260624_全球伺服器_gs_global-server

PDF 原檔：260624_全球伺服器_gs_global-server_original.pdf

原始內容

Global Server: Raising both AI and General servers; better mix and higher memory costs drive CSP spending

We raise our global servers TAM to re fl ect the ramp-up of AI server racks, and higher memory costs driving CSP spending. We refresh our forecasts for (1) AI server racks (NVL72-equivalent), (2) AI servers (8-GPU equivalent), (3) General servers, (4) AI servers implied GPU vs. ASIC, and (5) leading US and China cloud capex. We expect AI server rack (NVL72-equivalent) shipment (incl. Nvidia and AMD) to reach 55k/ 105k/ 163k in 2026-28E, increasing by 16% / 20% in 2027 / 28E, re fl ecting our positive view on AI server racks, which carry more intense design and stronger GPU interconnection. On ASIC AI servers, we expect growing ASIC adoption, with ASICs accounting for 50%/ 52%/ 55% of AI chips in 2026-28E (vs. our previous estimates of 43% / 50% / 55% in 2026-28E). Our global AI server volume forecast implies demand for AI chips at 19m/ 27m/ 32m in 2026-28E, which increases from our last update due to our revised outlook on AI server racks (NVL72-equivalent) and AI servers (8-GPU equivalent) across ASIC and GPU base. We forecast a positive demand trend for GPU and ASIC platforms, and new AI servers based on new networking architectures from 2027 which we expect will further enhance the advancement in computing performance. We expect the continuing AI infrastructure cycle to support growth through 2028E.

AI servers: Buy Wiwynn / Wistron (ODM), Hon Hai / FII (ODM), LandMark (Silicon photonics), VPEC (Silicon photonics), AVC / Fositek (liquid cooling), Auras (liquid cooling), King Slide (rail kits), Chenbro (chassis), EMC (CCL), GCE (PCB), Eoptolink (Optical connections), TSMC (foundry; on CL), MPI (probe cards), WinWay (sockets), Aspeed (fabless), and Hon Precision (FT handler).

Allen Chang

+852-2978-2930 | allen.k.chang@gs.com Goldman Sachs (Asia) L.L.C.

Verena Jeng

+852-2978-1681 | verena.jeng@gs.com Goldman Sachs (Asia) L.L.C.

James Schneider, Ph.D.

+1(212)357-2929 | jim.schneider@gs.com Goldman Sachs & Co. LLC

Michael Ng, CFA

+1(212)902-8618 | michael.ng@gs.com Goldman Sachs & Co. LLC

Katherine Murphy

+1(212)902-1151 | katherine.a.murphy@gs.com Goldman Sachs & Co. LLC

Ronald Keung, CFA

+852-2978-0856 | ronald.keung@gs.com Goldman Sachs (Asia) L.L.C.

Giuni Lee

+82(2)3788-1177 | giuni.lee@gs.com Goldman Sachs (Asia) L.L.C., Seoul Branch

Anmol Makkar

+1(212)357-1366 | anmol.makkar@gs.com Goldman Sachs & Co. LLC

Daiki Takayama

+81(3)4587-9870 | daiki.takayama@gs.com Goldman Sachs Japan Co., Ltd.

Ting Song

+852-2978-6466 | ting.song@gs.com Goldman Sachs (Asia) L.L.C.

Yifan Hu

+852-2978-0996 | yifan.hu@gs.com Goldman Sachs (Asia) L.L.C.

Zorayda Montemayor

+1(212)357-6403 | zorayda.montemayor@gs.com Goldman Sachs & Co. LLC

Evelyn Yu

+886(2)2730-4187 | evelyn.yu@gs.com Goldman Sachs (Asia) L.L.C., Taipei Branch

Ryan Huang, CFA

+886(2)2730-4084 | ryan.huang@gs.com Goldman Sachs (Asia) L.L.C., Taipei Branch

Goldman Sachs does and seeks to do business with companies covered in its research reports. As a result, investors should be aware that the fi rm may have a con fl ict of interest that could a ff ect the objectivity of this report. Investors should consider this report as only a single factor in making their investment decision. For Reg AC certi fi cation and other important disclosures, see the Disclosure Appendix, or go to ed as research e92c7a75ab8b4efbba794e6b187208c8

Exhibit 1: AI servers implied GPU vs. ASIC volume

Source: Goldman Sachs Global Investment Research

Related reports:

Exhibit 2: Global servers value TAM revisions

Source: Goldman Sachs Global Investment Research

e92c7a75ab8b4efbba794e6b187208c8

Servers: AI server racks +190%/+89%/+56% YoY in 2026-28E; General server volume +14%/ +6%/ +6% YoY in 2026-28E

Source: Company data, Goldman Sachs Global Investment Research

Server TAM estimate revisions

AI server racks: We raise our forecast for AI server racks powered by NVIDIA to 92k / 148k in 2027 / 28E, with our estimates being raised by 19% / 22% in 2027 / 28E, while keeping AI server racks powered by AMD largely unchanged at 5k / 13k / 15k in 2026-28E. With continuous speci fi cation upgrade, we expect the dollar content per AI server rack will continue to grow. Overall, we expect the AI server rack global value TAM to grow at 118% CAGR in 2025-28E to US$561bn, or 51% of the global server value TAM by 2028E. By vendor, we continue to expect global leading CSPs to be the major customers, with leading ODM direct players as the major suppliers. We expect Hon Hai / FII to maintain the leading market position among ODMs and with continuous speci fi cation upgrade driving its global market share from 55% in 2026E to 69% by 2028E.

