260822_亞洲硬體_jpm_asian-hardware-technology

PDF 原檔：260822_亞洲硬體_jpm_asian-hardware-technology_original.pdf

圖片清單（已驗證 2026-08-24）

原始內容

ora

Asian Hardware Technology

Demand outlook strengthens across AI and general servers; supply constraints the key swing factor

Albert Hung AC +886 2 2725-9875 albert.hung@jpmorgan.com J.P. Morgan Securities (Taiwan) Limited

William Yang AC william.yang@jpmorgan.com J.P. Morgan Securities (Taiwan) Limited

Anthony Leng +886 2 2725-9240 anthony.leng@jpmorgan.com J.P. Morgan Securities (Taiwan) Limited

See the end pages of this presentation for analyst certification and important disclosures, including non-US analyst disclosures. J.P. Morgan does and seeks to do business with companies covered in its research reports. As a result, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of this report. Investors should consider this report as only a single factor in making their

Company

Delta

J.P. Morgan

Lenovo

Quanta

Asus

Analyst

Albert Hung

Albert Hung

Albert Hung

MSI

Asia tech valuations

Compal

Albert Hung

Inventec

ASPEED

Wistron

21Vianet

Accton

Lotes

TSMC

Hon Hai

Mediatek

UMC

ASE Holding

Novatek

Powerchip

Vanguard

GDS

Chipbond

Alchip

GUC

Nanya Tech

Rating

OW

OW

oW

N

N

OW

N

YTD return

%

81%

221%

20%

67%

31%

41%

46%

P/E

FY26E

40.5

12.1

12.4

11.1

15.9

18.4

11.1

P/B

FY26E

13.7

4.2

4.9

2.2

1.1

5.3

2.0

EPS (LCS)

FY26E

43.1

2.4

5.6

343.8

12.7

Meesnthenu

Pric

FY27E

26.4

10.8

13.4

FY27E

ROE (%)

FY26E

FY27E

37%

28%

45%

38%

15%

2.8

7.6

Company

Unimicron

J.P. Morgan

Yageo

Win Semi

EMC

Analyst

Jerry Tsai

Jerry Tsai

Jerry Tsai

Jerry Tsai

Tripod

Asia tech valuations

Flexium

TSMT

Chicony Power

CHPT

Sunonwealth

E Ink

Auras

AES

Kinik

GPTC

Rating

OW

OW

OW

OW

OW

OW

N

YTD return

%

418%

143%

107%

264%

176%

53%

-34%

P/E

FY26E

58.9

30.8

47.3

53.1

51.1

15.6

NA

FY27E

25.6

17.7

35.0

27.9

30.8

12.1

NA

P/B

FY26E

13.4

6.4

3.6

29.2

22.8

3.8

NA

FY27E

9.6

5.0

3.3

16.9

16.8

3.3

NA

ROE (%)

FY26E

26%

21%

8%

65%

50%

27%

N/A

EPS (LCS)

FY26E

FY27E

19.4

18.2

8.0

112.6

41.9

31.0

N/A

44.5

5

Ticker

Price target

(LC)

Mkt. cap

AI server: Robust demand outlook, Rubin generation ramp on schedule k units

Total datacenter Al chip volumes

YoY growth

High end datacenter GPU volumes

YoY growth

Nvidia

2.1022

800

775

2023

69%

3,632

1,931

1,911

2024

115%

7,815

4,950

4,500

2025

32%

10,320

6,150

5,500

15,895

2026E

54%

8,225

34%

7,425

2027E

25,631

10,781

61%

31%

9,281

Robust AI server demand outlook supported by strong Capex spending

x86-based

AMD

25

•  We model 45% AI chip shipment CAGR in 2025-28.

Oberon-like %

1,485

800

23%

80

10%

1,392

1,500

88%

750

50%

31,173

2028E 25-28E CAGR

45%

22%

12,673

10,673

18%

15%

9,606

90%

2,000

1.067

33%

1,400

•  GPU server demand remains robust this year on resilient US CSP AI spending and pent-up Neocloud/enterprise demand.
•  We see longer demand visibility into next year on strong demand outlook and sustained U.S. CSP capex spending growth.

ASIC chip volume

1,701

•  We forecast a 60% CAGR for AI ASIC chip shipments in 2025-28, driven by strong AWS Trainium/Google TPU demand.

Chip per server

Total Al server volumes

YoY growth

Source: J.P. Morgan estimates for all years, including historical. Note: Total AI server includes AI ASIC + high-end GPU AI servers.

27%

25%

45%

60%

1,911

20

-20%

4,374

2150%

24%

156%

141%

450

650

44%

$1,000

$800

58%

90%

80%

70%

Upbeat 2026 AI capex outlook, extended visibility into 2027

$400

$200

$300

$100

29%

40%

•  U.S. hyperscalers raised 2026 capex outlook and provided positive commentary on 2027.

$900

76%

79%

80%

Trillion USD

S0.3

•  JPM U.S. hardware team now forecast +80%/+50% YoY growth of datacenter capex in 2026/27, vs. previous estimates of +63%/+40% YoY, which is the third raise this year. We see ~10% upside to the estimates after the recent updated outlook for leading hyperscalers.
•  While negative FCF among leading hyperscalers over the next few quarters is inevitable, we see a few positive indicators to support continued AI investments, including robust core businesses, strong AI demand indicators, and decent returns of AI businesses with continued improvement.
•  Datacenter capex from Tier 2 providers (e.g. emerging neocloud and other data center-related companies) is now becoming a much more significant cohort.
•  We expect the group to grow capex by 65%/39% YoY to US$90bn/US$130bn in 2026/2027.
•  We estimate US$5.5tn of AI capex over 2026-30, with US$4.1tn funded through debt. We expect the IG corporate debt market to finance more than US$2.1tn of data center investments over the next five years

Top 4 US CSP/ISP cloud capex trend

Source: 650 Group and J.P. Morgan U.S. team estimates. Units in US$bn.

Anticipated AI Capex Funding Sources (2026-2030 total)

Source: J.P. Morgan U.S. team estimates.

$2.1

S0.4

$1.4

$5.5

700

2,000

J.P. Morgan

600

1,500

500

400

300

40%

450

400

350

35%

30%

27.1%

35.4%

23.7%

300

38.0%

34.5%

42%

35%

28%

AI-led accelerating cloud revenue growth with margin expansion despite capex hikes

500

200

100

0

12%

15%

100

50

Source: Company data, J.P. Morgan estimates.

Source: Company data, J.P. Morgan estimates.

Source: Company data, J.P. Morgan estimates.

Source: Company data, J.P. Morgan estimates.

16%

11%

15%

49%

14%

45%

15%

60%

48%

14%

41%

37.0%

37.4%

33.5%

mi eapenenturemireotmell eot e mene ly tegten

296

Own-firm 3%

396

8%

34%

11%

8%

39%

6%

14%

9%

morrerote lot hi carenenterennieeounent

Coding

31%

Customer Access

Product Development

19%

26%

4.5%

24%

42%

39%

5.8%

22%

26%

9%

5%

5%

7%

4%

90% of APAC respondents on board with AI, with upside in spending in the near future

• Very concerned • Somewhat concerned • Neutral • Somewhat optimistic • Very optimistic

2%

Labor Productivity

14%

392%

40%

40%

Last 12M

Next 12M

•  JPM Asia Pacific Artificial Intelligence Implementation Survey 2026 (link) suggests that AI investment is already mainstream, with 88% of APAC respondents reporting AI expenditure over the last 12 months, while only ~3% said they had not invested and ~7% were unsure. Key findings include:
•  AI investment is mainstream and growing fast, but leadership is uneven.
•  AI stated value case is productivity-led but with a growth mindset; so far cost reduction is secondary.
•  Execution capability will increasingly separate AI winners; most respondents were optimistic about profit implications.
•  Average AI spend is expected to rise from 4.5% of expenses plus capex over the last 12 months to 5.8% over the next 12 months.

