backlog US$496bn, and Microsoft commercial remaining performance obligations US$678bn. The report cautions that rising rates and greater risk perception in some AI-infrastructure funding segments make financing an ongoing issue to monitor. Within Asian technology supply chains, JPMorgan sees semiconductor production equipment as especially attractive heading into 2027. It expects equipment to become a key bottleneck as semiconductor vendors expand capacity. TSMC is expected to step up investment in N2, N3 and A14, while Intel, Samsung Foundry, Terafab and trailing-edge foundries also increase capex. DRAM and NAND capex is expected to rise rapidly in 2027-28 as clean-room capacity comes online, process migration accelerates and suppliers serve customer LTAs; China memory investment is also expected to accelerate over the next two years. Tight tool capacity, higher input costs and expedited orders are already supporting equipment price increases. JPMorgan’s equipment picks are Advantest, Tokyo Electron, Chroma, Grand Process Technology and ASMPT. The report identifies IC substrates and advanced packaging as enduring AI bottlenecks. Tier-1 substrate supply is concentrated, including at Unimicron and Ibiden, and new capacity can take up to about 2.5 years to build. Leading OSATs are also constrained by clean-room availability. Higher substrate content in AI-accelerator packages and greater advanced-packaging intensity are expected to widen the supply-demand gap, with further upside from agentic-AI-related server CPU demand and adoption of substrate-intensive EMIB-T packaging from late 2027. JPMorgan therefore prefers Unimicron and Ibiden among substrate suppliers and ASE among advanced-packaging leaders. For AI interconnects, the report expects data transfer among GPUs, CPUs, memory and storage to become a key cluster-efficiency constraint. It favors near-packaged optics in the next two years because it is a simpler near-term solution and expects Google, AWS and NVIDIA to adopt NPO-based solutions. Co-packaged optics is described as the longer-term destination, but one requiring more supply-chain maturation, particularly for scale-up applications. Memory remains fundamentally constructive but more contested from an investor perspective. JPMorgan expects supply-demand balance to remain tight until at least 2028 and pricing to continue rising into 2027. However, it highlights concerns around HBM specification reductions, algorithmic-efficiency efforts that reduce memory use per function, and shifting KV-cache storage toward DDR or NAND. It expects investor sponsorship to remain relatively lukewarm until a new model-development cycle materially raises memory demand per function. Finally, JPMorgan expects stronger percentage price increases in 2027 than in 2026 for foundries, especially 8-inch mature foundries, OSAT, substrates, semiconductor equipment and high-end CCL. It expects TSMC to raise prices across all process nodes in 2027, versus leading-edge nodes only in 2026, and estimates 8-inch mature-foundry price increases of 10-15%, compared with 8-10% in 2026. Memory and PCB materials are the exceptions where the pace of price gains is expected to slow.
Analysis framework
JPMorgan starts with AI-compute unit economics and the effect of open-source models, then assesses regulatory and funding constraints on hyperscaler capex. It translates that demand outlook through the semiconductor value chain, identifying capacity bottlenecks, expected investment cycles, pricing conditions and preferred Asian supply-chain segments. It also uses public-cloud backlog, capex, operating-cash-flow, leverage, earnings-revision and valuation comparisons to support the sector view.
Methodology notes
Supply-demand analysis of AI compute, semiconductor capacity, substrates, packaging, memory and equipment.
The report links AI-compute demand and constrained manufacturing capacity to expected capex, bottlenecks and pricing outcomes across technology supply chains.
AI-compute economics transmitted from model vendors and cloud providers to semiconductor equipment, packaging, substrates and optics.
JPMorgan uses the chain from token demand and hyperscaler spending to identify which upstream Asian technology suppliers may benefit.
Forward P/E comparison for Asia technology excluding memory.
The report references forward P/E as part of its assessment that technology-hardware valuations remain compelling.
Forward P/B comparison for Asia technology.
