October9,2025 05:53 PM GMT

Investor Presentation | Asia Pacific

Greater China Semiconductors: Global Al Supply Chain Updates; Key Opportunities in Asia Semis

WeupgradedourIndustryViewtoAtractive going into2H25,andwereiterateourpreferencefor Alovernon-Al.Withtarifson semisandFXimpactconcernsbehindus,weexpect thegrouptore-ratefurther.Itisalsoagood timetopreview keyinvestment themesfor2026.Seeourreportfordetails:Greater China Semiconductors: Alsemi strength continues; upgrading Industry Viewto Attractive

Charlie Chan

Equity Analyst Charlie.Chan@morganstanley.com

+886 2 2730-1725

Daniel Yen, CFA

Equity Analyst Daniel.Yen@morganstanley.com

+886 2 2730-2863

MORGAN STANLEY ASIA LIMITED+

Daisy Dai, CFA Equity Analyst Daisy.Dai@morganstanley.com

+852 2848-7310

MORGAN STANLEY TAIWAN LIMITED+

Tiffany Yeh
Equity Analyst
Tifany.Yeh@morganstanley.com

+886 2 7712-3032

Lucas Wang
Research Associate
Lucas.Wang@morganstanley.com

+886 2 2730-2875

MORGAN STANLEY ASIA LIMITED+

Ethan Jia Research Associate Ethan.Jia@morganstanley.com

+852 3963-2287

GREATER CHINA TECHNOLOGY SEMICONDUCTORS

Asia Pacific Industry View

Attractive

Morgan Stanley does and seeks to do business with companies covered in Morgan Stanley Research.As a result, investors should be aware that the firm may have a conflict of interestthat could affectthe objectivity of Morgan Stanley Research.Investors should consider Morgan Stanley Research as only a single factor in making their investment decision.

For analyst certification and other important disclosures, refer to the Disclosure Section,located at the end of this report.

+= Analysts employed by non-U.S.afiliates are not registered with FlNRA,may not be associated persons of the member and may not be subject to FINRA restrictions on communications with a subject company, public appearances and trading securities held bya research analyst account.

· Top ideas:

OW: · Al: TSMC (Top Pick), Aspeed, Alchip, KYEC, ASE, FOCI, Himax, ASMPT, AlRing ·Memory (Al rippling effect): Winbond (Top Pick), GWC,Phison, Nanya Tech,APMemory, GigaDevice,Macronix · Non-Al: · Smartphone/Glasses: Novatek, OmniVision, Realtek · China WFE: NAURA Tech, AMEC, ACMR · Specialty: Silergy, SG Micro, Yangjie · Raw wafer: GlobalWafers

·EW/UW: MediaTek, UMC, ASMedia, Vanguard, WIN Semi, Parade ·Al cannibalization: 2H25 recoveryis gradual. Historicaly,the decline in semi inventorydays was a positive signal for semi stock price appreciation.

DeepSeek triggers inferencing Al demand,but are domestic GPUs suficient? DeepSeek demonstrating cheaper inferencing. However,shipments of Nvidia B30 could dilute the domestic GPU supply chain.

· China fab supply ramp: This has led to a prolonged downcycle of mature node foundry and niche memory.

· Long-term demand drivers:

·Tech diffusion: Alsemidemandreacceleration-thanks togenerativeAl-which proliferates todiferentverticalsoutsideof the semi industry.

· Tech deflation: We expect“price elasticity" to stimulate demand for tech products.

Valuation Comparison: Foundry, Back-end, Memory, IDM and Semi Cap

Ticker		Price Curr	Price	Upside/	Dividend Yield	FCF Yield	Market	3MAvg. Daily	P/E Ratio (x)			EPS Growth		ROAE			P/B Ratio (x)
	10/9/2025		Target				Cap (US$M)	Trading	2024					2024
		ency		Downside Rating	(%)	(%)			2025e	2026e	2024	2025e	2026e	2025e	2026e	2024	2025e	2026e
Foundry TSMC		TWD			1.5%					21.1	40%				30%		6.9
2330.TW 2303.TW			1,388.0	-4%		1.8%	1,178.953	1,245.1	31.8	24.6		29%	17%	30%	31%		8.7	5.7
0981.HK		TWD HKD	48.0	5%	5.7%	5.3%	18,198	50.6	12.1	17.3 12.8	-23%	-30%	35%	13%	9% 11%		1.5 1.5	1.4
5347.TWO			40.0	-52%	0.0%	-5.0%	69,306	817.3	NM	NM NM	-46%	105%	72%	2%	5% 8%		NM NM	NM
Vanguard 1347.HK		TWD	82.5	-19%	4.0%	3.1%	5,393	20.1	26.8	23.5 24.4	-14%	14%	-3%	12%	12% 11%		2.5 2.5	2.5
Hua Hong 6770.TW	85.2	HKD	38.0	-55%	1.5%	-4.6%	18.723	304.2	NM	NM NM	-82%	189%	111%	1%	3%	NM	NM	NM
PSMC 3105.TWO WIN Semi	30.3	TWD	22.0	-27%	0.0%	5.6%	3.992	28.8	NM	NM	224%	4%	-95%	-9%	5% -10% -1%	1.4	1.5	1.5
8086.TWO AWSC	101.5	TWD TWD	70.0 65.0	-31%	0.0%	-3.8%	2.264	13.6	NM 56.2 39.8	30.1	NM	41%	32%	2%	3% 4%	1.1	1.2	1.2
Mean:		115.0		-43%	0.9%	-3.9%	714	14.2	43.4 71.1	48.0	NM	-39%	48%	7%	4% 9%		2.9 2.9	2.8
Median:					1.7% 1.2%	-0.2% -1.0%		42.1 31.8	44.8	34.1	17% -19%	39%	27% 34%	7% 5%	7% 4%		3.0 2.8	2.5
									24.6	24.4		22%			7%		2.0 2.0	2.0
Memory
603986.SS Giga Device	225.5	CNY	234.0	4% 0	0.3%	4.1%	20.805	136.2	81.7	47.7	585%	67%	71%	7% 10%
2408.TW Nanya Tech	98.5	TWD	90.0	-9% 0	0.0%	13.6%	664.7 9.586 157.4	NM	380.1	6.8	-59%	-128%	NM	-2%	16% 0% 24%		9.0 8.2 1.8	7.2 1.5
2344.TW Winbond 8299.TWO Phison	43.5	TWD	42.0	-3% 0	6.3%	16.6%	6,174 93.4	NM	NM	25.3	-123%	-522%	NM	0% -1%	8%	2.1	1.9	1.9
SIMO.0 Silicon Motion	882.0 94.4	TWD	800.0 88.0	-9% 0	3.7%	1.6%	5,732 49.6 25.4	24.7	37.9	19.6	102%	-35% 94%		17% 10%	18%	3.7	2.0 3.6	3.2
2337.TW Macronix	27.5	USD TWD	29.0	-7% 0 0	3.7% 0.0%	2.1% 17.2%	3.192	31.4	24.4	18.8	64%	29%		16%	18%	4.1	3.7	3.3
Mean:				5%			1.610	16.9 NM	NM	307.5	90%	8% -105%		-9%	0%	1.2	1.3	1.5
Median:					2.3% 2.0%	9.2% 8.8%		64.1	131.0	71.0	110%	-97%		5%	14%	3.7	3.8	3.1
								31.4	59.8	22.4	77%	-13%	51%	4%	5% 17%	2.9		2.5
Backend 3711.TW ASEH 600584.SS JCET	179.0	TWD	188.0	5% 0	2.7%	6.0%	24,526	82.9 24.2	21.5