29%

13%

90

80

Arm-based

70

60

50

40

30

20

10

GB300 NVL72

NLI< oom rack quarterly shipments (k units)

2H25

Niula oro servel rodumay

2H26

GPU: HBM3e: 288GB, TDP: up to 1.4kW

1.5x FP4 performance

Nvidia GPU roadmap: Rubin on schedule, Kyber architecture facing challenges

1.5x FP4 performance

17

19

•  Vera Rubin system ramps up on schedule in late 3Q to 4Q26
•  Our research indicates limited impact on the Vera Rubin ramp schedule from the recent leading PCB maker yield issue.
•  We estimate 70-80k NVL72 rack shipments this year, comprising 60-65k GB300 rack shipments and ~10k VR rack shipments.
•  We continue to forecast ~1.5mn Rubin chip production this year (board level output), but expect only ~10k VR NVL72 rack shipments due to production cycle time and component shortages.
•  We expect 85-95k NVL72 rack shipments in 2027, with VR200 rack being the majority.

Nvidia GPU server roadmap

Source: J.P. Morgan, Company data.

21

27

Vera-Rubin NVL72

Dual CX9 1.6T

GPU: HBM4: 288GB, TDP: up to 2. 3kW

CX8 800G

2H24

1425

GB200 NVL72

CX7400G

GPU: HBM3e: 192GB, TDP: 1.2kW

20

2H27

Vera-Rubin Ultra NVL72/576

77

GPU: HBM: TBD., TDP: up to 2.6kW

4*CX9, 3.2T

2H28

Feynman generation

CX10

In USS

NVL72 ODM BoM

J.P. Morgan

MVA fees

NVL72 ODM prices

ODM margin

GB200

3,261,618

144,527

3,406,145

4.29

GB300

3,934,640

148,114

4,082,754

3.69

VR200

6,300,268

174,883

6,475,151

2.7%

Deep dive into NVL72 BOM cost: GPU and Memory to be the key cost driver

Compute tray

Switch tray

5,480,100

485,964

•  We estimate the ODM price of the GB300/VR200 rack to be ~US$4.1/6.5mn, ~20%/90% higher than that of the GB200 server, primarily driven by GPU and memory price increases and content upgrades in the networking components and power and thermal systems.

Mix %

2,781,108

3,450,132

•  We expect AI server ODM GMs to decline to 3.6%/2.7% of GB300/VR200 vs. 4.2% in GB200.

Power supply system

1%

1%

1%

•  However, we see MVA dollars (GP dollars per rack) increasing in VR200, due to a more complex server architecture and longer testing time.
•  We estimate ~US$148k/ $175k for the MVA of GB300/VR200 vs. ~US$145k for GB200.

Total memory

HBM

NAND

DRAM (LPDDR)

Power supply system (power shelf/power whip/busbar)

Thermal (coldplate/manifold/QD)

Mechanical (chassis/ slide rail)

MVA

Others (compute tray contents excl. superchip, switch tray, etc.)

Source' .l P Moraan estimates

NVL72 BoM - Summary table

1.4%

2.5%

2.7%

Source: J.P. Morgan estimates.

J.P. Morgan

Google

TPU - inference

Total

Amazon

Inferentia

Total

2023

1,020

300

2024

900

420

Chip forecast (k)

2025

2026

1,500

2,600

100

Chip units per server

1/4/8

8

1/6/12

2023

255

36

291

38

2024

Implied server units (k)

225

338

2025

2026

375

650

113

53

88

463

13

•  We forecast AI ASIC chips to grow 77%/88% YoY in 2026/2027, mainly driven by strong AWS Trainium and Google TPU demand.

Meta

2028

1300

1900

600

199

199

•  As AI inference activities scale, we expect more end customers to develop its own AI inference ASIC solutions, which could further expand the AI ASIC server TAM.

MTIA

OpenAl

AI ASIC chip

Apple

AI ASIC chip

Grand total

Yoy

500

300

1,000

24

500

8

JPM ASIC chip/server estimates

300

50

500

13

6

63

Source: J.P. Morgan estimates.

150

200

4,170

6

8

21

38

42

63

125

175

825

2027

4,600

8,500

3,900

2028

5,200

10,000

4,800

2027

1150

488

1638

Microsoft

Microsoft

Meta

Major

Hon Hai

Quanta

Secondary

Quanta

Hon Hai

Quanta

Ai server vom maula " Niula s oro systell

ASIC servers

Nvidia NVL72

GB300

Secondary

Quanta

Hon Hai

Meta

Al

Quanta

Quanta

Consolidated market landscape in AI server ODM space

Google

Dell

Hon Hai

Quanta, Wiwynn, SMCI, Dell

Hon Hai

Quanta, SMCI

Major

Hon Hai

Quanta

Quanta

Secondary

Quanta

Hon Hai, (Wiwynn)

Hon Hai, Jabil

Hon Hai, Jabil, Inventec (board)

Hon Hai

Quanta

Quanta

Hon Hai, Wiwynn, (SMCI), (Dell), (Inventec)

Celestica (L6-L11), Flex (L6-L11), Hon Hai (L10.5), Inventec (L6)

Hon Hai (L10), Inventec (CPU board)

•  Hon Hai and Quanta stand out as preferred picks in the NVL72 supply chain, while Celestica and Wiwynn lead among ASIC plays.

Source: J.P. Morgan estimates. Note: Flex/ Inventec also supplies CPU boards for the MI300/400 system.

•  While we see tier-2 ODMs (e.g. Pegatron, Compal) managing to break into the space, complex design capabilities, frequent product upgrades, and high working capital requirements should keep the competitive landscape concentrated among a few major players.

Major

Secondary

AI server ODM matrix - AMD's GPU system

AI server ODM matrix - AMD's GPU system

Source: J.P. Morgan estimates. Note: Flex/ Inventec also supplies CPU boards for the MI300/400 system.

Microsoft

Meta

Oracle

GB200

Hon Hai

CPU

Hon Hai

VR200

General Servers: Multi-year strong demand driven by AI, key swing factor on supply constraints

Source: Nvidia.

• OBSERVE

CUM

•  As we migrate towards the era of agentic AI, focusing on test-time scaling (i.e. long thinking), CPUs are playing an increasingly important role in orchestration, running tools and skills, storage, and security. SECURITY & GOVERNANCE

• ACT

CLIM

•  For example, the Vera CPU not only sits in the NVL72 rack systems, but also serves as a standalone CPU across the entire Vera Rubin platform, including storage servers and CPU servers.

MEMORY

•  We attribute the strong general server demand this year (30-40% YoY growth on our estimates) primarily to the emerging demand for agentic AI.

NICIA AL Platform

TAT-LEN

Agent = LLM + Harness; CPU plays a critical role in the agentic AI ecosystem

Source: Nvidia.

NEMOTRON

OPEN MODELS

NEMOCLAW

Agentic AI CPUs becoming a new driver for server CPU demand

PROMPT

TOOLS & SKILLS

CUDA-X

LIBRARIES

Unit in millions

1. AI SERVER HEADNODE

Global Al accelerator Shipments (mn)

YoY

Avg accelerators per CPU in headnodes

Headnode Server CPUs Needed (mn)

2024

7.6

4.1

1.8

2025

9.6

27%

3.0

3.2

2026E

16.1

67%

2.3

7.1

Accelerating server CPU growth cycle on AI inference cycle

Traditional server unit for non Al workload (mn)

YoY

2%

3%

5%

2027E

25.2

2.2

56%

11.3

59%

13.1

5%

2028E 2025-28 CAGR

32.5

50%

1.9

29%

16.8

48%

13.8

5%

•  We segment CPU demand into three buckets-(1) AI server headnode CPUs, (2) regular server CPUs, and (3) agentic-AI CPUs.

