Foundation
86 2 2730-1725
+852 2848-6918
86 2 2730-2863
+852 2848-7310
+852 2239-7731
+852 3963-1592
86 2 7712-3032
86 2 2730-2875
+852 3963-2287
tors
tive
th
s a result,
a conflict of
Stanley
ley
vestment
closures,
d of this
t registered
member
Not for redistribution without written consent of Morgan Stanley
M
Greater China Semiconductors | Asia Pacific
China's AI Accelerators – Who's
Poised to Win?
We initiate coverage on Cambricon and Iluvatar at Overweigh
and MetaX at Equal-weight.
Bull vs. bear cases for China AI accelerators: Who will capture China’s inferen
driven substitution cycle? In a recent report, we argued that China is narrowing
US lead in AI compute not simply at the chip level but also through system-level
innovation, supply-chain localization, and increasingly attractive inference
economics. We continue to believe this path will lift China’s domestic AI accelera
self-sufficiency rate to 86% by 2030, reshaping the global competitive landscape
AI semiconductors over the next decade.
Over the past month, our channel checks have turned incrementally more
constructive on China’s AI accelerator industry. 1) At our China Summit, major
LLM developers such as MiniMax and Zhipu signaled their willingness to adopt
domestic AI chips as long as token economics are competitive. 2) Our recent field
trip indicated that the Nvidia GPU supply in China tightened after the SMCI-relat
disruption, redirecting incremental demand toward domestic alternatives. 3) Stro
spot demand for Nvidia RTX 5090 in China suggests AI inference demand remain
robust. 4) Rising token prices and GPU rental prices also point to a still-tight
compute market. The main negative datapoint: price competition appears to be
arriving earlier than we had expected, as some vendors have started cutting price
to gain share. 5) WAIC will be held in Shanghai in July 2026, where we expect to
next-generation Chinese AI accelerator products, especially from Iluvatar.
Stock calls: We initiate coverage on Cambricon and Iluvatar at Overweight, an
MetaX at Equal-weight. We believe all three are well positioned to benefit from
China’s accelerating AI chip localization trend, though each offers a differentiated
investment case.
• Cambricon (Overweight; PT Rmb1,588): We view Cambricon as the lead
domestic AI inference chip play, supported by strong CSP customer
anchoring, proven hardware-software co-optimization, and solid positioni
in large-scale cloud inference deployments.
• Iluvatar (Overweight; PT HK$600): We view Iluvatar favorably for its
diversified supply chain strategy, stronger supply visibility, and growing
exposure to cloud customers.
• MetaX (Equal-weight; PT Rmb758): We see MetaX as a differentiated
domestic GPGPU vendor, with relatively strong CUDA-like software
compatibility, and a more scalable near-term manufacturing path. Howev
its valuation is less attractive than peers.
Key risks to our view include: 1) slower AI demand growth, 2) earlier-than-expec
April 26, 2026 09:50 PM GMT
pricing pressure, 3) policy/export headwinds.
若有侵权，请联系删除。公众号
ng
er,
ted
Morgan Stanley does and seeks to do business wi
companies covered in Morgan Stanley Research. A
investors should be aware that the firm may have
interest that could affect the objectivity of Morgan
Research. Investors should consider Morgan Stan
Research as only a single factor in making their in
decision.
For analyst certification and other important dis
refer to the Disclosure Section, located at the en
report.
+= Analysts employed by . affiliates are no
with FINRA, may not be associated persons of the
and may not be subject to FINRA restrictions on
Morgan Stanley Taiwan Limited+
Charlie Chan
Equity Analyst
@ +8
Morgan Stanley Asia Limited+
Gary Yu
Equity Analyst
@
Morgan Stanley Taiwan Limited+
Daniel Yen, CFA
Equity Analyst
@ +8
Morgan Stanley Asia Limited+
Daisy Dai, CFA
Equity Analyst
@
Henry Zhao
Research Associate
@
Joanne Lau
Research Associate
@
Morgan Stanley Taiwan Limited+
Tiffany Yeh
Equity Analyst
@ +8
Lucas Wang
Research Associate
@ +8
Morgan Stanley Asia Limited+
Ethan Jia
Research Associate
@
Greater China Technology Semiconduc
Asia Pacific
Industry View Attrac
t,
ce-
the
tor
for
AI
ed
ng
s
s
see
d
ing
communications with a subject company, public appearances
and trading securities held by a research analyst account.