Source: Company data, Morgan Stanley Research (e) estimates; data as of October 9, 2025.

Valuation Comparison: Fabless, Power Semis, FPGA and Analog

Dividend FCF Market 3M Avg. P/E Ratio (x) EPS Growth ROAE P/B Ratio (x)DailyPrice Curr Price Upside/ Yield Yield Cap TradingTicker Company 10/9/2025 ency Target (%) (%) (US$M) (US$M) 2024 2025e 2026e 2024 2025e 2026e 2024 2025e 2026e 2024 2025e 2026eFabless

Smartphone/Wireless Semis
2454.TW MediaTek 1,345.0 TWD 1,388.0 3% E 5.4%						0.9% 66,474 302.9 20.1 20.8 18.2						38%		-3% 14%		27%	25% 27%		5.2 5.2 4.5
603501.SSOmniVision 150.0 CNY			170.0 13%	0	0.5%	2.1%	24,962	428.8 54.9		40.4	29.7	485%		36%	36%	15%	17% 20%		7.4 6.4 5.4
300782.SZMaxscend 84.3 CNY			55.00 -35%	U	0.4%	-2.2%	6,250	114.2 NM		NM	52.9	-64%		-18%	160%	4%	3%	8%	4.4 4.3 4.0
603160.SSGoodix 85.1 CNY			65.0 -24%	U	0.5%	-1.4%	5.418 71.2 64.6			66.2	59.4 266%			-2%	11%	7%	7%	7%	4.5 4.3 4.1
6462.TWOEgis 123.0 TWD			130.0 6%	E 0.7%		-16.3%	354	14.5 NM		69.9	15.7	28%		-114%	346%	-11%	2%	6%	1.1 1.1 1.0
IP/Design Service
3661.TW Alchip 3,515.0		TWD	4,288.0 22%	0 0.7%		5.1% 9,063		274.7 44.5		48.0	24.5	83%		-7% 96%		22%	15% 26%		NM NM NM
301269.SZEmpyrean Technology 131.8		CNY	115.0 -13%	E 0.1%		NA 9.943		114.0 NM NM			NM	-45%		184% 102%		2% 6% 11%			NM NM NM
3443.TW GUC 1,515.0		TWD	1,480.0 -2%	0 0.0%		4.7% 6,411		92.1 59.1 52.5			41.3	-2%		13% 27%		33%	32%	34%	NM NM NM
6531.TW APMemory 392.5 TWD			415.0 6%	0 2.1%		4.1% 2.006 33.8 40.3 NM 29.5						9%		-75% 453%		16%	4%	20%	6.1 6.5 5.5
6533.TW Andes 295.0 TWD			608.0 106%	0 0.0%		0.3% 472 8.6 NM				NM NM -102%				-5828%	-552%	0%	-2% 10%		3.1 3.2 2.8
PC/Cloud/Display Semis
3034.TW Novatek 424.0		TWD 550.0 30%		0	5.6%	14.3% 8,147		42.0 12.7		16.8	15.5	-13%		-24%	8%	30%	23%	24%	3.8 3.9 3.7
2379.TW Realtek 555.0		TWD	650.0 17%	0	3.1%	5.0%	8.988	40.7 18.6		20.5	17.2	67%		-9%	19%	32%	25%	27%	5.4 4.9 4.3
688008.SSMontage 157.1 CNY			++#VALUE!	+	0.4%	-0.7%	24,909	538.4 126.2		NM	NM	214%		NA	NA	13%	++	++	15.7 NA NA
5269.TW ASMedia 1,460.0 TWD			1,600.0 10%	E	1.0%	0.9%	3,440 47.2 28.4			23.0	19.5	60%		24% 18%		14%	14%	14%	3.3 3.0 2.7
5274.TWOAspeed 5,690.0 TWD			6,100.0 7%	0 1.7%		1.3% 6,790 73.0 83.6 60.8 47.9						155%		37% 27%		54%	56% 58%		NM NM NM
4966.TWO	Parade 781.0	TWD	747.0 -4%	E	2.1%	1.6%	1,955	38.1 24.0		22.0	18.5	27%		9% 19%		12%	12% 13%		2.8 2.6 2.4
2458.TW	Elan Micro 132.5	TWD	120.0 -9%	E	5.1%	9.6%	1,198	8.9	14.0	17.4	15.6	27%		-20%	12%	29%	22%	24%	3.9 3.9 3.6
HIMX.O	Himax 9.1	TWD	8.8 -3%	0	5.7%	8.3%	1,554	9.6	19.9	21.8	19.5	58%		-9%	12%	9%	9%	10%	1.8 1.8 1.8
Analog IC
6415.TW	Silergy 249.0 TWD		378.0 52%	0 0.2%		1.3%	3,035	23.0 42.2		28.3 17.3		212%		49% 64%		7%	10% 14%		2.8 2.6 2.3
300661.SZSG Micro 85.2		CNY	90.0 6%	0	0.4%	1.1% 7,283		210.7 NM		82.6	50.2	80%		-1% 65%		12%	12% 19%		NM NM NM
MCU
4919.TW Nuvoton 62.8		TWD	60.0 -4%	E	3.4%	14.5%	832	5.7 NM		NM	20.4	-91%		-94%	9418%	1%	0% 8%		1.7 1.7 1.6
688018.SS	Espressif 216.2	CNY	211.0 -2%	0	0.8%	20.3%	4.623	120.3 NM		70.8	49.4	149%		41%	43%	17%	21% 26%		NM NM NM
300327.SZ	Sino Wealth 27.4	CNY	17.7 -35%	U	0.8%	6.6%	1,293	46.1 69.8		49.3	32.9		-28%	41%	50%	7%	10% 14%		5.2 4.8 4.4
FPGA
002049.SZ	Unigroup Guoxin 92.8	CNY	50.0 -46%	U	0.3%	1.5% 10.951		286.6 66.8		42.7	33.7		-53%	57% 27%		10%	14% 15%		6.4 5.6 4.9
1385.HK	Shanghai Fudan 45.1	HKD	32.8 -27%	0	0.6%	0.1% 6.905		37.8 64.4		44.4	33.1	-20%		45%	34%	10%	13% 15%		6.3 5.5 4.7
688107.SS	32.3 CNY		32.4 0%	E	0.0%	-1.0% 1,800 25.8 NM				NM	NM 4%			-66%	-133%	-15%	-6% 2%		NM NM NM
Mean: 1.6%Median: 0.7%
Mean: 1.6%Median: 0.7%						3.2%1.4% 44.5				43.5 29.5 27%				-219%-2% 31%		12% 13% 15%			4.4 4.3 4.0