Yoy

Yoy

YoY

3%

5%

5%

5%

•  As AI applications transition from training toward inference, we expect agentic-AI and AI headnode CPU demand to drive a 'super' general server CPU cycle. 0.0 0.0 0.17 1.6 0.45 7.3 0.68 17.1 0.83 27.0 155%
•  We forecast total server CPU shipments to grow from 26mn units in 2025 to 68mn by 2028 (38% CAGR) based on a top-down methodology, triangulated with bottom-up estimates. 16% 42% 42% 32%

Total server CPU shipment excl. Al server headnodes sen

YoY

20.3

22.5

29.2

40.1

51.1

31%

•  We forecast a 74%/5%/155% CAGR of AI server headnode CPU/ regular server CPU/ agentic AI server CPU demand in 2025-28.
•  We forecast a 53% CAGR of server CPU revenue TAM in 2025-2028, driven by the fast-growing server CPU demand and a ~10% CAGR of server CPU ASP expansion.

Source: J.P. Morgan estimates. Gartner® data for 2024 total server CPU shipments. Global AI accelerator shipments are based on our CoWoS estimates by the semi team.

74%

5%

12.5

Al accelerator unit (mn)

NVDA

HGX (x96 base)

AMD

NVL72 (ARM base)

TPU

v4/5/6/7

v8

vg

AWS

Tm 2

2025

6.2

2.

.8

3

.8

0.7

.8

7

2026E

8.9

3.0

5.9

0.7

27

cautious

2027E

10.0

2.1

.9

1.5

8.9

.7

Fast-growing AI accelerator shipments fueling the AI headnode CPU demand

Inferentia 2

Tm 3

Tm 4

Meta

Others

Total Al accelerator units (mn)

Avg number of accelerators per CPU

HGX (x86 base)

NVDA

NVL72 (ARM/ base)

AMD

TPU

v4/5/6/7

v8

v9

AWS

Tm 2

Inferentia 2

Tim 3

Tm 4

Meta

Microsoft

Others

Blended ratio

Al headnode CPU units (mn)

NVDA

AMD

TPU

AWS

Meta

Microsoft

Others

Total

0.6

1.3

3.5

•  Strong accelerator demand + Rising CPU attach rate = Exponential AI server headnode CPU growth.

0.2

0.4

0.8

Source: J.P. Morgan estimates for all years, including historical. AI accelerator unit estimates are based on our CoWoS estimates by the semi team.

0.5

3

0.1

5

2024

4.5

4.3

587

74

2028E

11.6

.0

2

.8

2.3

.6

Note

14

12

0

J.P. Morgan

64%

68%

66%

80%

70%

60%

50%

Super server cycle underpinned by the recent uptick in x86 CPU vendor revenue growth

20%

60

47

•  We have seen accelerating growth in Intel/AMD server datacenter CPU revenue growth in the past three quarters, driven by strong server demand.
•  AMD also cast an upbeat long-term outlook for server CPU TAM (US$220bn by 2030 or ~50%+ CAGR in 2025-2030), which underpins our research of the multi-year super server cycle across hyperscalers and the positive demand signals across the supply chain.

2024

2025

2026E

2027E

2028E

Source: Gartner®, J.P. Morgan estimates.

120

100

Source: J.P. Morgan U.S. semi team estimates, Company data. Note: AMD's datacenter revenue excludes GPU revenue.

100

5.0

4.0

3.0

2.0

1.0

-

volvel yuartery orptenlo, Xod allu tot dellu

18% 2121%

8%

29%9%

17%8%

15%

4°8% 1%

6%0%

40%

30%

20%

10%

30

25

20 - 3%

An AI-led multi-year general server upcycle

-5%.

-9%

-13%

-20%

10 -

5

30%

25%

20%

10%

15%

5%

-5%

0%

-10%

-15%

•  1Q26 U.S. CSP server shipments were better-than-seasonal at up QoQ and the overall server shipments appear to have grown double-digit QoQ in 2Q26. 2216
•  Initial outlook indicates server shipments should continue to grow into 2H26, led by U.S. CSP server demand.

• YY

-Y/Y

•  We have picked up incremental positive signs in server demand over the past 3-6 months with order visibility extended into next year.
•  Lotes and ASPEED's order visibility has already extended into 2027. Memory vendors also collected 50-80% demand growth feedback from the U.S. CSPs for next year. The key swing factor is supply chain constraints.
•  We anticipate an accelerating shipment growth of 25% YoY next year, driven by accumulated backlog from this year and strong server demand.
•  By customer type, we forecast 40-50% growth/ single digit decline to flattish for CSPs/enterprise server shipments in 2026/27, respectively.
•  In the longer run, we expect server shipments to sustain at double digit growth in 2028-29, driven by the multi-year AI-led server demand.

Source: Gartner®, J.P. Morgan estimates.

Source: Gartner®, J.P. Morgan estimates.

-4%

13%

vorvol emtteer ortpttollle anle tet urolte

22% 25%

19%

-3%

1%

2%

7%

-18%

6%

4%

9%

Datacenter Power: Secular TAM expansion in AI datacenters

350

60

300

250

40

200

150

30

100

20

0

0

Datacenter new power installed and YoY growth (GW)

Server power supply market landscape

Delta's server power supply TAM (US$bn)

42%

Others, 22%

50

36%

45%

40%

Datacenter Power TAM reaches US$50bn by 2028E, growing at 80% CAGR

Chicony Power,

19%

25%

•  We forecast Delta's AC/DC server power revenue to grow at 75-80% CAGR in 2025-28, largely in line with our server power supply TAM CAGR at 80%. 9 5%

5

5

6

0%

•  Our datacenter power TAM is driven by three compounders:

2022

2023

2024

2025

2026E

2027E

2028E

—YoY

•  Accelerator shipment growth: 40-50% CAGR in 2025-28E based on our latest CoWoS updates
•  TDP growth: 30-50% CAGR on continued spec upgrades
•  Power architectural change: Rising power rack adoption (JPMe 20% in Vera Rubin and 50%+ in VR Ultra cycle)

Datacenter new power installed and YoY growth (GW)

Source: IEA, J.P. Morgan estimates.

Server power supply market landscape

Source: Company data, J.P. Morgan estimates. Note: Lite on's server PSU revenue is based on company guidance. Chicony's implied market share based on J.P. Morgan coverage analyst Jerry Tsai's revenue estimates for the company (3412 TT, OW).

Delta's server power supply TAM (US$bn)

Source: J.P. Morgan estimates for all years, including historical.

AI ASIC blended TDP trend (Watt)

Datacenter Al chip shipment growth (mn units for volume)

Nvidia's high-end GPU TDP

J.P. Morgan

35

3000

1,800

30

2500

1,600

25

20

2000

1500

15

1000

10

1,000

800

5

500

600

-

0

400

200

0

115%

140%

2600

120%

25-27 = ~35-40%

2025-28 CAGR = 45-50%

20-25 CAGR = 25-30%

2300

1,600

Datacenter Power growth driver 1 and 2: Accelerator shipment growth + TDP expansion

16-20 CAGR = 15%

54%

1400

•  We forecast ~45% CAGR for AI accelerator shipments based on the latest CoWoS update (link) (up from our prior estimate of ~30% CAGR). 500

0%

•  We expect a ~30% CAGR for AI accelerators' TDP expansion, driven by continued spec upgrades to address training and inference workloads.

2025

2026E

2027E

2028E

•  We see a faster TDP expansion for AI ASIC chips as hyperscalers start to leverage their ASIC solutions for AI training/inference tasks.

Source: Company data, J.P. Morgan estimates.

