UBS: AI Infrastructure Investment Spreading from GPUs to the Entire Supply Chain
Nashnova编辑部
After surveying 15 core vendors at OCP APAC Summit 2026 in Taipei, UBS concludes that AI investment is shifting from accelerators themselves to everything needed to keep them running — power, packaging, networking, cooling, and rack architecture are all becoming independent growth vectors.
Why is the "just buy GPUs" era over?
UBS's central finding: AI infrastructure is moving to a whole-system approach — compute, memory, networking, power, cooling, security, packaging, storage, and rack architecture must all be optimized together.
This means → the old logic of "buying GPUs equals doing AI" no longer holds. Every layer around the GPU is becoming a separate capex line item.
In plain terms = the GPU is the engine, but the engine is now so powerful that the fuel lines, cooling, and transmission all need a redesign — and money is flowing outward along that chain.
Power upgrades: why is 400V/800V DC the most certain path?
Microsoft, Google, Nvidia, Delta, and others all point the same way: starting in 2027, power electronics will migrate to standalone power sidecars — dedicated power modules pulled out of the rack — delivering 400V/800V DC directly to IT racks.
UBS industry feedback: penetration starts at roughly 10% with Nvidia's Rubin platform and rises further at Rubin Ultra. Google and Meta's ±400V ASIC systems may move faster, with penetration projected at 30%–40%.
This means → high-voltage DC raises rack density while cutting copper use, cable bulk, and conversion losses — one upgrade addresses multiple bottlenecks, which is why the industry consensus is so strong.
Advanced packaging: why has the bottleneck shifted from "shrinking transistors" to "assembling chips together"?
TSMC, ASE, Applied Materials, and Broadcom all stress that AI performance increasingly depends on advanced packaging. The toolkit now spans CoWoS, CoPoS, SoIC, fan-out/FoCoS, CPO — co-packaged optics — and STCO (system-technology co-optimization).
The critical bottlenecks are no longer just logic process nodes; they include yield, warpage, substrate supply, thermal design, and optical alignment.
In plain terms = the race used to be about etching circuits smaller. Now it is about stacking multiple chips together like precision Lego — the assembly step is rising fast in both difficulty and value.
Networking and optics: why is data movement the next chokepoint?
As speeds climb, copper reach shortens, pluggable optics lose reliability, and chip-edge bandwidth falls short. Co-packaged optics (CPO) — placing optical components directly next to the chip so data travels as light, not electricity — is emerging as the solution.
ASE frames CPO as a natural extension of advanced packaging: integrating optics, ASICs, switch chips, HBM, and power delivery into a tighter system-level architecture.
This means → after "compute faster," "move data faster" is the new bottleneck — and the strategic value of optical-interconnect players is being repriced.
In the Agentic AI era, why do CPUs matter more, not less?
AMD and Arm both note that Agentic AI — AI agents that autonomously call tools and manage multi-step tasks — does not just need GPUs. It also sharply increases CPU orchestration demand for memory, storage, RAG/vector databases, tool calls, and system management.
AMD's data: the traditional inference CPU-to-GPU ratio is about 1:4; new architectures are converging toward 1:1. This means → CPU compute inside data centers will roughly double, yet GPU rack demand does not shrink.
In plain terms = GPUs are the workers doing the heavy lifting; CPUs are the dispatchers. More workers need proportionally more dispatchers, or the whole site stalls.
Who does UBS favor? Where will the money flow?
Semiconductors: UBS prefers TSMC, ASE, ASpeed, and MediaTek — benefiting respectively from advanced packaging, CPO/optical packaging, BMC/security chips, and custom silicon with system-level integration.
Hardware and components: server ODMs Foxconn, Quanta, Wistron, Wiwynn; infrastructure players Delta, King Slide, and BizLink — benefiting from higher rack density, 800V DC power, liquid cooling, and cable/interconnect upgrades.
This reflects UBS's core thesis: the AI investment diffusion has already begun, but whether valuations across each segment fully price in that diffusion will be the market's key debate going forward.
Content is for reference only, not financial advice.