P/E multiple - NVIDlA re-rated in 2023

P/E multiple - ASlC group once exceeded NVIDlA, and then de-rated with NVIDIA

Memory Stock Prices Are a Leading Indicator for Logic Semis; We Have an Attractive Industry View on GC Semis

Memory share price YoY peak leads logic semis

Share price performance over the past three years (TSMC vs. SK hynix)

TSMC: One-year forward P/E chart

We estimate Al semis to account for \~34% of TSMC's 2027e revenue

Al semi is the major growth driver

Another US\$3-4 trillion Al capex in the rest of this decade

Global Semiconductor Revenue: Alvs.Non-Al

Logic semi foundry utilization at 70-80% in 1H25 - Still not fully recovered

Semi supply chain days of inventory declined in 2Q25
Source: SIA, FactSet, company data, Morgan Stanley Research (e) estimates.

Excluding NVIDIA's Al GPU revenue, non-Al semi growth was slow at only 10% Y/Y in 2024

Historically, when inventory days went down, semi stock index went up

DDR4 Shortage Persisting into 2H26

Quarterly supply breakdown (mn Gb)
Source: Company data, Morgan Stanley Research estimate

Quarterly demand breakdown by product (mn Gb)

Quarterly oversupply/undersupply ratio vs. Nanya and Winbond pricing Q/Q change

NOR flash demand and supply growth rates
NAND wafer spot pricing and module prices

NOR flash demand and supply growth rates

Source:ChinaFlashMarket,Morgan Stanley Research estimate

Customer 12-inch wafer inventory (logic vs. memory; August 7, 2025)
Source: Sumco, Morgan Stanley Research

China Al and Domestic GPU Supply

China DeepSeek: Inference Demand Is the Key Driver of Future Capex

We forecast the top 6 companies' capex to grow 62%YoY to Rmb373bn

SMIC 7nm and 10nm/14nm revenue mix

China GPU performance comparison table


Brand	Huawei		Cambricon		Hygon	Iluvatar	Enflame	Alibaba
Product	Ascend 910B	Ascend 910C	MLU370	MLU590	Deep Compute IIl	Big Island	120	pingtouge
Node	SMIC 7nm (n+2)	SMIC 7nm (n+2)	7nm	7nm	7nm	TSMC 7nm	TSMC 7nm	12nm
FP16 (TFLOPS)	320	800	72	256	192	147	128	205
FP8/Int8 (TFLOPS)	640	1,600	192	512	392	294	256	825
HBM usage	HBM2e	HBM2e	HBM2	HBM2e	HBM2e	HBM2	HBM2	SRAM
HBM memory (GB)	64	NA	24	NA	NA	32	32	NA
HBM bandwidth (GB/s)	1,600	3,200	NA	NA	NA	1,200	300	NA
GPU-GPU networking bandwidth	PCle Gen5	PCle Gen5	PCle Gen4: 200GB/s	PCle Gen5	PCle Gen5	PCIE Gen4	64GB/s	NA
TDP (W)	400W	NA	70/150W	NA	350	300W	300W	NA

Source: Tom's Hardware, FactSet, company data, Morgan Stanley Research (e) estimates.

Comparison between NVL72 and CloudMatrix 384
Huawei CloudMatrix 384 A3 SuperPod

	Nvidia	Huawei
	NVL72	CloudMatrix 384
Al processor	GB200	Ascend 910C
CPU	Nvidia Grace	Huawei Kunpeng 920
FP16/BF16 (PFLOPS)	180	215-307
FP8/Int8 (PFLOPS)	360	430-614
FP4 (TFLOPS)	720	NA
Total memory size (TB)	30	49
Total memory bandwidth (TB/s)	576	1229
Interconnection	NVLink Gen5: 1.8TB/s	UBLink: 784GB/s
Launch date	March 2024	April 2025

Racks of CloudMatrix 384

Source: Company data, Morgan Stanley Research

We expect China cloud Al TAM to be US\$48bn in 2027

We expect China local GPU can almost fulfill China Al demand by 2027e
Source: Gartner, Morgan Stanley Research (e) estimates.