Datacenter AI chip shipment growth (mn units for volume)

Source: J.P. Morgan estimates for all years, including historical.

Source: J.P. Morgan estimates for all years, including historical.

J.P. Morgan

Real server Power consumption excl. redundancy (kw

Design power capacity per rack

Nvidia's Al server rack power roadmap and estimated power content changes

DGX A100

71

13

DGX B200/B300

14

20

GB200 NVL72

120

198

GB300 NVL72

140

288

Datacenter Power growth driver 3: Power supply configuration change

Redundancy spec

PSU specification (kw)

PSU ASP (USS)

Components

SOKW BBU (20kW*4)

Rack chassis

Liquid cooling busbar

Power whip

N +1 (3+1)

3.3

Nvidia's AI server rack power roadmap and estimated power content changes

VR200 Power shelf

210

444

Power shelf

4*111kw

N+2 (4+2)

18.5

2,600

VR Ultra

900

1500

800VHVDC power rack

10*150kw

N+2 (4+2)

25

3,750

DC/DC cross link cable

Others

Power vendor assembly and testing

3,600

3% 1004

1% 125A

Power rack BoM cost breakdown (800V 660kw)

Source: Nvidia, J.P. Morgan estimates. Note: 1 660kW power rack can support 2 Vera Rubin NVL72 racks.

600

600

12

6

N+2 (4+2)

N+2 (4+2)

N+1 (5+1)

N+2 (4+2)

N+2 (4+2)

DGX H100/H200

10

H20

5

J.P. Morgan

10 kV / 13.8 kV /

20 kV / 34.5 kV

GPU

Today

Source: Delta Electronics.

Power architecture migration in Al servers

Distribution

Rack busbar

Server

Datacenter Networking: Whitebox Switching Enters a New Growth Cycle

J.P. Morgan

35

AI-driven Networking upgrades expand whitebox switching market opportunity

•  We estimate ~US$6bn datacenter networking serviceable addressable market (SAM including generic ethernet switch, AI back-end and front-end networking) for the top 4 CSPs for Accton in 2025 and expect the market to grow at a ~75% CAGR to reach ~ $33bn by 2028, driven by
•  Strong AI-driven general purpose server demand
•  AI front-end and back-end networking upgrades
•  AI-led accelerating network upgrade cycles

2022

2023

2024

2025

2026E

2027E

2028E

Source: J.P. Morgan estimates, Gartner®. Note: Calculation of Accton' s datacenter switch SAM only includes whitebox ToR switches for scale-out fabrics. AI ASIC scale-out networking switch SAM only includes AWS ASIC back-end ToR switch TAM, while we include AI front-end scale-out SAM into generic whitebox Ethernet switch categories..

Accton's datacenter switch revenue SAM and YoY growth

CSP General server shipment trend and CAGR

238%

Switch ODM-direct vs. server ODM-direct spending growth

J.P. Morgan

K units

21,000

40%

45%

AI-driven general purpose server demand sets the stage for switching growth

127%

30%

•  We expect general server shipments to accelerate over the next few years on AI-driven inference demand.

9,000

•  We estimate CSP's general server shipment growth of 40%/25%/20% in 2026/27/28.

50%

6,000

0%

3,000

•  There could be upside risks to our general server shipment assumptions driven by fast-growing agentic AI demand.

-11%

5%

•  Over the past 2-3 years, CSPs have prioritized AI server investment and data center build-outs, while networking has lagged.

2025

2026E

2027E

2028E

•  Gartner ® data shows ODM-direct server spending increased 207% / 127% in 2024/25 versus ODM direct switch spending growth of 18%/103% over the same period.
•  We expect CSPs to step up networking spending to reduce latency and improve data transmission rates, presenting huge growth opportunities for whitebox Ethernet switches.

CSP General server shipment trend and CAGR

Source: Gartner®, J.P. Morgan estimates.

Switch ODM-direct vs. server ODM-direct spending growth

Source: Gartner®.

-50%

300%

43%

Gb

Tb

500

3500

3000

400

2500

2000

300

1500

200

1000

100

500

0

0

Front-end network and back-end network

Nvidia's GPU scale out bandwidth - NVL72/144 rack

Nvidia GPU scale out bandwidth - per chip

Cloud

Backbone

Frontend

Super Spine

461

3200

AI/ML drives both front-end and back-end networking upgrades

Spine

•  Higher bandwidth requirements and data volumes in AI datacenters will drive continued upgrades in the traditional frontend networks and create an incremental back-end GPU fabric.
•  Exponential growth of Nvidia's GPU scale-out bandwidth.
•  We forecast a ~164% CAGR for the GPU back-end Ethernet ToR switch TAM over 2025-28.

Spine

•  Examples of front-end network upgrades:
• -BlueField 3 400G in GB200/300 moving to BlueField 4 800G in VR200
• -CSP's in-house NICs moving from 200G to 400G

Nvidia's GPU scale out bandwidth - NVL72/144 rack

Source: J.P. Morgan estimates, Company data

Front-end network and back-end network

Source: Cisco

Nvidia GPU scale out bandwidth - per chip

Source: J.P. Morgan estimates, Company data

Blended ASP of Accton's datacenter switch port

Switch port ASP growth by customer type

Datacenter switch technology roadmap

J.P. Morgan

30%

Speed

US$,%

250

100G

20%

10%

150

0%

100

-10%

800G

-20%

50

-30%

1.6T

2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025E 2026E 2027E

35%

20%

18%

40%

Port ASP expansion driven by AI-led accelerating network upgrade cycles

9%

14%

20%

•  A shortening network bandwidth upgrade cadence is supportive of data center switch ASP expansion. -5% -9% -5% -9% 0%
•  7 years for 100G/200G transition versus <2 years for 800G to 1.6T

-24%

•  Rising mix of higher-speed ports supports ASP expansion

• YoY

•2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 • 2025

•  We see Accton as a key beneficiary, given its strong positioning in whitebox data center switches. (JPMe ~40% of total revenue in 3Q26).
•  We forecast a ~20% CAGR in Accton's switch port ASP over 202528, primarily driven by the 800G/1.6T upgrade cycle.

Datacenter switch technology roadmap

Source: 650 Group, J.P. Morgan estimates

24%

Switch port ASP growth by customer type

Source: J.P. Morgan estimates.

Source: Gartner®.

Blended ASP of Accton's datacenter switch port

Source: J.P. Morgan estimates, Company data

PCs: Unit downside in sight, but margin/value better than feared

100

J.P. Morgan

80

60

40

20

49%

8%2%

70%

50%

11% 11%

9%

Better-than-feared PC demand in 1H26 likely driven by pull-in orders, cautious on 2H26 PC demand 10% 150 5% 8% 0%

-9971%1%1%

-1%

2%1%1%

0%

-8%

-3%

50

-14%

-15%

-10%

•  1H26 PC shipments appear better-than-expected, likely due to continued Win10 replacement demand and the pull-in of orders ahead of the memory price hikes. -50 -15% 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026Е2027E2028E2029E -20%
•  We are cautious on PC end demand in 2H26 due to pull-forward demand in 1H26, the negative price elasticity from key component cost hikes, and late cycle of Win10 PC replacement. ·Q/Q - Y/Y -YN YN
•  We anticipate a sub-seasonal shipment trend in 2H26 and forecast a 3% HoH decline for PC shipments.
•  We forecast an 8% PC shipment decline this year, comprising 14%/4% consumer/commercial PC shipment decline.

PC + Chromebook quarterly shipments and QoQ/YoY trend

Source: Gartner®, J.P. Morgan estimates.

PC incl./excl. Chromebook annual shipments and YoY trend

Source: Gartner®, J.P. Morgan estimates.