：趣享财经
mailto:@
mailto:@
mailto:@
mailto:@
mailto:@
mailto:@
mailto:@
mailto:@
mailto:@
M Foundation
2
Executive Summary – Don't Underestimate China's AI
Compute Ecosystem
Booming China AI GPU market
China’s AI GPU market is entering a more commercially grounded phase, with the debate
shifting from whether domestic chips can participate to which vendors will win meaningful
share as inference demand scales. In our view, two structural forces shape the market: (1)
a rapid rise in AI inference demand, driven by commercialization across consumer and
enterprise applications, and (2) persistent export controls, making localization a long-
duration feature of China’s AI compute market rather than a temporary policy response.
Together, these forces expand the addressable market for domestic AI accelerators and
improve the probability of sustained substitution. This aligns with our framework that
China's AI chip TAM could reach US$67bn by 2030, with domestic self-sufficiency rising to
86%.
Our core thesis remains that China’s localization strategy is gaining traction: scaling
domestic chips, foundries, packaging, and equipment capabilities to partially offset
process-node disadvantages. In the bull case, domestic AI semis broaden from inference
into selected training workloads, software ecosystems improve faster than expected, and
some vendors achieve overseas adoption or indirect export opportunities. In the bear case,
product differentiation fades, pricing pressure intensifies earlier than expected, and the
sector moves toward commoditization and consolidation.
Exhibit 1: China AI accelerators' market cap trend
-
50
100
150
200
250
300
Jul-25 Aug-25 Sep-25 Oct-25 Nov-25 Dec-25 Jan-26 Feb-26 Mar-26
Cambricon Hygon MetaX MooreThreads Biren Iluvatar CoreX
(US$ bn)
Source: FactSet, Morgan Stanley Research
Exhibit 2: China AI accelerators' revenue trend
Source: Company data, Morgan Stanley Research estimates
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 3
Exhibit 3:
"10 Dragons" of Chinese AI GPGPU vendors. We focus on Cambricon, MetaX , Iluvatar
company research in this report
Source: Company data, Morgan Stanley Research
China's AI compute industry can be competitive globally, given strong
system design and infrastructure
More broadly, we believe China’s AI GPU race is no longer just a chip-specification contest.
While domestic silicon still trails the US by roughly two generations at the chip level, the
effective gap is narrowing through multi-die design, advanced packaging, rack-scale system
architecture, optical networking, and software-hardware co-optimization. This is why we
think system-level competitiveness matters more than ever. In a market increasingly
dominated by inference and utilization, the vendor that delivers the best real-world token
economics at acceptable software migration cost is likely to win customer budgets, even
without leading-edge process technology.
From an investment perspective, this leads to a simple conclusion: the sector should not
be valued as a monolithic policy theme. Instead, investors need to distinguish between
vendors with a realistic path to shipment scale, ecosystem credibility, and pricing
discipline, and those that may struggle to convert technical potential into durable
revenues and margins. We therefore evaluate the group through a two-dimensional
framework of economics × execution, combining TCO, token cost, TPS, and performance-
per-dollar with qualitative factors such as foundry access, software ecosystem maturity,
CSP relationships, and roadmap credibility. In our view, that framework remains the most
effective way to separate likely winners from those at risk of being marginalized as the
industry consolidates.
What has become clearer through our recent channel checks is that economics, not
ideology, is driving adoption. At our China Summit, major LLM developers indicated they
are willing to deploy local GPUs as long as token cost is competitive. This aligns with our
若有侵权，请联系删除。公众号：趣享财经
M Foundation
4
core finding that domestic accelerators already offer materially lower TCO than Nvidia
products available in China and that leading Chinese chips can achieve broad cost-per-
token parity in inference workloads. In other words, purchasing decisions are increasingly
made on deployable economics rather than absolute peak silicon performance. This
matters because China’s AI demand is becoming more inference-heavy, more recurring, and
more utilization-driven, which structurally favors solutions optimized around cost
efficiency, software adaptation, and availability rather than headline benchmark
leadership.