Local GPU revenue could grow to Rmb136bn in 2027e,enabled by SMiC's leading node capacity

	2024	2025e	2026e	2027e
SMIC capacity for Local GPU(kwpm)		7	13	18
910B capacity (kwpm)		3	0	0
910C capacity (kwpm)		1	6	4
910X capacity(kwpm)	2200	0	2	7
Other GPU capacity		3	5	7
Die per wafer 910B Die per wafer 910C	78	78	78	78
	39	39	39	39
Die per wafer 910X	39	39	39	39

Average yield rateof910B (%)	30%	30%	50%	70%
Average yield rate of 910C (%)	0%	5%	30%	50%
Averageyield rate of 910x (%)	0%	0%	15%	30%
Wafer inventory from previous year (k)	150
910B Die annual production (k)	562	842	0
910C Die annual production (k)		23	842	0
910X Die annual production (k)	0 0	0	140	936 983

910B price (Rmb th)	50	50	0	0
910C price (Rmb th)	0	110	70	50
910X price(Rmbth)	0	0	140	70
Revenue from 910B(Rmbmn)	35,580	42,120	0
Revenue from910c(Rmb mn)	0	2,574	58,968	0 46,800
Revenue from910x(Rmbmn)	0
		0	19,656	68,796
Other GPU revenue (Rmbmn)	7,367	13,763	15,463	20,486
Total GPU revenue (Rmb mn)	42,947	58,457	94,087	136,082
YY (%)		36%	61%	45%

=oundry Process Roadmap and Performance Gap: TSMC vs. Intel vs. Samsung vs. SMIC

Logic density comparison: TSMC vs. Intel

Logic density comparison: TSMC vs. Samsung

Foundry process road map
Source: Company data, Morgan Stanley Research estimates.

RTX Pro 6000 series products allhave better FP8 performance than H20...

...while H20 possesses greater memory bandwidth
Specs of NViDlA's chip available to China

Nvidia
Product	L40S	RTX 6000 Ada	H20	RTX Pro 6000	A100	H100
Node	TSMC N4	TSMC N4	TSMC N4	TSMC N4	TSMC 7nm	TSMC N4
FP16 (TFLOPS)	362	364	148	250	312	989
FP8/Int8 (TFLOPS)	733	729	296	500	624	1,979
FP4 (TFLOPS)	NA	NA	NA	1,001	NA	NA
DRAM interface	GDDR6	GDDR6	HBM3	GDDR7	HBM2e	HBM3
Memory size (GB)	48	24	96	96	80	80
Memory bandwidth (GB/s)	864	672	4,000	1,792	1,935	3,350
GPU-GPU networking bandwidth	PCle Gen4: 64GB/s	NVLink: 100GB/s	NVLink: 900GB/s	PCle Gen5: 128GB/s	NVLink: 600GB/s	NVLink: 600GB/s
TDP (W)	350w	300W	400W	600W	300-400W	700W
Launch date	Aug 2023	Sep 2022	Q1 2024	Mar 2025	May 2020	Apr 2022

Source: Company data, Morgan Stanley Research

RTX Pro 6000 Blackwell GPU at Computex

ConnectX-8 with Built-in PCle Gen6 Switch

	Nvidia				Huawei	Huawei
Product	L40S	RTX 6000 Ada	H20	RTX Pro 6000	Ascend 910B	Ascend 910C
Node	TSMC N4	TSMC N4	TSMC N4	TSMCN4	SMIC 7nm (n+2)	SMIC 7nm (n+2)
FP16 (TFLOPS)	362	364	148	250	320	800
FP8/Int8 (TFLOPS)	733	729	296	500	640	1,600
FP4 (TFLOPS)	NA	NA	NA	1,001	NA	NA
DRAM interface	GDDR6	GDDR6	HBM3	GDDR7	HBM2e	HBM2e
Memory size (GB)	48	24	96	96	64	128
Memory bandwidth (GB/s)	864	672	4,000	1,792	1,600	3,200
GPU-GPU networking bandwidth	PCle Gen4: 64GB/s	NVLink: 100GB/s	NVLink: 900GB/s	PCle Gen5: 128GB/s	PCle Gen5	PCle Gen5
TDP (W)	350w	300W	400W	600W	400W	310W
Launch date	Aug 2023	Sep 2022	Q1 2024	Mar 2025	Q3 2023	Q2 2025

Source:Foxconn, Morgan Stanley Research

China semi equipment imports growth rebounded to +10% yoy (3MMA) in July-25

Semi Lithography equipment imports from Netherlands see slight uptick in July-25
Source: China's General Administration of Customs, Morgan Stanley Research.

Semi equipment imports from most major countries declined in Jan-Jul 25, except for Singapore

16x Trainium per hour

Source: Company data for US east 2b, 2c,and US central 1, Morgan Stanley Research.

H100 per GPU per hour

12x Inferentia per hour

TSMC CoWoS and CPO

Al Semi Supply Leading Indicator

· Demand trigger: Generative Al ·Growth limitations:

Budget
Energy -> in the US
3.Chip Capacity -> in China
4.Regulation

Semiconductor solutions: · Moore's Law · CoWoS/SolC HBM · CPO Custom chip GaN HVDC 800V · Growth perspective:

1． Inference vs. Training Al semis
2． Edge vs. Cloud Al semis
3． Custom ASIC vs. Al GPU

·Growth perspective:

1． Edge Al semis: 22% CAGR, 2023-30
2. Inference Al semis: 55% CAGR, 2023-30
3. Custom Al semis: 39% CAGR, 2023-30

Edge Al semis could grow slightly faster than cloud

Total Al semi revenuebreakdown: Cloud vs. Edge
... and Custom Al chips to outgrow general purpose

Cloud semi breakdown:Al Training vs.Inference

Within Cloud Al, Inference Al chips to outgrow training..
Source: Company data, Gartner, Morgan Stanley Research (e) estimates.