-1%

-10%

350

1%

12%

/2%

2%

2%

20%

200

350%

300%

150

100

12%

6%

15%

10%

10%

20%

15%

10%

Positive ASP y-y trend from 1Q24

400%

350%

3Q16-4Q17

Darents Mile

3Q23-3Q24

Commercial PC demand more resilient driven by Windows 10 expiry and lower price sensitivity from memory-led product price hikes 0%

50%

0%

-50%

50

-100%

-

50%

0%

40

-50%

-100%

Source: WSTS, J.P. Morgan estimates.

-5%

Source: Gartner®, J.P. Morgan estimates.

NAND ASP Y/Y% change vs previous assumptions

Source: WSTS, J.P. Morgan estimates.

Source: Gartner®, J.P. Morgan estimates.

DRAM ASP Y/Y% change vs previous assumptions

Source: WSTS, J.P. Morgan estimates.

-10%

140

14%

25%

J.P. Morgan

20%

15%

10%

5%-

0%

-5%

-10%

11%

Unit downside in sight, but PC brands' margin/value better than feared

2%

1%

0%%

•  Asian PC brands delivered better-than-expected PC margins in 1H26, driven by better-than-feared PC demand, continued product price hikes, and product mix adjustments. -3%% 1%

-6%

-6%

•  Our research indicates 20-30% YoY growth of PC ASP in 1H26. This, along with the lower-cost PC inventory at hand, drove meaningful margin expansion for brands.

• Notebook ASP Y/Y

•  In contrast, we saw PC ODM margin dilution in 1H26 due to the ASP hikes, while profit dollars remain intact.

Courco: Cartnor@) | D Mornan estimatocl

•Acer —

Microstar

Caurco: Camnany data l D Maraan ectimatae Aeer'e 20.1026 data is hacad en Rlnnmhora

•  Looking ahead, while we have seen weakening PC demand in 2H26, we expect PC brands margin trend likely to be better than feared due to continued product mix adjustments with pricing actions.
•  We expect tier one PC brands' margins to be more resilient in 2H26 due to better supply chain capability.

Blended PC ASP and YoY trend

Source: Gartner®, J.P. Morgan estimates.

Source: Company data. J.P. Morgan estimates. Acer's 2Q-4Q26 data is based on Bloomberg consensus estimates.

23%

17%

11%

9%

CPU

Memory total

SSD

DRAM

2025

250

75

35

40

% of BoM

26%

8%

4%

4%

Note

Raptor Lake

512G

LPDDR5 16G

2026

300

310

125

185

% of BoM

24%

25%

10%

15%

Note

Arrow Lake

512G

LPDDR5 16G, JPMe $11.6/GB

NB BoM analysis supports a 20-30% PC product price hikes by brands in 1H26

Panel

45

5%

13.4"

43

3%

13.4", assume 5% decline

•  We estimate that the BoM cost of high-end NBs has increased by ~30% over the past year, mainly driven by the price hikes in memory, non-key semis (analog ICs), software license fees, and other components (e.g. passive components, PCBs).

Dongle

Software

Others

Warranty

•  We estimate memory cost to increase by US$200-250 after 3-5x DRAM/NAND price hikes over the past year, driving the memory as % of total NB BoM to rise to 20-30% vs. HSD % a year ago. 237 92 19% 7% MSFT lincense fee cost hike by 5-10%
•  We also assume a 5% price hike for other components & non-key semis.

45

100

1254

•  The BoM cost increases have also driven PC brands to implement 20-30% price hikes in 1H26 to protect margins
•  We believe mid to low range PCs could have a higher sensitivity to the memory BoM % increase given the lower PC price band.
•  PC brands have mitigated the impact through spec downgrades and product mix adjustments.

Source: Gartner®, J.P. Morgan estimates.

4%

8%

MVA, logistics etc.

25,000

J.P. Morgan

20,000

27%

25%

10%

40%

28%

16%

60,000

50,000

40,000

Graphics/Motherboards - Soft end demand in the absence of new product ramps

10,000

5,000

-

-12%

-8%

20,000

-14%

40%

28%

16%

4%

8%

•  We have seen weakening end demand for PC motherboard and VGA in 1H26 due to memory supply shortage and negative price elasticity from product price hikes.

-

-32%

-32%

•  Looking ahead, we expect a sub-seasonal shipment growth (~5% QoQ growth) for the top 3 MB/VGA vendor shipments in 3Q26 and continue to forecast a double-digit shipment decline for the full year MB/VGA shipments. · Aggregated Top 3 motherboard vendor shipment (K unit)
•  We see limited catalysts for the market given the lack of new product launches in the next six months.
•  This, coupled with the memory-led end demand weakness, implies potential margin pressure for PC component brands.

Source: J.P. Morgan estimates, Company data.

Source: J.P. Morgan estimates, Company data.

18%

29%

Individual stock one-pager - Albert Hung

250

200

150

100

50

0

•  Rising power consumption of AI servers

86%

71%

53%

40%

53%

2%

Delta: Key beneficiary of secular AI datacenter power and thermal trend

2% 5%

10%

59%

2%

8%

6%

22%

58%

3%

7%

7%

24%

•  We believe Delta will be a major beneficiary of the fast-growing datacenter power supply TAM (JPMe +80% CAGR in 2025-28), driven by

3%

3%

2024

2023

•  Non-PSU content increase from power rack adoption and HVDC material upgrades

• Server Al power - DC/DC

2026E

2027E

• Liquid cooling i Server Al power - AC/DC

•  Delta possesses the leading position (40%+) in datacenter power (esp. in high-end market) and liquid cooling space.
•  We model above-seasonal revenue growth in 3Q/4Q at +16%/+14% QoQ respectively, driven by order push-out from 2Q26, new AI ASIC project ramp in 2H, 110kW power shelf / HVDC product ramp in 4Q, and product price hikes in late 3Q26.
•  We anticipate a gradual margin recovery helped by product price hikes and favorable pricing terms for new products.
•  We expect the stock to regain momentum over the next three months as datacenter related revenues re-accelerate. Stay OW.

Source: Company data, J.P. Morgan estimates.

Source: Company data, J.P. Morgan estimates.

100%

40%

35%

2025

2028E

65%

55%

15%

5%

-5%

J.P. Morgan

20,000

155%

109%

180%

• 130%

50

40

ASPEED: Supply unlock in 2H26 drives revenue upsides

30%

•  We have seen potential revenue upside near term, driven by robust general server demand, new AI server product cycles, and better supply in 2H26.

25

•  We forecast 19%/33% QoQ revenue growth in 3Q26/4Q26, and early feedback suggests momentum could extend into 1H27.
•  We estimate ~38% shipment CAGR for ASPEED's regular BMC over 2025-28, broadly in line with our server CPU shipment growth forecast.
•  ASPEED's BMC shipments should outgrow CPU volumes due to higher exposure to U.S. CSPs, market share gains, and expanding applications.
•  We expect AST2700 penetration to accelerate in 2027 (JPMe ~25%/50% in 27/28). From the meaningful ASP premium for AST2700, the ramp should be a clear tailwind for ASPEED's top-line growth.
•  ASPEED remains a key beneficiary of the secular server CPU upcycle in our coverage. Stay OW.

Source: Company data, J.P. Morgan estimates.

Source: Company data, J.P. Morgan estimates.

AST2500

AST2600

>4x content growth

24

AST2700

43

10

8

90

80

6,000

5,000

J.P. Morgan

70

60

4,000

50

40

30

20

-

10

-

(1,000)

37%

37%

120%

12%

8%

9.3% 7. 1%6.5%

110%

-0.5%0.4%

• 90%

50%

30%

4%

60%

-21% -1% 1.0%

63%

-2.8%

• 0%.

Lenovo: ISG margin inflection confirmed; AI pipeline doubles

• 0%

•  Lenovo's AI server pipeline more than doubled to US$54bn exiting Jun-Q (vs. US$21bn last Q). Management expects meaningful revenue conversion in 2HFY27.

