Morgan Stanley: Rising Risk of "Underinvestment" in AI Supply Chain, Top Picks in Compute & Networking
nashnova research
Morgan Stanley's latest research note argues that the risk of underinvestment in the AI supply chain now exceeds the risk of overinvestment, ranking AI compute as the top allocation priority, followed by networking, with memory requiring selectivity.
Why does Morgan Stanley now say "too little" is more dangerous than "too much"?
GPT-6 Astra's launch reinforced the "AI bottleneck" narrative — compute supply can't keep pace with model iteration.
Yet investor positioning has declined and macro uncertainty has risen, pulling capital back.
This means → Supply-side tightness plus demand-side hesitation makes underexposure the bigger risk. Morgan Stanley says lean in, not pull back.
Why does AI compute rank first?
Morgan Stanley calls AI compute its highest-conviction theme: the entire compute stack shows both shipment growth and earnings-revision upside simultaneously.
TSMC's AI semiconductor business sustains a 70%–80% CAGR; AI revenue already exceeds 30% of TSMC's total in 2026. Full-year capex is raised to $60–64 billion, with revenue guidance at 40%+ YoY growth.
ABF substrates — the critical middle layer connecting chips to circuit boards — are seeing demand shift from PCs to AI servers. The price-hike cycle arrived earlier and stronger than expected; top pick is Unimicron.
MLCCs — multilayer ceramic capacitors, the tiny components next to chips that stabilize power delivery — are surging in per-server value as AI server power density rises. Top pick is Murata; Samsung Electro-Mechanics benefits from both ABF and MLCC tailwinds, supporting roughly 29% earnings CAGR through 2026–2028.
Why does networking rank second — and why not just bet on optical?
As accelerator clusters (scale-up) keep expanding, networking bandwidth demand grows with them.
But Morgan Stanley flags that the migration to optical interconnects is more gradual than expected. A pure bet on near-term co-packaged optics (CPO — embedding optical modules directly next to chips to shorten transmission distance) volume is premature.
This means → Morgan Stanley prefers a portfolio that straddles optical, electrical, and architecture-enabling technologies. Top picks include Corning, Lumentum, Coherent, Keysight, and others.
Can you still buy memory?
Morgan Stanley expects the memory cycle to enter its late stage by Q4 2026, with DRAM monthly price YoY gains peaking.
Q3 DRAM contract prices are tracking at roughly +20% QoQ; Q4 drops to +5%–10%, with limited visibility beyond that.
On the supply side, DRAM bit shipments are up 36% year-to-date, tracking at roughly 31% full-year growth — above the prior 25% forecast. Over the past nine months, Morgan Stanley has raised its 2026 DRAM wafer-fab equipment forecast by 47%.
In plain terms = Fundamentals haven't collapsed, but the fattest stretch is behind us. From here, it's about picking the right names — Morgan Stanley favors CXMT for its structural share-gain story.
What's worth watching outside AI?
Analog semiconductors are Morgan Stanley's most attractive early-cycle diversification play. Top picks: STMicroelectronics, NXP, Renesas.
The first half of 2026 confirmed the industry entered the early upswing. Morgan Stanley expects improvement to last at least through H2: pricing improving, restocking beginning, book-to-bill staying constructive.
This means → This is a cycle-turn opportunity with no AI dependency — useful for hedging overconcentration in AI names.
What worries Morgan Stanley most?
AI beneficiary stocks no longer clearly outpace non-AI tech on earnings-revision momentum.
Consensus expects AI beneficiaries to grow 2027 earnings by 56% with further margin expansion. Morgan Stanley sees this combination as unlikely; a more realistic growth rate may be 53%.
From 2027 onward, AI capex growth will most likely decelerate — not turn negative, but the second derivative of growth falls.
In plain terms = The near-term driver for share prices is no longer "how big is the story" but "can earnings deliver" and "can the stock absorb a downward revision."
市场有风险，内容仅供研究参考，不构成投资建议。