Cloud semi breakdown:General Al vs. Custom Al

Morgan Stanley cloud capex tracker estimates nearly US\$582bn of cloud capex spending in 2026 (Purely Top 10 listed global CSPs; no sovereign Al)

Nvidia CEO shared the estimate of global cloud capex to reach US\$1 trilion in 2028 (including sovereign Al)

Our supply chain data-driven bull case assumption cloud Al Semi TAM may grow to US\$235bn in 2025e

Al semi is the major growth driver

Global Semiconductor Revenue: Al vs.Non-Al

Capex to EBlTDA ratio has likely surged since 2024

Capex intensity is also set to increase rapidly

Monthly tokens processed

Datacenter/HPC semi revenue from NVIDIA and AMD

NVIDIA GB200/300 rack supply share (2025)
GB200/30o NVL72-equivalent rack supply share

NVIDIA GB200/300 rack demand share (2025)
NvidiaGB200 serverdemand share,2025e

NVIDIA GB200/300 Rack Output Estimates

GB200/300 NVL72 Rack Ramp Quarterly Trajectory

Industry-wide GB200 NVL72 Rack Monthly Output
TSMC expected to produce 5.1mn chips in 2025, with full-year GB200 NVL72 shipments expected to reach 30l

Source: Morgan Stanley Research. te:i accounting forrack assembly andtestimesforL11).Estimates fromAsia Pacific tech team'sdownstream estimates.

Global CoWoS capacity expansion by year-end and by vendor

Global CoWoS consumption, by customer

Key changes for 2026 CoWoS allocation in a chart

Key changes on 2026 CoWoS allocation in a table (see numbers marked in red)

(kwafer)	2023	2024	2025e	2026e (new)	2023	2024	2025e	2026e
NVIDIA	53	200	425	685	45%	54%	63%	59%
TSMC			390	605
CoWoS-L			390	590
CoWoS-S			0	15
CoWoS-R			0	0
Non-TSMC			35	80
Amkor			35	60
CoWoS-S			20	0
CoWoS-R		1	15	60
ASE/SPIL			0	20
CoWoS-S			0	0
CoWoS-R			0	20
Broadcom	23	68	85	210	20%	18%	13%	18%
TSMC			83	205
CoWoS-L			0	50
CoWoS-S		1	83	155
ASE/SPIL			2	5
CoWoS-L			0	0
CoWoS-S			2	5	6%	11%
AMD	7	40	60	105			9%	9%
TSMC			60	80
CoWoS-L			0	70
CoWoS-S		F	60	10
ASE/SPIL			0	25
CoWoS-L			0	25
CoWoS-S MediaTek			0	0				1%
AWS +Alchip	9			10	8%	4%	1%	6%
Intel Habana	0	16	5 9	70 0	0%	2%	1%	0%
Marvell		18			1%	5%	11%	4%
	1		75	45	1%	0%	0%	1%
GUC	1	1	2	14	1 17%	3%	3%	1%
Others	20	10	19	15	100%	100%	100%	100%
Total demand	117	370	680	1,154

Source: Company data, Morgan Stanley Research (e) estimates.

Al semi wafer consumption, by customer, 2025e

Al semi wafer consumption, by customer, 2026e

Al chip vendor	Product name	CoWarsacity capacity allocation	Chips per Chips per CoWoS wafer shipments (k)	Implied Implied	Compute die Compute die size	Compute die Compute die Geometry		Warer Watferpriew consumption	Warer Wafer price (uss)	Wafer revenue TAM (Ussmn)
Al chip vendor	Product name	CoWarsacity capacity allocation	Chips per Chips per CoWoS wafer shipments (k)	Implied Implied	Compute die Compute die size		units	Warer Watferpriew consumption	Warer Wafer price (uss)	Wafer revenue TAM (Ussmn)
AI GPU (2026e)		(k wafers)						(k wafers)
NVIDIA	B300	235	14	3,290	850	4nm	2	261	21,945	5,730
NVIDIA	Rubin R200	355	8	2,840	850	3nm	2	225	24,300	5,477
AMD	MI355	10	12	120	110	3nm	8	3	24,300	85
AMD	MI400	80	10	800	110	2nm	8	39	28,125	1,091
AI ASIC (2026e)
Google	TPU v6 (Trillium)	20	20 20	400	600	5nm	2	18	19,609	352
Google	TPU v7 (Ironwood)	120		2,400	700	3nm	2	157	24,300	3,812
AWS	TPU v8 (Zebrafish)	10	20	200	800	3nm	2	15	24,300	363
AWS	Trainium 2.5	20	16	320	600	5nm	2	9	20,543	187
xAI	Trainium 3	70	17 20	1,190 200	700	3nm	2	41	26,089	1,061
Microsoft		10			645	4nm	1	5	21,945	107
Microsoft	Maia 200	4	29	116	700	3nm	1	3.0	24,300	74
	Maia 300	25	11 10	275 500	850	2nm	1	14.6 39.7	28,125 24,300	409
Meta	MTIA 3 (Iris)	50	13	130	850 750	3nm 3nm	2 4	18.2	28,125	964 512
Open Al	Nexus (Titan 1)	10						849
Total		1,109								20,224

Source: Company data, Morgan Stanley Research estimates. Note: Estimates are compiled using our supply chain checks.

HBM consumption by customer in 2026e

NVIDIA still consumes most of the HBM supply in 2026e

Al chip vendor	Product name	capacity allocation (k wafers)	Chips per CoWoS wafer shipments (k)	Implied	HBM chip density (GB)	HBM chip Total HBM size units	(GB)	HBM generation	HBMvendor	Total HBM demand (k GB)
AI GPU (2026e)
NVIDIA	B300	235	14	3,290	36	8	288		HBM3e 12hiHynix/Micron/Samsung	947,520
NVIDIA	Rubin R200	355	8	2,840	36	8	288		HBM4 Hynix/Micron/Samsung?	817,920
AMD	MI355	10	12	120	36	8	288	HBM3e 12hi	Samsung	34,560
AMD	MI400	80	10	800	36	12	432	HBM4	Samsung	345,600
AI ASIC (2026e)
AI ASIC (2026e)	TPU v6 (Trillium)	20	20	400	24	4	96	HBM3e 8hi	Samsung/Micron	38,400
Google	TPUv7(Ironwood)	120	20	2,400 200	24	8	192	HBM3e 8hi	Hynix/Samsung	460,800
Google	TPUv8 (Zebrafish)	10	20		36	6	216	HBM3e 12hi	Hynix/Micron	43,200
AWS	Trainium 2.5	20	16	320	24	4	96	HBM3e 8hi	Hynix/Samsung/Micron	30,720
AWS	Trainium 3	70	17	1,190	36	8	288	HBM3e 12hi	Hynix/Samsung/Micron	342,720
xAI		10	20	200	24	6	144	HBM3e 8hi	Samsung?	1,152
xAI	Maia 200	4	29	116	16	4	64	HBM3	Samsung	7,424
Microsoft	Maia 300	25	11	275	36	8	288	HBM4	Samsung?	79,200
Meta	MTIA 3 (Iris)	50	10	500	36	6	216	HBM3e 12hi	Hynix/Samsung	108,000
Open Al Total	Nexus (Titan 1)	10 1,109	13	130	36	6	216	HBM3e 12hi	Hynix/Samsung	28,080 3,285,296