Exhibit 4: Relative strengths of US and China AI industries
0
2
4
6
8
10
Wafer front-end
Chip packaging
Memory: HBM, LPDDR5
Server system
Optical networking
Software optimization
(LLM)
AI datacenter space
Power supply
Policy support
China US
Source: Morgan Stanley Research
Exhibit 5: Domestic chips have lower TCO and comparable per token cost (AI LLM inference) vs. NVIDIA's processors for China
-
H200 A100 H20 910B 910C 950PR MLU 370 MLU580 MLU 590 MLU 690 C500 C600 S5000 BI-V100 BI-V200 PPU P800
NVIDIA Huawei Cambricon MetaX Illuvatar T-Head Kunlun
TCO of 10MW capacity (US$ mn) Per token cost (US$ cent) (Right-axis) (Lower is better)
- China AI chip TCO could be 30-60% less than Nvidia's AI
processors.
- Per token cost of top Chinese AI accelerators can match
with or surpass Nvidia's processors for China.
Source: Company data, Morgan Stanley Research estimates
Who's poised to win in China AI GPUs?
China’s AI accelerator ecosystem spans merchant chip vendors, sovereign-backed players,
and captive chip design houses linked to major cloud service providers (CSPs). We assess
this ecosystem's competitors in a global GPU/ASIC context, compare relative positioning
across performance, cost, and execution, and apply a consistent valuation framework to
identify stocks with the most attractive risk/reward. Our field work with CSPs suggests
that while per-token cost is the single most important KPI, software optimization and
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 5
strategic customer partnerships matter even more than we had assumed.
Based on recent shipment trends, customer allocation, market share evolution, and the
earlier-than-expected onset of price erosion, we believe the next phase will be defined less
by theoretical peak performance and more by commercial execution, software readiness,
and customer capture.
From a competitive perspective, we believe the market should be segmented by customer
type. For top-tier CSPs and leading LLM developers, the primary decision metric is
increasingly per-token cost, but that KPI alone is not sufficient. In practice, software
maturity, framework compatibility, cluster-level optimization, and the depth of strategic
partnerships play a decisive role in order allocation. For sovereign AI, telecom, SOE, and
government-linked demand, supply security, domestic controllability, and policy alignment
carry greater weight. This creates room for different winners across end-markets. In our
view, vendors with strong CSP co-development relationships and credible software stacks
are better positioned to win high-volume cloud inference deployments, while vendors with
stronger domestic supply chain visibility or government relationships may be better
positioned in sovereign and public sector projects.
Exhibit 6: Order placement and potential orders for domestic AI accelerators developers, according to our industry checks
Source: Morgan Stanley Research
Within this framework, we see meaningful differentiation among Cambricon, MetaX,
Iluvatar. Cambricon stands out on the ASIC/DSA path, where its inference performance,
customer anchoring, and hardware-software co-optimization support strong deployment
economics, particularly in large-scale cloud use cases. Iluvatar is differentiated by its
diversified foundry strategy, better supply visibility, and a pragmatic path to customer
migration through software compatibility. MetaX is one of the more credible domestic
GPGPU players, in our view, due to relatively stronger CUDA-like software compatibility
and a manufacturing path that may prove more scalable near term.
In short, Cambricon looks strongest on current cloud inference traction (second only to
若有侵权，请联系删除。公众号：趣享财经
M Foundation
6
Huawei Ascend), Iluvatar on supply-chain resilience plus commercial optionality, and
MetaX on scalable GPGPU positioning.
