Goldman Sachs Outlines Five Key Semiconductor Debates: WFE Cycle May Extend Through 2028
nashnova research
Ahead of the Communacopia tech conference in San Francisco, Goldman Sachs laid out five core debates spanning AI compute, equipment cycles, memory supply-demand, analog recovery, and EDA monetization — sketching a semiconductor upcycle that could last through at least 2028, with direct implications for equipment, memory, analog, and EDA investment timing.
Where does AI compute spending land — GPUs or custom chips?
Goldman expects hyperscaler capex to remain strong, with Agentic AI — AI agents that autonomously execute tasks — emerging as a fresh demand catalyst.
This means → inference costs keep falling + applications like code generation are showing quantifiable ROI = the economic case for AI infrastructure spending is being validated, no longer a "build first, justify later" bet.
On chip architecture, general-purpose GPUs still dominate near-term, but ASICs (application-specific integrated circuits — chips purpose-built for a single task) will steadily gain share. Put simply = competition shifts from raw performance to a three-way trade-off of performance, cost, and power.
Nvidia is expected to discuss its Rubin product cycle; Broadcom will likely focus on AI networking and custom XPUs; AMD may detail its MI4XX rack-scale solution and ROCm software stack.
Can the equipment upcycle really last until 2028?
Goldman sees the current WFE — wafer fabrication equipment, the full toolkit for making chips — upcycle lasting at least through 2028.
Key growth engines: DRAM + leading-edge logic/foundry + advanced packaging; NAND and mature-node logic capex are also gradually recovering.
The market's core debate: can 2027 WFE growth exceed the roughly 35% expected for 2026, and can the cycle extend beyond 2028? This means → if the answer is yes, equipment-stock valuations get structurally re-anchored higher.
Goldman flags non-traditional customers like Terafab entering the market + Intel's investment cycle re-accelerating as potential sources of incremental demand. Deposition, etch, advanced packaging, inspection, and metrology are all segments to watch.
Why could memory chips keep rallying?
Goldman's supply-demand model shows DRAM shortfalls of 5.0%, 5.9%, and 3.9% for 2026–2028; NAND gaps of 4.4%, 4.6%, and 3.0% over the same period.
In plain terms = even as new capacity comes online, the memory market could stay undersupplied for three straight years — the gap is not narrowing, it persists.
This means → most memory companies may channel roughly 50%–100% of excess free cash flow into capital returns; improving cash flow + rising shareholder payouts could reshape how the market values memory stocks.
On the technology side, progress on HAMR — heat-assisted magnetic recording, which uses a laser to heat the disk surface for higher storage density — and commercialization of new NAND architectures like HBF could alter the supply picture.
Has the analog chip recovery peaked already?
As of June, analog chip shipments are running about 5% above the long-term trend line. Goldman argues this does not signal the cycle is near its end.
This reflects four years of shipments running below trend, which built up a large demand-repair backlog — volumes have only just crossed back above the normal line.
Auto and other traditional end-markets are improving, while AI data centers add new incremental demand. Put simply = those two forces stacking together could make this analog recovery last longer than the market expects.
The key variable: whether price recovery can translate into revenue and margin expansion — if it does, there is still room for valuation upgrades.
Where does EDA's $3.7 billion upside come from?
Goldman estimates Agentic AI could generate roughly $3.7 billion in incremental annual revenue for the EDA — electronic design automation, the software tools used to design chips — industry by 2030.
This means → that upside is not yet priced into sell-side earnings estimates and may start materializing as early as the second half of 2026.
The logic chain: rising AI-chip customization → greater design complexity → worsening engineer shortage → EDA firms use Agentic AI to automate parts of design, verification, and optimization, converting efficiency gains into commercial value.
In plain terms = the more bespoke chips become, the more design work there is and the fewer engineers to do it — AI tools become a hard necessity. Whether EDA can turn this wave into real revenue growth is the next key proof point.
市场有风险，内容仅供研究参考，不构成投资建议。