Source: Company data, Morgan Stanley Research (e) estimates

NVIDIA testing demand - Revenue contribution to KYEC

1Q24 2Q24 3Q24 4Q24 1Q25 2Q25 3Q25 4Q25 2024 2025 2026
Units (k)
H-series	800	1,000	1,400	1,200	650	200	250		4,400	1,100
B-series with GDDR7	-	-		：	：	：	600	1,400	-	2,000	5,000
B-series with HBM	-	-	-	400	1,000	1,100	1,300	1,600	400	5,000	3,000
R-series											2,400

Testtime(assumeHopper testingtimefor~6mins)
H-series (k mins)	4,667	5,833	8,167	7,000	3,792	1,167	1,458		25,667	6,417	-
H-series testtime(seconds)	350	350	350	350	350	350	350	350
B-series with GDDR7(k mins)							3,500	8,167	-	11,667	29,167
B-series with GDDR7 test time (seconds)							350	350
B-series with HBM(k mins)		-		4,667	11,667	13,750	19,500	26,667	4,667	71,583	50,000
B-series with HBMtest time (seconds)				700	700	750	900	1,000
R-series (k mins)											56,000
R-series test time (seconds)											1,400

Testhourrate(forthelargestcustomerincludingburn-in)	180	180	180	190	190	180	180	180

Largest customer revenue to KYEC(US$mn)	14.0	17.5	24.5	36.9	49.0	44.8	73.4	104.5	92.9	271.6	406.9
%of KYECMSe Revenue	7%	8%	11%	16%	21%	16%	24%	31%	11%	24%	27%

KYEC revenue may increase significantly in 2025/26e
Source: Company data, FactSet, Morgan Stanley Research (e) estimates.

KYEC: 12-month forward P/E

CPO Could Help Enhance Data Transmission Speeds and Reduce Power Consumption

CPO is the most power saving solution

Source: Broadcom,ASE, Morgan Stanley Research.

Design Service: Custom Chip Projects Mapping Table

	Alchip Inferentia2/Tranium1	Global Unichip	Broadcomm	Marvell Trainium2/Trainium2.5	TSMCdirect Graviton1(16nm)	Samsung direct	SMICdirect	MediaTek	Socionext Inteldirect Al fabric chips (18A)	VeriSilicon
AWS-Annaumna	（7nm） Trainium3(3nm）			（5nm)	Graviton2(7mm) Graviton 3				Xeon6(3nm)
AWS-Annaumna	Kindleprocessor Amazon-Lab126	Echo processor(5nm)			NitroNetworking Chip
Google	Echo processor(12nm)	Security Chip Axion2 CPU (3nm) Axion3CPU(IP Onliy)	TPUv1/TPUV2 TPUV3/TPUV4 TPUV5/TPUv6e (Trillim）	MapleCPU (5nm) Axion3CPU (?nm)	Axion1CPU (5nm) Tensor (3nm)	Tensor (7nm) Tensor (5nm)		TPUv3
Microsoft		Maia 100(5nm) Cobalt 100(5mm)	TPU v7p(lronwood;3nm) TPUv7e(3m）	Maia 200 enhancement				TPU v8p (3nm) ro TPUvse (3nm)
Meta	Oculus ASIC	Maia 200(3m)	MTIAv1	（3nm） Switch Networking			Oculus ASIC
Meta	MTIAv4?	MTIAv4(CoWoS) MTIAv5(SOW)	MTIAV2 MTIAVS MTIAv3.5	AINetworking MTIAv4?			MTIA v4?	CPU (3nm)？
Apple			MTIAv? CustomNetworking Chip
Sony	TVASIC SmartphoneISP	DSCASIC						PS5 South Bridge
Tesla/xAI	D1/Dojo(7m)	Alaccelerator a1t g			D1/Dojo(7nm) D2/Dojo(5m)	Autopilot 3.0/FSD1 (14nm)
Li Auto					AI5(3nm)	Autopilot4.0/FSD2 Ag
GM-Cruise	ADAS High-end (5nm) ADAS Low-end (5nm)									ADAS
Intel-Habana	Gaudi1(16nm) Gaudi2/Goya 2（7nm)	Goya1(16nm)						5nm
OpenAI	Gaudi3/Goya3（5nm）
Baidu			AI Custom Chip
Alibaba-T-head					Kunlun2（7nm)	Kunlun1(14nm) Kunlun3（5nm)
Tencent		Generation1			Hanguan Zixiao(12m)
ByteDance			Al accelerator(7nm)					Al accelerator(3nm)		Al accelerator
			5nm?						x86CPU
Nvidia								G810

CSPs Still Need Custom Chips Even With NVIDIA Providing Powerful Al GPUs

Global CSP cloud Al custom chips

Latest evidence of Al ASlC: AWS Trainium3

Trainium Forecast from MS Asia Research team

k Units	2023	2024	2025	2026	2027	2028
Trainium1/Inferentia2	300	600	30
Trainium2		300	1,200	320
Trainium3			20	1,190	1,500	250
Trainium4					50	1,300
Total	300	900	1,250	1,510	1,550	1,550

TPU Forecast from MS Asia Research team

k Units	2023	2024	2025	2026	2027	2028
v5	500	2,400	250
v6 (Trlium)			1,500	300
v7 (Ironwood)			-	2,500	1,200
v8 (N; 3nm)				200	1,000	200
N+1 (2nm)					800	2.200
N+2 (2nm)						800
Total	500	2,400	1,750	3,000	3,000	3,200

Google TPU keeping pace with Nvidia GPU

Source: Company data, Morgan Stanley Research.

ASIC volume forecast by major companies

Total cost of ownership to train Llama-3 40oB model

Total cost of ownership to inference

What's Next?

