Goldman Sachs Significantly Raises Server Market TAM, Upwardly Revising 2027/28 Estimates by 31%/45%
nashnova research
Goldman Sachs raised its 2027/2028 global server market forecasts by 31% and 45% respectively, with AI server TAM reaching $1 trillion by 2028. This means → Wall Street is repricing the AI infrastructure cycle from tentative to full upgrade.
What is driving this upgrade?
Three forces are pushing forecasts higher simultaneously: hyperscaler capex expansion, sustained AI infrastructure investment, and rising memory-chip costs.
In plain terms = the buyers are spending faster, and the servers themselves are getting more expensive — volume and price are both moving up.
Goldman raised both general-purpose and AI server estimates — this is not a single-category story but an industry-wide upcycle.
General-purpose servers — why is the "supporting act" also rising?
Goldman raised 2027/2028 general-purpose server value estimates by 17%/11%, with shipment volumes and average selling prices both revised upward.
Shipments are forecast to grow 20%, 12%, and 8% year-on-year from 2026 to 2028, reaching roughly $271 billion by 2028 — a 2025–2028 CAGR of 22%.
This means → general-purpose servers are no longer just "legacy business holding steady." Memory-chip price increases plus AI-workload spec upgrades have turned them into a standalone growth driver.
Even so, general-purpose servers will account for only 17% of the total market by 2028 — AI servers remain the headline.
How does the AI server market reach one trillion dollars?
For rack-scale AI servers — systems designed as whole-cabinet supercomputers — Goldman raised NVL72 shipment forecasts by 5%/1%, projecting 110,000 and 164,000 units in 2027/2028.
Nvidia racks account for 97,000/149,000 units; AMD racks for 13,000/15,000 — Nvidia remains dominant.
The global AI server rack market is projected to grow at a 163% CAGR to $1 trillion by 2028, accounting for 63% of the total server market.
In plain terms = three years from now, for every $100 spent on servers, $63 goes to AI servers.
Why are ASICs the hidden variable?
8-GPU AI servers — the standard configuration running eight GPUs in parallel — saw 2027/2028 shipment estimates raised 10%/29% to 2.7 million/3.3 million units, driven by rising demand for ASICs (application-specific integrated circuits — chips custom-designed for particular AI tasks, as opposed to general-purpose GPUs).
Goldman projects global ASIC penetration of roughly 60% by 2028, with tier-one hyperscalers at about 50%.
This means → cloud giants are no longer just buying Nvidia's off-the-shelf GPUs. They are increasingly designing their own chips — whoever achieves volume-production yields first gains a cost advantage.
The 2025–2028 CAGR for this segment is forecast at 38% — slower than rack-scale AI servers, but on a larger base with higher certainty.
Where exactly is the hyperscaler money going?
Goldman forecasts US hyperscaler capex growing 84%/52%/13% year-on-year from 2026 to 2028, reaching roughly $1.397 trillion by 2028.
China's leading cloud platforms are projected to grow 105%/21%/14% over the same period — 2026 growth even exceeds that of US peers.
Spending is concentrated on GPU/ASIC/CPU chipsets, high-bandwidth memory (HBM), and networking, cooling, and power components.
This reflects a multi-year spending curve, not a one-off pulse — and both the US and China are accelerating in parallel.
Who benefits across the supply chain — what does Goldman's "buy" list say?
Goldman issued a full-chain "buy" list: ODMs include Hon Hai/Foxconn Industrial Internet and Wiwynn/Wistron; optical modules include Zhongji Innolight and Eoptolink; PCB/copper-clad laminates cover Shenghong Tech, Shengyi Technology, TUC, and Taiyo Yuden.
Liquid-cooling names include Auras Technology, Forcecon, and Apaltek; silicon photonics names include LandMark Optoelectronics and Ase Industrial; the list also features TSMC, King Slide, Aspeed Technology, and others.
This means → Goldman sees AI server growth benefiting not just chipmakers but every link from assembly to optical interconnects, thermal management, and testing.
The key verification checkpoints: whether hyperscaler capex actually lands on schedule, and whether custom ASIC volume-production yields can support the penetration curve Goldman projects.
市场有风险，内容仅供研究参考，不构成投资建议。