AI Capex Revised Up: Memory, Packaging, and Power Supply Seize Pricing Power
Taylor Wilson
UBS now sees 2026 cloud capex growing 77% year-on-year, but the money is funneling through three bottlenecks — memory, advanced packaging and power — and whoever controls the delivery gate takes structural pricing power.
Where is the money going — and what does the strong-vs-weak split mean?
UBS's latest forecast: 2026 cloud capex at +77% YoY, 2027 at +19% — revised up by 7 and 16 percentage points respectively from prior assumptions.
Server shipments are expected to grow 19% and 17% YoY in 2026–2027; but PC and smartphone shipments have been cut to roughly −11% and −10%.
This means → AI and legacy end-devices compete for the same upstream pool — memory, advanced process, PCB, MLCC. The AI chain claims the supply; the legacy chain absorbs the price hikes.
In plain terms = the pie is growing, but only the AI half is expanding; the phone-and-PC half is getting squeezed.
Why are memory and passive components taking pricing power?
Samsung Electro-Mechanics may deliver a 55% operating-profit CAGR across FY26–28. Murata's AI-server MLCC — multilayer ceramic capacitors, tiny components that stabilize current on circuit boards — demand CAGR has jumped from ~30% to nearly 80%, with revenue potentially doubling over two to three years.
The pricing window for HBM — high-bandwidth memory, high-speed stacked memory designed for AI chips — enterprise SSDs and high-end passives is stretching from a traditional cycle into a structural pricing advantage.
This means → these segments are no longer cyclical plays that give back gains in a downturn; tight supply is delivering long-duration pricing power.
Why is advanced-packaging demand diversifying?
CoWoS — TSMC's advanced packaging that tiles multiple chips tightly together — used to serve almost exclusively NVIDIA GPUs. Now Google TPU, AWS Trainium and MediaTek ASICs are all queuing for capacity simultaneously.
J.P. Morgan notes MediaTek ASIC revenue at roughly $2 billion in 2026, implying $7–12 billion in 2027. TPU v9 supports both CoWoS-L and EMIB-T packaging paths; MediaTek captures extra value on the EMIB-T route.
This means → MediaTek is not picking up a side contract — it is structurally migrating toward an AI custom-silicon platform. Advanced-packaging capacity competition has shifted from one dominant buyer to multiple contenders.
Why are power and connectors moving from supporting roles to delivery gates?
AI rack power density keeps climbing. Delta Electronics (HVDC and data-center power), Chroma ATE (CPO and high-end power testing), Accton Technology (800G-to-1.6T switch migration) and Lotes (connectors) have moved from auxiliary suppliers to critical choke points that determine whether an AI cluster goes live on time.
In plain terms = no matter how fast the chip is, if the power supply can't keep up, connectors are short, or test gear is booked out, the whole rack stays dark — the power chain now holds bargaining chips equal to the silicon chain.
How do the three investment lines break down?
Line 1 — buy the project gate: Alchip, GUC, MediaTek, Accton, Lotes, Delta Electronics, Chroma ATE and Hon Hai. These companies sit on must-pass nodes from AI design to delivery.
Line 2 — buy bottleneck pricing: Samsung Electronics, SK hynix, Samsung Electro-Mechanics, Murata, plus the HBM, DRAM/NAND, ABF substrate, MLCC and high-end substrate chain. Tight supply gives them active pricing power.
Line 3 — stay cautious on the legacy end-device chain: display panels, low-bargaining-power phone/PC components, and brands or module makers that cannot pass memory cost increases through to consumers. This reflects the fact that within the same AI cycle, upstream price hikes are ultimately absorbed by the weakest links downstream.
Bull, base, bear — where are the dividing lines?
Bull: capex keeps rising and projects deliver on time. The AI chain broadens from GPUs to ASICs, networking, power and test equipment; Asia-Pacific tech enters a structural bull market.
Base: demand stays strong, but HBM, CoWoS, ABF and data-center deployment constraints push revenue recognition later. Companies with high order visibility hold valuations; pure-concept names get compressed first.
Bear: falling inference prices lead cloud operators to reassess capital returns. Project delays plus memory price hikes weigh on end demand; the market is willing to pay only for bottleneck assets with supply discipline and strong cash flow.
This means → the single key metric over the next four quarters is whether rising AI capex can push through supply bottlenecks and ultimately show up in revenue, gross margins and cash flow.
Content is for reference only, not financial advice.