Inference and Edge Al

The five key edge Al hurdles Battery & Power Challenge: Edge devices are power constrained Consumption Enablers:Provide energy eficient compute Processing Power & Challenge:Edgedevices lack cloud-level compute&memory Memory Enablers:Expandon-devicecompute&memorycapability Cost hal Form Factor hllEe Application Use Challenge: Edge Al "Killer Apps" are yet to be seen Cases Enablers:Develop/leverage toolkitstoexpandapps

2.5D CoWoS solution

Wafer on Wafer solutions; DRAM on top of the SOC die
Source: Company data, Winbond, Morgan Stanley Research

Benefits from WoW Stacking

WoW technology could be a key solution for bandwidth and power improvement

WoW could potentially reduce the BOM cost requirement for Al functions
Source: Company data, Morgan Stanley Research

Ne Estimate a US\$6bn TAM by 2030, Substantial for Niche Memory Players

In our base case, we assume 257% market CAGR, 2025-30e, for WoW stacking products

Top 4 CSPs (Amazon, Google, Microsoft and Meta) capex was up 67% Y/Y in 2Q25CY
Source: Company data, Morgan Stanley Research (e) estimates.

Capex to EBITDA ratio should remain stable at around 50%

Aspeed earnings revision breadth

Global auto semis are dominated by EU, Japan and US vendors; Greater China's supplies in aggregate account for less than 5%...

... while China EV penetration is ahead of the global average

Self-sufficiency rate is uneven among different segments; we see opportunities in power discrete, ADAS and MCU

We expect China auto semis to achieve 28% selfsufficiency in 2027

OSAT capex as % of sales back to historical high level driven by advanced packaging

Future GW-scale data center requires higher reliability power discrete

JCET has highest overseas revenue exposure among China OSAT Overseas revenue exposure in 2024
Source: SIA, Company data, Morgan Stanley Research

JCET local revenue growth lower vs. its peers in past five years

SiC: 6/8-inch Pricing Decline Further, While AR Glass and Al Data Centers To Be Future Driver

6/8-inch SiC substrate price trend

We expect SiC penetration in EVs to grow to 45% by 2027
SiC used in 800V Data Centers

Meta Orion, utilizes a MicroLED+SiC waveguide display solution

1st Stage (Grid- 2nd Stage (800V- 3nd Stage (48V
800V) 48V) Al) ·UHV SiC ·HV GaN & SiC ·80-200V GaN ·80-200V GaN

Important note regarding economic sanctions.This research references U.S.Executive Order 14032 and/or entities or securities that are designated thereunder.Executive Order 14O32 may prohibit U.S.persons from buying certain securitiesofentities namedinthis note/presentation.Thedatasetforth inthis researchisfor informationalpurposes and does not represent Morgan Stanley's view as to whether ornot anyof theinstruments discusedinthis note are subject to sanctions.Anyreferencesinthis reporttoentities,debt orequityinstruments that maybecoveredbysuch sanctions shouldnot be readas recommending or advising as toanyinvestment activitiesin relation to such entities or instruments.Users ofthis report are solely responsiblefor ensuring thattheir investment activities in relation to any sanctioned entities and/or securities are carried out in compliance with applicable sanctions.

Important noteregarding export controls.This note references export controls maintained bythe U.S.Departmentof Commerce Bureau of Industry and Security (BIS)and/or entities that may be subject to export control restrictions. These controls may restrict export,re-export or transfer (in-country)of certain items covered by the Export Administration Regulations (EAR).This disclaimerdoes not represent Morgan Stanley's view as towhetheror notthe export,re-export or transfer (in-country)of any items is subject to export controls laws.Investors are solely responsibleforensuring thattheir investment activities in relation toanyentities arecarriedoutincompliance with applicable export controls laws.

Disclosure Section

are made by the duly qualified research analysts.

tative or Morgan StanleyResearchat1585Broadway, (Attention:Researchanagement),NewYork,NY,36UA

Research Management), New York, NY 1OO36 USA.

Analyst Certification

specific recommendations or views in this report: Charlie Chan; Daisy Dai, CFA; Tiffany Yeh; Daniel Yen,CFA.

Global Research Conflict Management Policy

tanley.com.br

Important Regulatory Disclosures on Subject Companies

Technolgtatf)d

WPG Holdings.

Corp.

i trading desks.

inthiseportadealibtecriadivitoftb Certain disclosures listed above are also for compliance with applicable regulations in non-US jurisdictions.

STOCK RATINGS

buyorsellastock shoulddependonindividualcircumstances (suchas theivestor's existing holdings)andotherconsiderations.

Global Stock Ratings Distribution

(as of September 30,2025)

eStockRatingsdescribedbelowaplytoorganStanleysFundamentalEquityReseachanddonotapplytoDebtResearchproducedbythFir.

veeh

Coverage Universe			Investment Banking Clients (IBC)			Other Material Investment Services Clients (MISC)
Stock Rating Category	Count	%of Total	Count	% of Total IBC	% of Rating Category	Count	% of Total Other MISC
Overweight/Buy	1499	41%	389	46%	26%	702	41%
Equal-weight/Hold	1618	44%	375	44%	23%	782	45%
Not-Rated/Hold	4	0%	1	0%	25%	1	0%
Underweight/Sell	577	16%	88	10%	15%	234	14%
Total	3,698		853			1719

of total"column may not add up to exactly1OO percent.

Analyst Stock Ratings

e Unless otherwisespecified,the timeframe for price targets included in Morgan Stanley Research is 12to18 months.

Analyst Industry Views

t t AC Asia Pacific ex Japan Index.

Important Disclosures for Morgan Stanley Smith Barney LLC Customers

conflict of interest.

Other Important Disclosures

s Management determine that a diferent publication schedule is appropriate based on current conditions.

d

securities/instruments.

narkets transactions performed by Morgan Stanley or the profitability or revenues of particular trading desks.

those discussed in Morgan Stanley Research.Derivatives may be issued by Morgan Stanley or associated persons.

beenrevieetosilai

/make investment decisions that are inconsistent with the recommendations or views in this report.

et

s directed at Sophisticated investors only.

from your sales representative.

and is not intended for Retail Customers as defined by the QFCRA.

solely to this information stated here may not bring about outcomes that fit your expectations.

ote

continuing analysis or additional reports relating to such issuersand/or individualfixed income products.

