Citi Roadshow: AI Hardware Demand Expanding from GPUs to Networking, CPO, and Packaging
nashnova research
Citi wrapped a US investor roadshow with one core message: AI hardware growth is spreading from GPUs into networking chips, CPO, advanced packaging, and power infrastructure. Taiwan's supply chain still has strong visibility — but the cycle ultimately hinges on whether hyperscalers can earn back what they're spending.
Where is the money flowing beyond GPUs?
Citi's headline call: AI hardware growth is broadening from GPUs and accelerators into networking chips, custom ASICs, co-packaged optics (CPO), advanced packaging, and power infrastructure.
This means → investors can no longer track AI hardware through a single GPU lens. The supporting layers around GPUs are becoming growth engines in their own right.
TSMC's commonly cited AI revenue metric covers GPUs and ASICs but understates networking chips and the CPU uplift that agentic AI may drive. In plain terms = TSMC's AI opportunity is larger than the headline numbers suggest.
How wide is TSMC's AI opportunity really?
Citi sees TSMC at the centre of virtually every major AI hardware shift — networking chips, AI CPUs, silicon photonics (chip technology that moves signals with light instead of electricity), and advanced packaging all contribute beyond the GPU/ASIC headline.
Growth in 2027 is expected to be even stronger: full-year 2nm revenue, rising 3nm AI chip shipments, and potential price increases could drive volume, mix, and ASP simultaneously.
On Intel, Citi notes its advanced-node capacity must prioritise its own CPUs, limiting room for external AI foundry clients. Intel's 14A ecosystem still faces a long qualification and design-in cycle before it can deliver meaningful external volume.
Advanced packaging: who captures TSMC's outsourcing wave?
Citi expects TSMC to focus resources on wafer fabrication and differentiated packaging like SoIC, outsourcing more back-end steps — benefiting ASE Technology (日月光投控) and SPIL.
ASE is not solely reliant on TSMC outsourcing. It is building its own full-flow advanced packaging capability, including FoCoS, to capture high-value AI packaging revenue more directly. Management expects leading-edge advanced packaging revenue to roughly double.
This means → as chiplets, HBM (high-bandwidth memory), silicon photonics, and complex interconnects proliferate, packaging value and technical barriers rise in tandem — potentially driving structural margin improvement for ASE.
When does CPO actually arrive?
CPO — co-packaged optics, integrating optical communication modules directly into chip packages — addresses both bandwidth and power consumption bottlenecks. Citi calls it one of the clearest structural shifts in AI hardware.
Citi separates two timelines that the market often conflates: switch-level CPO is closer to production and should ramp first; integrating CPO directly alongside GPU packages still needs two to three years.
The near-term bottleneck is external laser sources and fibre array unit (FAU) capacity, explaining why Lumentum and Corning are expanding aggressively. Citi sees Largan's precision automation capabilities positioning it well for the highest-difficulty fibre-array alignment step.
MediaTek and GUC: how big is the Google TPU uplift?
Citi expects MediaTek's role in future Google TPU generations to expand significantly, targeting an initial 15%–20% share with room to grow.
For the current TPU v8 generation, MediaTek handles I/O and SerDes — serialiser/deserialiser, the high-speed data links between chips. The next generation adds partial compute functions. This means → MediaTek's revenue per die rises materially, so its revenue growth could far outpace TPU shipment or share gains alone.
Citi flags a risk: a future Google–MediaTek project is expected to use Intel's EMIB-T technology, which remains under-validated at hyperscale volumes. Early yield and throughput may be constrained.
Hon Hai and power architecture: what drives the next leg up?
Hon Hai's growth this year comes from higher AI rack shipments. Next year, as the supply chain shifts from GB300 to Vera Rubin, per-rack value rises — meaning revenue and profit can grow meaningfully even if shipment growth slows.
On power architecture, Citi sees 800V DC power (800VDC) as strategically attractive but not a hard requirement for Rubin. Greenfield data centres have more incentive to design around 800VDC; existing facilities can stick with alternatives. In plain terms = 800VDC will spread gradually, not flip on overnight.
Citi stresses: regardless of whether 800VDC is adopted, power management chip and capacitor shortages remain a binding constraint. Delta Electronics holds structural advantages in high-power data-centre infrastructure, but the actual pace of 800VDC adoption may lag market expectations.
AI infrastructure spending cycle: can it last beyond 2028?
BULL
High supply-chain visibility
Taiwan's supply chain from TSMC to ASE has clear order visibility for the next two years.
Growth engines broadening
From GPUs to networking, CPO, packaging, and power — the investment thesis is widening, not narrowing.
Rising unit values
2nm ramp, rack upgrades, packaging complexity — every layer is seeing price increases.
BEAR
Commercial returns unproven
Whether hyperscalers can earn back their AI spending is a validation point that hasn't arrived yet.
New-tech production risk
Intel EMIB-T, 800VDC, and other technology paths lack validation at mass-production scale.
The biggest long-term risk
Citi explicitly flags commercial-return validation as the single largest uncertainty spanning the entire cycle.
In plain terms = near-term orders are real, but whether the full cycle plays out depends on hyperscalers earning back what they've spent — and that card hasn't been turned over yet.
市场有风险，内容仅供研究参考，不构成投资建议。