•  We estimate110%/50% YoY ISG revenue growth in FY27/28.
•  We expect a significant OPM improvement for ISG business to 9.3% in FY27 from 0.4% in FY26, driven by favorable pricing and OP leverage.
•  Beyond FY27, we anticipate a sustained ISG growth momentum on the multi-year server cycle with continued margin improvement on better scale.
•  IDG profitability also appears better than feared, as Lenovo has been able to outgrow its peers through better supply chain management and scale.
•  Looking ahead, we expect continued growth in 2QFY27 as strong ISG/SSG momentum more than offsets seasonal IDG softness.
•  We expect ISG to become a key earnings driver over the medium term. IDG profitability also appears better than feared. Stay OW

Lenovo's ISG revenue and YoY Trend

Source: Company data, J.P. Morgan estimates.

Lenovo's ISG operating income & OPM trend

Source: Company data, J.P. Morgan estimates.

100

80

60

40

20

62%

68%

74%

140%

120%

100%

80%

450

400

350

300

80%

160%

140%

120%

100%

Accton: Riding AI-driven networking upgrade and unlocking new TAM in server racks

2% 70%

20%

20%

-6%

•  We forecast Accton to outgrow the whitebox switch market, with a 2025-28 datacenter switch revenue CAGR of ~92% versus an industry TAM CAGR of ~75%, supported by -40%

2027

2019 2020 2021 2022 2023 2024 2025 2026

• Accton's datacener switch revenue (NT$bn)

•  Better-positioned in high-end datacenter switch market (400G+)
•  Consolidated switch ODM market & Accton's market share rebound in Meta 800G projects

Source: Company data, J.P. Morgan estimates.

Source: Company data, J.P. Morgan estimates.

•  Amazon and Meta raised capex guidance for this year and the implied 2H26 capex uptick indicates sustained revenue momentum for Accton.
•  We believe market expectations on AWS T3 supply chain remain undemanding and anticipate the stock to ride on T3 ramp up in 2H26.
•  Beyond 2026, Accton plans to expand its capacity by 30%+ next year, which could help sustain Accton's growth momentum in the next 1-2 years.OW.

Accton's earnings and yoy trend

Source: Company data, J.P. Morgan estimates.

Accton to outgrow datacenter switch market with 92% rev CAGR

Source: Company data, J.P. Morgan estimates.

-YoY

2028

120%

93%

137%

85%

' 600

500

400

300

119%

250%

• 200%

• 150%

300

250

200

Wiwynn: Well positioned across GPU, AI ASIC, and general servers

200

100

50%

•  We expect Wiwynn to outperform on new AI server ramps (both ASIC and GPU projects) and a multi-year general server upcycle.

50

•  There could be a re-rating opportunity on the new U.S. CSP ASIC project ramp.

0

•  We expect GPU server to re-accelerate in 4Q26, supported by the initial ramp of Nvidia/AMD GPU servers.

GPU server revenue (NT$bn)

3%

223%

166%

•  Overall, we forecast AI server revenue growth of ~73%/~110% in 2026/27.
•  Importantly, our industry research suggests Wiwynn is participating in the bidding process for Google's TPU v8/v9 projects (TPU trays).
•  We model a 26% CAGR in Wiwynn's general server shipments over 2026-28, supported by inference-driven server demand.
•  We expect the stock to continue outperforming over the next 3-6 months on new AI server ramps and a multi-year general server upcycle. Stay OW.

Source: Company data, J.P. Morgan estimates.

Source: Company data. J.P. Morgan estimates.

0%

2026E

2027E

• GPU server % of total revenue

22%

25%

2028E

30%

25%

20%

15%

10%

5%

0%

20,000

15,000

Loles o earmings ana yoy trena

80%

66%

29% 31% 27%

20%

90%

70%

• 50%.

25%

Lutes o server gross prom and server or controution 7

20,000

16,000

18,000

14,000

47%

Lotes: Near-term GM headwind, eyes on strong server demand and AI optionality

•  Near-term margin weighed down by raw material price increases, depletion of low-cost inventory, and ramp-up costs for new products.
•  We expect stable GMs and a mild OPM rebound in 3Q26, with a more meaningful margin recovery in 1H27 as new product cycles ramp.

2020

2021

2022

2023

2024

2025

2026E 2027E

•  We view general server demand intact and keep our forecast of 20-30% server shipment growth over 2025-28, despite near-term supply uncertainty.
•  2027 order outlook remains strong with new platform penetration rates likely to reach 30%+

Source: Company data, J.P. Morgan estimates.

Source: Company data, J.P. Morgan estimates.

•  Lotes has gained market share in the existing Intel/AMD server platforms due to its competitive pricing strategy.
•  Looking ahead, management expects 35-40% m/s in the next-gen Intel/AMD server platforms.
•  We believe AI products could bring potential upside to Lotes' business, driven by SOCAMM / QD ramp and mass production of OE socket in 2H27.

Source: Company data, J.P. Morgan estimates.

Lotes's server gross profit and server GP contribution %

Source: Company data, J.P. Morgan estimates.

69%

80%

70%

48%

63%

J.P.Morgan

Apple Supply Chain

(mn units)

iPhone EMS build (LHS)

YoY (RHS)

(mn units)

iPhone 15 mini - 2U21

iPhone 13 - 2021

"Phane SES

3

iPhone EMS build plan

iPhone 15 Plus - 2023

iPhone 16 Plus - 2024

iPhone 16 - 2024

iPhone 16 Pra -2024

iPhone 16 Pra Max - 2024

"Phane 16e iPhone 17 - 2025

iPhone 17 Air - 2025

iPhone 17 Pro - 2025

iPhone 17 Pra Max - 2025

iPhone 17e

Foldable iPhone - 2026

iPhone 18 Pra - 2026

iPhone 18 Pro Max - 2026

Mix (%)

iPhane 13 - 2021

¡Phane 15 mini - 2021

"Phane SEJ

iPhone 14-2022

Phane 15 - 2023

iPhone 14 Plus - 2022

tane 15 Plus - 20

чане 16 - 210

iPhone 16 Plus - 2024

iPhane 16 Pra -2024

iPhane 16 Pra Max - 2024

iPhone 16e iPhone 17 - 2025

iPhone 17 Air - 2025

iPhone 17 Pro -2025

iPhone 17 Pra Max - 2025

iPhone 17e

Foldable iPhone - 202G

iPhone 18 Pro - 2026

iPhone 18 Pra Max - 2026

New model (on units)

2021 - 13 series

2022 - SE3

2022 - 14 series

2023 - 15 series

2024 - 16 series

2025 - 16e

2025 - 17 series

2026 - 17e

2026 - 18 series

10

Source: J.P. Morgan estimates. Note: As this data is based on our channel checks, both actuals and forecasts are from J.P. Morgan.