Registatt ofreturnstoinvestorsvestmentinsecuritiesmarketaresbcttomarketrissReadalteeateddocumentscarefullyeoreivesting

INDUSTRY COVERAGE: Greater China Technology Semiconductors

Charlie Chan
ACM Research Inc (ACMR.O)	0 (03/07/2023) US$41.99
Advanced Micro-Fabrication Equipment Inc (688012.SS)	0 (11/06/2023) Rmb319.11
Advanced Wireless Semiconductor Co (8086.TWO)	U (07/14/2025) NT$115.00
Alchip Technologies Ltd (3661.TW)	0 (05/14/2021) NTS3,515.00
Andes Technology Corp (6533.TW)	0 (08/04/2022) NT$295.00
ASE Technology Holding Co.Ltd. (3711.TW)	0 (09/15/2024) NTS179.00
Global Unichip Corp (3443.TW)
GlobalWafers Co Ltd (6488.TWO)	0 (07/27/2024) NTS1,515.00
Gudeng Precision (3680.TWO)	0 (09/19/2025) NTS495.50
Hua Hong Semiconductor Ltd (1347.HK)	E (07/16/2025) NTS339.00
King Yuan Electronics Co Ltd (2449.TW)	E (05/09/2025) HK$85.20
Maxscend Microelectronics Co Ltd (300782.SZ)	0 (03/03/2023) NT$176.00
MediaTek (2454.TW)	U (01/11/2021) Rmb84.28
Nanya Technology Corp. (2408.TW)	E(09/23/2025) NT$1,345.00
NAURA Technology Group Co Ltd (002371.SZ)	0 (09/18/2025) NTS98.50
OmniVision Integrated Circuits Group Inc (603501.SS)	0 (11/06/2023) Rmb458.05
Phison Electronics Corp (8299.TWO)	0 (05/09/2025) Rmb150.02 NT$882.00
SG Micro Corp. (300661.SZ)	0 (03/11/2025) 0 (06/17/2024)
Silergy Corp. (6415.TW)	Rmb85.16
SMIC (0981.HK)	0 (02/13/2025) NT$249.00
TSMC (2330.TW)	E (04/09/2025) HK$83.50
UMC (2303.TW)	0 (02/07/2022) NTS1,440.00
Vanguard International Semiconductor (5347.TWO)	E (10/28/2024) NTS45.65 NTS102.00
WIN Semiconductors Corp (3105.TWO)	U (08/05/2024) U (07/14/2025) NTS101.50
Daisy Dai, CFA
ASMPT Ltd (0522.HK)	0 (07/24/2025) HKS95.35
China Resources Microelectronics Limited (688396.SS)	E (06/19/2025) Rmb58.65
Elan Microelectronics Corp (2458.TW)	0 (10/03/2025) NTS132.50
Empyrean Technology Co Ltd (301269.SZ)	E(01/17/2025) Rmb131.82
Hangzhou Silan Microelectronics Co. Ltd.(600460.SS)	U (08/25/2025) Rmb32.54
JCET Group Co Ltd (600584.SS)
Shanghai Anlogic Infotech Co Ltd (688107.SS)	U (09/25/2024) Rmb47.00
Shanghai Fudan Microelectronics (1385.HK)	E (05/14/2024) Rmb32.33
SICC Co Ltd (688234.SS)	0 (03/07/2025) HKS45.14
StarPower Semiconductor Ltd (603290.SS)	E(09/03/2025) Rmb84.90 Rmb115.80
Unigroup Guoxin Microelectronics Co Ltd (002049.SZ)	0 (03/01/2022) U (01/10/2023) Rmb92.78
Universal Scientific Ind. (Shanghai) (601231.SS)	E(10/23/2024) Rmb22.88
AP Memory Technology Corp (6531.TW)	0 (07/11/2025) NTS392.50
	NTS1,460.00
ASMedia Technology Inc (5269.TW)	U (10/03/2025)
Aspeed Technology (5274.TWO)	0 (06/09/2025) NTS5,690.00
Egis Technology Inc (6462.TWO)	E (03/26/2024) NTS123.00
Espressif Systems (688018.SS)	0 (05/15/2023) Rmb216.18
GigaDevice Semiconductor Beijing Inc (603986.SS)	0 (05/15/2025) Rmb225.49
Macronix International Co Ltd (2337.TW)	0 (09/18/2025) NTS27.50
Montage Technology Co Ltd (688008.SS)	+ Rmb157.08
Novatek (3034.TW)	0 (03/18/2024) NT$424.00
Nuvoton Technology Corporation (4919.TW)	E (08/06/2025) NT$62.80
Parade Technologies Ltd (4966.TWO)	0 (10/03/2025) NTS781.00
Powerchip Semiconductor Manufacturing Co (6770.TW)	E (06/13/2024) NTS30.30
Realtek Semiconductor (2379.TW)	NTS555.00
Shenzhen Goodix Technology Co Ltd (603160.SS)	0 (06/19/2025)
Sino Wealth Electronic (300327.SZ)	U (07/14/2025) Rmb85.09
Winbond Electronics Corp (2344.TW)	U (03/31/2025) Rmb27.40
WPG Holdings (3702.TW)	0 (03/11/2025) NTS43.45 E (11/16/2023) NTS62.60
Duan Liu
Dosilicon Co Ltd (688110.SS)
Shenzhen Longsys Electronics Co Ltd (301308.SZ)	U (09/06/2024) Rmb109.30
Tiffany Yeh	0 (03/11/2025) Rmb186.20

AllRing Tech Co.(6187.TWO)	0 (09/23/2025) NTS376.00
FOCI Fiber Optic Communications Inc (3363.TWO)	0 (01/15/2025) NTS372.50
Himax Technologies Inc (HIMX.O)	0 (05/09/2025) US$9.03
Silicon Motion (SIMO.0)	0 (05/06/2024) US$93.86

Stock Ratings are subject to change. Please see latest research for each company. * Historical prices are not split adjusted.

全球人工智能供应链最新动态；亚洲半导体的关键机遇.pdf免费PDF下载