e92c7a75ab8b4efbba794e6b187208c8

Exhibit 3: AI server racks global shipment by chipset platforms

Source: Goldman Sachs Global Investment Research

Exhibit 4: AI server rack shipment by ODM / OEMs

% in the chart indicates Hon Hai / FII global market share in AI server racks

Source: Goldman Sachs Global Investment Research

AI servers: We raise AI server (8-GPU) equivalent shipment by 18% / 24% / 11% in 2026-28E to 1.9m / 2.4m / 2.6m in 2026-28E, mainly driven by rising ASIC-based AI server demand. The custom-design provides ASIC-based AI servers optimized for AI computation with better e ffi ciency, such as lower latency, or memory savings via reduced expensive read/write cycles. We continue to see new generation ASICs launched by global leading CSPs, supporting AI server growth: we expect the AI servers (8-GPU) global value TAM to grow at 26% CAGR in 2025-28E to US$295bn, or 27% of global server value TAM by 2028E.

% in chart indicates AI server (8-GPU and NVL72) revenues mix

Source: Goldman Sachs Global Investment Research

% in chart indicates ASIC shipment contribution to total AI chips

Source: Goldman Sachs Global Investment Research

General servers: We raise general server revenue growth from +15% / +6% / +6% YoY in 2026-28E to +28% / +17% / +11% YoY in 2026-28E driven by high memory costs and ongoing speci fi cation upgrade driving AI workloads. Intel launches Intel Xeon 6+ processors in Computex 2026 (link) to facilitate Agentic AI: CPU servers are served as the control plane, driving system-level performance and e ffi ciency under rising agentic AI workloads, while supporting data movement and sustained inferencing. We expect general servers' global value TAM to grow at 18% CAGR in 2025-28E to US$245bn, or 22% of global server value TAM by 2028E.

e92c7a75ab8b4efbba794e6b187208c8

Exhibit 7: General server shipment vs. ASP YoY in 2026-28E

Source: Goldman Sachs Global Investment Research

Exhibit 8: x86 server shipment vs. ASP YoY in 2021-25A

Source: IDC

CSP capex: Regarding customer spending, our US internet team forecasts that leading US cloud service providers (CSPs) in aggregate will grow their capex budgets by 76%/ 35%/ 8% YoY in 2026-28E, with our estimates being raised by 8% / 27% / 25% in 2026-28E to US$1,145bn by 2028E. Within the China internet space, we expect leading Chinese cloud platforms to grow capex by 80%/ 20%/ 18% YoY in 2026-28E, with our estimates being raised by 37% / 44% / 55% in 2026-28E. The strong increase in CSP capex re fl ects the high memory costs and growing demand on AI workloads. In terms of capex structure, we estimate chipsets (GPU, ASIC, CPU) remains the major contributors, along with memory (HBM), and followed by networking and other components (e.g. cooling, power, chassis, rail kits, etc.). Compared to global CSPs, we estimate China CSP spending on chipsets would be more diversi fi ed across GPU and ASICs considering various local fabless suppliers and customers' varying demand on AI scenarios / tasks. Read more: Microsoft, Amazon, Meta, Alphabet, Oracle, Baba, Tencent, Baidu.

Exhibit 9: US CSP Capex trend

Data include Microsoft, Amazon, Meta, Alphabet and Oracle. For Microsoft, we use capex incl. fi nancial lease. In FY24, 50% of the capex of Microsoft was spent on land and 50% on AI/Cloud.

Source: Company data, Goldman Sachs Global Investment Research

Exhibit 10: China CSP capex trend

Data include Bytedance, Tencent, Alibaba, Baidu. We calculate Bytedance (private) relevant operating metrics by analyzing the industry and companies that our China Internet team cover, and then extrapolating them to Bytedance.

Source: Company data, Goldman Sachs Global Investment Research

e92c7a75ab8b4efbba794e6b187208c8

(2026E)

GPU + HBM

38%

ASIC+ HBM

$31,273

BITERDAR BEED BEI

31%

Exhibit 11: CSP capex structure

Global CSP Capex US$786.1bn

DITTODA COOD ACCE

GPU+HBM

$23,827

23%

(2026E)

EEL RELEEE

FEEEEEEEE

EEEEEEE

DITTOBOTE COOD ECLI

BUITTODOTE ERED NECE

INTERD

ПОВОЛ СОБО ДОЦі СО 0CO)

DIRED OBIE CEDO ACCE

DDAT CDDO СОСООТЕ СОДОКОТ

DITTODOTE COOD ECCI

General servers Others (Utility,

DITTODOTE CEDD ECCO

Chassis, rail

BUTTODO CERO COCU

DI DODORE CEDO COL

UDITODOTE CONDIRET

ODM/OEM

CPU & CPU Networking

NODAT СЕГО ПО

Cooling

China CSP Capex US$102.1bn

Source: Goldman Sachs Global Investment Research

TODD

Power

BITTOD

e92c7a75ab8b4efbba794e6b187208c8

PCs: -10% / +2%/ +3% YoY in 2026E-28E

Global PCs market opportunity

Shipments by markets do not include workstations.

Source: Company data, Goldman Sachs Global Investment Research e92c7a75ab8b4efbba794e6b187208c8

Smartphones: Shipments -10% / +3%/ +1% YoY in 2026-28E

Global Smartphone market opportunity

Source: Company data, Goldman Sachs Global Investment Research

e92c7a75ab8b4efbba794e6b187208c8

圖片清單（已驗證 2026-07-02）

回補驗證：僅涵蓋已被 lib 頁嵌入的圖片，非全量驗證。