1023

60

1%

2023

40

-17%

3023

50

-10%

1Q23|

88

11%

1Q24

49

-17%

2024

46

15%

3024

58

15%

4

7624

83

• 5%

1025

56

14%

2025

48

4%

3025

52

-10%

40,25

94

13%

1026

58

3%

2026E

55

16%

3Q2GE

52

0%

1Q2GE

85

• 10%

21

232

7%

2022

243

5%

2023

238

• 2%

2024

236

1%

2025

250

6%

2026E

250

0%

Launch time

Model

Front camera

Face ID (Rx)

• Resolution

Selfie

13e

Face ID

(Rx+Tx)

HD

1H27E

iPhone 13

iPhone 13 Air

Face ID + Sellie can

Face ID

(Rx+Tx)

HD

Face ID

(Rx+Tx)

HD

2H27E

iPhone 20 Pro iPhone 20 Pro

Max

Face ID + Seife cam

Face ID

Face ID

(Rx+Tx)

HD

iPhone 17 Pro

Max

Face ID

(Rx+Tx)

HD

1H26

17e

Face ID +

Selfie cam

Face ID

(Rx+ Tx)

HD

2H26E

iPhone 13 Pro iPhone 18 Pro

Max

Face ID + Selfie cam

Face ID

Face ID

(Rx+Tx)

HD

(Rx+Tx)

HD

Foldable iPhone

Dư ai cam

NIA

2H25

iPhone 17 iPhone 17 Air iPhone 17 Pro

Face ID + Se/fe cam

Face ID

(Rx+Tx)

HD

Face ID

Face ID

(Rx+ Tx)

HD

(Rx+Tx)

HD

iPhone 17, 18, 20 series roadmap, with continued lens spec migrations

• Elemerts

• Elemerts

GP

EP

•  Expect more iPhones to be launched in 2026E and 2027E:

Rear camera

Wicke

• Resolution

• Elemerts

• OIS

Telephoto

• Elemerts

Ultra wide

• Resolution

• Elemerts

TOF

• Resolution

• Elemerts

Triple cam +

Triple cam +

EP

Triple cam +

Triple cam +

•  4 models in 2026E: iPhone 17e in 1H26 + 1 st foldable iPhone / 18 Pro / 18 Pro Max in 2H26

48MPx

48MPx

6P

Wide

48MP x

GP

Single cam

Wide

48MPx

Triple cam +

ToF

(Rx+Tx)

HD

Metalens

Selfie

24MP x

Triple cam +

ToF

Wide

Wicke

48MPx, 7P unchanged, but a

48MP x

48MPx

48MP x

Foldable iPhone

Du ai cam

NIA

Selfie

24MPx

GP

Urclerdisplay

2-MPx

EP

Du al cam

Wide

48MPx

48MPx

48MP x

•  6 models in 2027E & onwards: iPhone 18e / 18 Air / 18 (base model) in 1H27 + next-gen foldable iPhone / 20 Pro / 20 Pro Max in 2H27

Sensor shift

Sensor shitt

No sensar shif

Sensor shif

Sensor shift

Sensor shif

•  Gradual lens spec improvements in iPhone 17-20 series:

1G3P + 1 prism 1G3P + 1 prism

Sensor shift

Folded

43VP x

1G3P + 1 prism 1G3P + 1 prism

Sensor shift

Folded

Sensor shift

Folded

43MPx

43VP x

1G3P + 1 prism 1G3P + 1 prism

•  17 series in 2H25: 1) front-end selfie camera will upgrade to 24MP (from 12 MP) for all 4 models, 2) folded camera (periscope) of Pro models will upgrade to 48 MP (from 24 MP in 16 Pro models). 6P ToF ToF 6P NIA NIA GP NIA NIA aperture ToF ToF 6P NIA

2MP x

2MPX

2MP X

2MPX

2MP ×

2MP X

•  18 series in 2H26E: 1) wide cam of Pro models will become 7P 'variable aperture', 2) foldable model will have dual front cam with underdisplay.
•  20 series in 2H27E: 1) wide camera of Pro models will remain at 7P with 48 MP resolution, but upgrading to a much larger lens size, 2) ultra wide cameras of Pro models will remain at 6P with 48 MP resolution, but upgrading to a larger aperture.

Source: J.P. Morgan estimates, Company data. Note: Red-colored text refers to spec upgrades.

No sensor shit

NIA

No sensor shift

NIA

Sensor shitt

NIA

Sensor shift

NIA

No sensor shit

NIA

GP

Underdisplay

24MPx

Liquid cooling components

J.P. Morgan

Cold plate module

(with inner manifold)

Liquid cooling component overview - take GB-series rack for example

Quick

cooling)

Source: Auras, Asus company reports, websites.

J.P. Morgan

NVIDIA GB300 Compute Tray

NVIDIA Vera Rubin Compute Tray

Source: Computex event.

Source: Company data, J.P. Morgan estimates.

Processor

Socket

Motherboard

Backplate ion report (link).

Appendix

Quanta

Compal

Compal

Wistron (excl. Wiwynn)

Lenovo

HP

Dell

10-15%

Wistron

China

80%

68%

70%

15-20%

India

Huagin

Other Asia Mexico

10%

10-15%

Vietnam

10%

Inventec

30%

30%

NB ODM capacity breakdown and market landscape

ASUS

Acer

<5%

40-45%

Pegatron

10%

35-40%

ASUSTek

5-10%

•  We believe NB ODMs will continue to expand the non-China NB capacity based on customer requests. We expect non-China NB capacity to reach 30%+ by end of this year. Lenovo 50-80% 10-20% 5% 5% 5%
•  Among NB ODMs, Hon Hai has the highest non-China NB capacity exposure (50% in Vietnam), followed by Wistron/Compal/Inventec at 30-40% non-China capacity.
•  In contrast, Pegatron and Quanta have slower capacity relocation, which could lead to potential market share losses.
•  There could be some margin impact from production relocation costs in the mid to longer term.

Source: J.P. Morgan estimates.

Quanta

Hon Hai

NB ODM/OEM capacity breakdown by region

Source: J.P. Morgan estimates, Company data. Note: PC brand capacity breakdown includes both in-house and outsourcing capacity. Other Asia includes Taiwan,

Thailand, India, etc.

U.S.

Wingtech/

Europe

Luxshare

Others/in-

RoW

house

60%

2%

CoWos Consumption by customer (k wafer per year)

NVIDIA

CoWoS forecast

CoWos-L

Outside TSMC (CoWoS-S/R)

AMD

CoWos-s

CoWos-L

Broadcom

CoWos-S at TSMC

CoWos-S outside TSMC

CoWos-L

Alchip + AWS

Marvell (incl AWS for Trn2)

MediaTek

Others

Overall Total CoWos Consumption yoy

Total TSMC

Total Others (OSATs)

Overall Total CoWos Consumption

Source: J.P. Morgan estimates.

CoWoS allocation breakdown

2027E

2028E

2026E

1196

1076

40

1735

1385

105

% Revision vs prior ests

2027E

1%

1%

100%

JPM Comments

2026 CoWoS-L at TSMC, others at OSAT, 2027 Rubin Ultra assuming 2-die-on-interposer solution.

TSMC adding some CoWo-R to support Vera and future IPX projects now.

1280

0%

4%

Majority of CoWos to migrate to COWOS-L by 2H25. Rubin Ultra likely to remain on CoWOS-L Feynman likely to

CoWoS allocation breakdown

2024

192

127

97

30

65

2025

400

385

30

355

15

2026E

725

685

10

675

40

Source: J.P. Morgan estimates.

7%

8%

2023

70

70

70

Compute tray

GB300 Superchip

SSD

Main SSD drives

Boot drives

Fans

BMC

PDB

Busbar

HMC

ASP (S)

85,000

960

1,152

576

20

16

GB300 server BoM analysis

Networking cards

BlueField 3

CX7/8

Cold plate

Thermal module

Manifold - GB300 server

Chassis

Universal Quick Disconnect

Cables/Connectors

Others

Slide rail

Compute tray BoM (S)

Total compute tray BoM cost (5)

MVA/ Assemblyfees

NVLink switch tray

N/Link Switch chip

CPu

Fans

Switch board

Thermal module

Manifold

Cold plate

Universal Quick Disconnect

BMC

Chassis

Others

2,500

3,000

Content (S) Mix of total BoN

170,000

86.0%

5,760

4.608

1,152

200

64

300

200

700

300

10,000

6,000

2.9%

2.3%

0.6%

0.1%

0.0%

5.0%

3.3%

11.6%

5.0%

5.1%

3.0%

Switch tray BoM (S)

MVA/ Assemblyfees

Total compute tray BoM cost ($)

GB300 (10-9-8) Rack BoM cost analysis

Compute tray

Switch tray

Power supply system

Power shelf

Power whip

Bus bar

BMC

Thermal module (rack level)

Manifold

Chassis (rack level)

NVLink cables / connectors

Top of rack (ToR) front-end switch

4,000

50,000

MVA/ Assembly fees

018300 Rack otal BOM mthout MVA (USS)

GB300 Rack total BoM (USS)

4.000

50,000

95%

0.1%

1.2%

JPMe for ASP

mgmt server, cable cartrage, etc.

