Morgan Stanley Tech Private Meeting: AI Hardware Chain Expanding from GPUs to Memory, MLCCs and Tantalum Capacitors
Miles Bennett
Morgan Stanley's closed-door tech conference concluded that AI hardware demand has spread beyond GPUs into memory, CPO, MLCCs and tantalum capacitors, with the market narrative shifting from "demand story" to "pricing delivery" — five checkpoints will determine whether this leg of the rally converts into earnings.
Memory stocks pulled back after a big run — is the cycle peaking?
Morgan Stanley views the post-Computex pullback as a "healthy reset," not a fundamental top — some memory names have rallied 70%–130%+ from their March lows, and profit-taking plus ETF deleveraging drove short-term volatility.
Underlying demand has not slowed: AI inference and Agentic AI — a new paradigm where AI autonomously executes multi-step tasks — are still accelerating memory consumption.
This means → the pullback is capital digesting gains, not the industry thesis breaking down.
How large are the DRAM and HBM price increases, really?
Near-term, Morgan Stanley expects Q3 DRAM contract prices up 10%–20% QoQ and NAND up 20%–30%.
The bigger story is HBM — high-bandwidth memory, the ultra-fast memory built specifically for AI chips: it is entering an LTA (long-term agreement) repricing phase, with 2027 HBM prices potentially up 50%–100% YoY.
In plain terms = large customers lock in multi-year contracts → prices stop swinging quarter to quarter → memory makers' earnings visibility jumps sharply.
This reflects a valuation regime change: if LTAs cover 50%–70%+ of supply, DRAM stocks could re-rate from the ~5× P/E typical of past cycle peaks to an 8–10× range.
DDR4 is a legacy product — why is it rallying?
Morgan Stanley raised its Q3 DDR4 price-hike forecast to 20%–30%, with a bull case of 30%–50% — well above prior market expectations.
The reason is straightforward: memory makers are prioritizing capacity for HBM and other advanced products, squeezing DDR4 marginal supply; Nanya Technology's new capacity is not expected until H2 2027.
This means → DDR4 price strength could last into 2027 — not because demand is booming, but because no one wants to allocate capacity to it.
CPO has sold off — is the optical-interconnect thesis broken?
CPO — co-packaged optics, integrating optical modules directly into chip packaging — has weakened recently. The core issue is yields lagging expectations: TSMC's PIC wafer output is ~500 wafers/month today, targeting 10,000 by Q1 2027, but back-end packaging yield sits at just 50%–60%, and ODMs assembling full CPO switches report end-to-end yields as low as 20%–50%.
In plain terms = the design direction is sound, but the "build it at scale" step is still ramping — near-term shipments will undershoot.
Morgan Stanley's view: the long-term architecture logic is intact — post-2028, more and more designs will need optical interconnects to break through copper's physical limits. Yield is an engineering problem, not a directional one.
MLCCs and tantalum caps — how is AI pulling even passive components along?
On MLCCs — multilayer ceramic capacitors, the most basic energy-storage components on a circuit board: Yageo disclosed that standard-product utilization tops 75% and specialty-product utilization tops 85%, both above prior guidance. Distributor inventory fell from 5.5 months to 4.8 months, drawn down by AI server demand.
The key driver is not "more units" but "more expensive units": from the B200 to the GB200 platform, MLCC dollar content is expected to rise 182%, with 47 μF+ high-capacitance specs up 210% — the value uplift comes from a richer mix, not volume alone.
On tantalum capacitors, Yageo has raised prices three times since April 2025; upstream tantalum pentoxide prices have surged 100%+ since late 2025. Morgan Stanley expects another round of hikes once NVIDIA and ASIC server volumes ramp in H2 2026.
What to watch next? Five checkpoints on one list
Checkpoint 1: Can Q3 DRAM/NAND contract prices deliver the expected increases?
Checkpoint 2: Will HBM LTA coverage ratios and 2027 pricing materialize?
Checkpoint 3: Does CPO back-end yield improve steadily from the 20%–50% range?
Checkpoint 4: Do MLCC makers' utilization rates keep rising and channel inventory keep falling?
This means → the pace at which these five checkpoints are met will determine whether this AI infrastructure-materials rally moves from narrative to earnings delivery.
Content is for reference only, not financial advice.