JPMorgan: AI Supply Chain Boom Spreading to Packaging, PCB, and MLCC — Bottlenecks Continue to Migrate
nashnova research
A J.P. Morgan Asia tech survey covering 37 companies and 115 investors reaches one conclusion: AI hardware demand is spilling from chips and HBM into packaging, substrates, liquid cooling and MLCC — but the winners will be suppliers with real technical moats, not the fastest capacity builders.
Where is the AI hardware money flowing now?
GPU and ASIC demand remains strong; cloud buyers show no meaningful pullback. Server supply chains report order visibility through 2027, with some firms already discussing early 2028 demand.
New buyers are entering: Neocloud and SpaceX are placing orders alongside traditional hyperscalers. This means → the demand base is broader than the market assumed — no longer a few cloud giants propping up the cycle.
Non-AI markets are mixed — industrial and auto are improving, PCs are more resilient than expected, and smartphones remain the weakest segment.
Why does the bottleneck keep migrating?
Advanced packaging is the clearest case: major Asian OSAT firms are raising 2026 capex to record levels, up over 80% year-on-year, yet capacity tightness may persist into 2027–2028.
In plain terms = the money is being spent, but output won't follow immediately. The bottleneck has shifted from cleanroom space to equipment — lead times for CoW (chip-on-wafer) advanced packaging tools stretch to 12 months, and traditional OSATs compete with TSMC for the same machines.
Substrates and PCBs face similar dynamics: high-end substrate contracts lock in volume, not price, and reprice by chip generation. EMIB-T — a packaging approach that skips the interposer — pushes more functions onto the substrate, demanding finer traces, more layers and higher yields.
Copper-clad laminate constraints are deepening too: low-end limited by glass fabric, high-end by copper foil. This reflects supply pressure cascading layer by layer up the chain.
When does liquid cooling actually ramp?
J.P. Morgan expects liquid cooling to re-accelerate in H2 2026, driven by Nvidia's VR200 volume production and ASIC programs such as Amazon Trainium 3 and Google TPU v8.
By 2027, ASICs could surpass GPUs as the larger source of liquid-cooling revenue. This means → suppliers' growth curves will shift from "follow Nvidia" to "follow multiple chip platforms."
Each program typically qualifies two primary suppliers plus one or two backups; limited certification slots help incumbents sustain gross margins.
Why has the MLCC suddenly become a critical component?
MLCCs — multi-layer ceramic capacitors, among the most common tiny components on a circuit board — are used in AI servers at volumes an order of magnitude higher than in cars or smartphones.
But volume is only the surface; the real uplift comes from spec upgrades: AI servers demand MLCCs that simultaneously deliver small size with high capacitance, high temperature tolerance (85 °C → 105 °C or 125 °C), high voltage (up to 1 200 V in some circuits), low ESL (parasitic inductance) and long-term reliability.
In plain terms = MLCCs used to be a "good enough" commodity part; now they are a "can't ship the server without it" bottleneck. The most advanced 1005-size products already reach 47 µF; the smallest form factor is 016008 — just 0.16 mm × 0.08 mm.
What power-delivery architecture will future servers use?
J.P. Morgan describes a layered setup: bulk MLCCs handle mainstream decoupling and energy storage; low-ESL multi-terminal or embedded MLCCs sit close to the die for high-frequency noise; silicon capacitors serve the most sensitive ultra-low-inductance nodes.
This means → the technology path is not winner-take-all but a stack sorted by capacitance, ESL, space and cost — each layer carrying different barriers and a different supplier landscape.
Low ESL can be achieved in multiple ways: die miniaturisation, reversed terminals, three-terminal or multi-terminal structures that shorten current paths, or embedding thin MLCCs directly under the IC inside the substrate to cut parasitic inductance.
Can Chinese suppliers close the gap?
In commodity MLCCs, mainland Chinese and Taiwanese makers have narrowed the distance significantly. But on advanced products — 0603, X6S decoupling, 1005 / 47 µF — gaps in materials, ultra-thin-sheet handling and manufacturing yield remain generational.
In the most advanced materials, Japanese firms are considered roughly a decade ahead; Murata's technology and patent lead is singled out. This reflects a core reality: the moat in high-end MLCCs is not equipment but the long-term co-optimisation of materials, equipment and process.
J.P. Morgan likens the current MLCC cycle to the "second inning" of a baseball game — server-related demand is still in its early stages. Suppliers are not chasing aggressive price hikes, prioritising customer relationships and long-term supply instead — a more restrained pricing strategy that helps avoid double-ordering, inventory build-up and over-expansion, thereby extending the cycle's duration.
市场有风险，内容仅供研究参考，不构成投资建议。