Source: Company data, J.P. Morgan estimates.

Unit

10

Note/Assumption

Each Bianca module has 2 G B200 Superchip

Each compute tray has 4 E1.S SSD drives, each SSD containing 3.84 TB storage, $0.3 per GB from Korea memory team

Each compute tray has 1 M.2 SSD boot drives, each having 1.92 TB storage, $0.3 per GB from Korea memory team

Each tray has 8-10 fans with each at $20 ASP.

DC-SCM and HMC board; two for two DPUs

Assuming adoption of Nvidia's DPU

Compute tray

VR200 Superchip - Strata board

SSD

Main SSD drives

Boot drives

BMC

Midplano

PDB

Busbar

HMC

125,000

ASP (S)

1,958

2,304

576

16

5,000

VR200 server BoM analysis

Networking cards

BlueField 4

CX9 1/0 board

Thermal module

Cold plate

Inner manifold

Universal Quick Disconnect

Chassis

Slide rail

Others

Compute tray BoM (S)

MVA/ Assemblyfees

Total compute tray BoM cost ($)

NVLink switch tray

NVLink Switch chip

CPU

Thermal module

Switch board

Cold plate

Manifold

Universal Quick Disconnect

BMC

Chassis

Others

Switch tray BoM (S)

6,800

4,000

6,000

250,000

Content (S)

9,792

9,216

576

48

5,000

300

200

700

300

34,000

Mix of total BoM

80.3%

3.1%

3.0%

0.2%

0.0%

1.6%

0.1%

0.1%

0.2%

0.1%

10.9%

MVA/ Assemblyfees

Total compute tray BoM cost (S)

Compute tray

Switch tray

Power supply system

110kw power shelt

Power whiD

Bus bar

BMC

Rack manifold

Chassis (rack level)

NVLink cables / connectors

Cable cartridge

53,996

1,699

24,000

97%

3.1% Assume 10% annual working capital fees for main chip, 2 month cycle time and 15% margin for other non-chip parts

0.4%

Top of rack (ToR) front-end switch

Others

MVA/ Assemblyfees

NVL72 Rack total BoM (US$)

4,000

25,000

4.000

25,000

0.1% Two ToR 400G 32 port switch, one mgmt 1G switch

Source: Company data, J.P. Morgan estimates.

1

0.4% mgmt server, fber patch panel, etc.

Unit

Note/Assumption

Each Strata module has 2 Rubin GPUs and 1 Vera CPU, 768GB LPDDR5 per board

Each compute tray has 4 E1.S SSD drives, 8 T per drive, $0.3 per GB from Korea memory tean

Each compute tray has 1 E1.S SSD boot drive, 2 T per drive, S0.3 per GB from Korea memory tean

DC-SCM and HMC board; one for DPU

11 Paladin connectors, M8+ PCB ~$1000, Connector content S500-1000, Amphenol margin

Assuming adoption of Nvidia's DPU

kunits

YoY

Top 4 US CSP general server board shipments

JPM bottom-up generic whitebox Ethernet ToR switch TAM

2022

7,167

2023

5,857

-18%

2024

6,144

5%

8,767

2025-

43%

2026E

12,274

40%

2027E

25%

18,411

2028E 25-28 CAGR

20%

Quantifying the TAM for generic whitebox Ethernet ToR switch (including AI Front-end)

Nhitebox ToR switch penetration rat

47%

46%

50%

60%

62%

64%

66%

•  Overall, we forecast a ~43% CAGR over 2025-28 for the general-purpose white-box Ethernet ToR switch revenue TAM, driven by (1) stronger general server demand, (2) rapid build-out of AI front-end network fabrics, and (3) ASP uplift from ongoing speed migration.

3%

0%

1%

0%

1 %

0%

1%

D%

JPM bottom-up generic whitebox Ethernet ToR switch TAM

0%

0%

25/5DGbps

20D/4DDGbps

20D/4DDGbp

Total

8DDGbps

1 Gbps

4DGbps

1 DG bps

25/5DG bps

20D/40DGbps

100Gbps

8DDGbps

Blended ASP per port ($)

1 DDGbps

8DDGbps

20D/40DGbps

Total TAM (US Smn)

3%

0%

4%

0%

8%

1%

YOY

Source:650 Group, Gartner®, J.P. Morgan estimates.

J.P. Morgan

0%

0%

2%

3%

15,342

28%

Al server shipment estimates (k racks)

GB200

Kyber NVL144

HGX

Hopper

Blackwell

Rubin

GB200

GB300

VR200

Kyber NVL144

HGX

Hopper

Blackwell

Rubin

GB200

GB300

VR200

Kyber NVL144

HGX

Hopper

Blackwell

Rubin

YoY

YoY

Nvidia's GPU ToR switch TAM estimates

2023

2024

30

2025

2026E

70

6

3

Nvidia's GPU ToR switch TAM estimates

2027E

10

3

60

38

2028E

44

Source: J.P. Morgan estimates, Company data

25-28 CAGR

J.P. Morgan

J.P. Morgan

400G switch

CPU

Fans

Unit price (S)

146

Switch BoM comparison - 400G vs. 800G bare-metal switch

% of BoM Note

2% Intel" Xeon® Processor D-1527

Volume

1

Total cost

146

What drives the ASP expansion of high-end datacenter switch?

PSU

Chassis

PCB

DRAM

300

300

•  We estimate a ~150% ASP increase for 800G switches versus 400G switches, driven by

NAND

Others

MVA

Subtotal

Total

•  2-3x higher Switch ASIC cost (Contributing ~70% of the total ASP increase)

20% Assume 20% GM for ODMI

•  PCB / CCL Materials upgrade to M8 from M6+, 3-4x CCL content growth
•  CPO adoption could also significantly increase the switch BoM and bodes well for ODM margins given the high complexity of component integration (FAU, OE, shuffle box, etc.). CPU Switch chip

PSU

Fans

Chassis

PCB

NAND

Others

DRAM

Subtotal

MVA

Total

1% 7+1, assume some spec upgrades for fans to dissipate more heats

Switch BoM comparison - 400G vs. 800G bare-metal switch

23

360

300

1500

184

300

720

1,500

Source: J.P. Morgan estimates, NAND/DRAM volume is on a per GB basis

0.3

Disclosures

Analyst Certification: The Research Analyst(s) denoted by an 'AC' on the cover of this report certifies (or, where multiple Research Analysts are primarily responsible for this report, the Research Analyst denoted by an 'AC' on the cover or within the document individually certifies, with respect to each security or issuer that the Research Analyst covers in this research) that: (1) all of the views expressed in this report accurately reflect the Research Analyst's personal views about any and all of the subject securities or issuers; and (2) no part of any of the Research Analyst's compensation was, is, or will be directly or indirectly related to the specific recommendations or views expressed by the Research Analyst(s) in this report. For all Korea-based Research Analysts listed on the front cover, if applicable, they also certify, as per KOFIA requirements, that the Research Analyst's analysis was made in good faith and that the views reflect the Research Analyst's own opinion, without undue influence or intervention.

All authors named within this report are Research Analysts who produce independent research unless otherwise specified. In Europe, Sector Specialists (Sales and Trading) may be shown on this report as contacts but are not authors of the report or part of the Research Department.