The report includes forward P/B as an additional valuation reference for the broader Asia technology sector.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Advantest (6857.T)Top pick and semiconductor-equipment beneficiary of rising foundry and memory capex.
- Strengths
- Expected equipment bottlenecks and tighter tool capacity.
- Comparison
- Included among preferred semiconductor production equipment vendors.
- Risks
- Capex funding or demand slowdown could weaken equipment orders.
- Tokyo Electron (8035.T)Top pick and semiconductor-equipment beneficiary of expanding foundry and memory investment.
- Strengths
- Expected broad equipment demand as capacity expansion accelerates.
- Comparison
- Included among preferred semiconductor production equipment vendors.
- Risks
- Capex funding or demand slowdown could weaken equipment orders.
- ASMPT (0522.HK)Top pick linked to semiconductor equipment and packaging-capacity tightness.
- Strengths
- Benefits from equipment demand and constrained clean-room capacity.
- Comparison
- Named among JPMorgan’s preferred equipment vendors.
- Risks
- Capex funding or demand slowdown could weaken orders.
- Unimicron (3037.TW)Preferred tier-1 IC-substrate supplier.
- Strengths
- Concentrated substrate capacity, long build lead times and rising AI-package substrate content.
- Comparison
- Preferred alongside Ibiden among tier-1 substrate vendors.
- Risks
- Demand could be affected by AI hardware efficiency or slower accelerator growth.
- ASE Technology Holding (3711.TW)Preferred advanced-packaging leader.
- Strengths
- Advanced packaging remains capacity constrained and AI-package intensity is rising.
- Weaknesses
- Clean-room availability constrains industry capacity.
- Comparison
- Identified as an advanced-packaging leader.
- Risks
- AI accelerator demand or packaging investment could slow.
Key data
- Hypothetical AI-token revenue per GWUS$20-40bn annuallyJPMorgan’s back-of-the-envelope estimate for a model vendor, versus US$10bn per GW per year in 2025.
- Reported frontier-model inference gross margin60-80%Reported range cited by the report.
- Vertical AI companies exceeding US$100m ARR80+Cited from a Sapphire Venture report as evidence of downstream AI value creation.
- Historical token-cost decline80-90%Historical decline in cost per token cited by the report.
- Large hyperscaler net debt to equity13%The report considers this low enough to leave funding capacity for capex growth.
- Public-cloud backlog in 2Q26Google US$514bn; AWS US$496bn; Microsoft US$678bnCompany-reported backlog or remaining-performance-obligation measures.
- Substrate capacity build lead timeUp to ~2.5 yearsA factor supporting persistent substrate tightness.
- Expected 8-inch mature-foundry price increase in 202710-15%Compared with 8-10% in 2026, according to the report.
Impact & implications
JPMorgan believes the absence of imminent capex-downturn signals, steady Asia-tech EPS revisions and supply constraints support the AI-infrastructure ecosystem despite poor sentiment. It sees the strongest supply-chain setup in semiconductor equipment, substrates, advanced packaging and near-term NPO optics; memory fundamentals remain positive but require greater confidence in future demand intensity.
Risks
- AI safety or alignment restrictions could delay the rollout of new models, even if training continues.
- Hyperscaler funding is a key risk area as free cash flow turns negative, rates rise and AI-infrastructure funding risk perception increases.
- Open-source competition and uncertain model-layer monetization remain market concerns.
- Memory demand could be reduced by HBM specification changes, algorithmic efficiency gains or shifting KV-cache storage to DDR or NAND.
What to watch
- Evidence of sustained monetization at the AI model and application layers, identified as a key catalyst.
- Hyperscaler operating-cash-flow growth, debt and equity issuance as funding sources for additional capex.
- The pace of AI model advancement and any safety or alignment restrictions on deployment.
- Equipment bottlenecks, clean-room capacity, substrate supply and the adoption timetable for NPO and EMIB-T.
- Whether future AI-model breakthroughs materially increase memory demand per function.