Near-term market tracker for China AI GPU demand
At the same time, near-term industry conditions have turned more favorable for domestic
vendors. Our recent field trip suggests Nvidia GPU availability in China has tightened,
creating more room for local substitution. Strong spot-market demand for Nvidia 5090
cards, alongside higher token prices and GPU rental prices, also points to resilient
downstream inference demand. These datapoints reinforce our view that the demand
environment remains robust, especially for customers that need immediate deployment
rather than waiting for supply normalization. The caveat: competition is intensifying faster
than expected. We are already seeing price cuts in parts of the market, implying the sector
may move into a market-share phase earlier than we had assumed. As a result, we expect
execution quality to matter more than ever – particularly in software optimization,
customer support, and strategic account penetration.
The World Artificial Intelligence Conference (WAIC) will be held in Shanghai in July 2026,
where we expect to see next-generation Chinese AI accelerator products, especially from
Iluvatar.
Exhibit 7: Nvidia's 5090 price keeps rising in China
10,000
15,000
20,000
25,000
30,000
35,000
NVIDIA gaming graphic card price in China
4090 Distributor price on TaoBao (Rmb) 4090 Tag price (Rmb)
5090D Distributor price on TaoBao (Rmb) 5090/5090D Tag price (Rmb)
Source: Taobao, Morgan Stanley Research
Exhibit 8: Average token price for China's mainstream AI LLMs
1Q25 2Q25 3Q25 4Q25 1Q26
Input (Rmb/mn token) Output (Rmb/mn token)
Source: Company data, Morgan Stanley Research
Exhibit 9: Surge in ByteDance (Volcano Engine/Doubao)
tokens indicates high AI demand
-
300
600
900
1,200
1,500
1,800
2,100
2,400
Apr-24 Jun-24 Aug-24 Oct-24 Dec-24 Feb-25 Apr-25 Jun-25 Aug-25 Oct-25 Dec-25
(t
ri
ll
io
n
)
Monthly tokens processed
Bytedance (Volcano Engine / Doubao) (GLM) China (National Data Administration)
Source: Company data, Morgan Stanley Research. ByteDance numbers represent monthly run-rate based
on daily numbers.
Exhibit 10: China CSP's capex will be a key demand driver for
China AI GPU
Source: Company data, Morgan Stanley Research (E) estimates
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 7
Identifying Winners With Our Framework: Cambricon and
Iluvatar Are Our Preferred Plays
Building on our performance and cost analysis, we apply a structured framework to assess
domestic AI chip vendors’ relative positioning, focusing on quantitative economics and
qualitative execution.
Our framework: economics × execution
We evaluate vendors across two key dimensions:
• Inference economics (quantitative) – including TCO, cost per token, TPS
performance, and performance per watt/dollar
• Execution capability (qualitative) – including access to leading-node capacity,
software ecosystem maturity, depth of CSP relationships, and product roadmap
soundness
In our view, sustained leadership requires strength in both. Vendors that excel in only one
– ., strong in silicon but with a weak ecosystem – are unlikely to achieve durable share.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
8
Exhibit 11: Comparison among Cambricon, MetaX and Iluvatar
Ticker 688256-SH 688802-SH 9903-HK
Product
MLU 220/270/370/580/590/690
(AI Training + inference)
C Series (AI training + inference)
N Series (AI Inference)
G Series (Graphic rendering)
Tiangai 100/150/200/300 (AI training + inference)
Zhikai 100 (AI inference)
GPGPU/ASIC ASIC GPGPU GPGPU
Chip suppliers
Process node for latest
products
7nm/N+2 12nm/N+1 7nm
Secured orders from
major CSPs
✔ X ✔
Sovereign fund as major
shareholder
X ✔ X
Per token cost
performance
2025 Revenue
(Rmb mn)
CNY 6,497 CNY 1,644 CNY 1,034
Profitablity ✔ X X
Source: Company data, Morgan Stanley Research
Cambricon: Leading in inference performance and customer
anchoring
Within this framework, we see Cambricon as one of the strongest positioned players on
the ASIC (DSA) pathway.
Quantitatively, Cambricon’s latest generation (., MLU590) delivers competitive
inference performance, with our TPS analysis suggesting meaningful outperformance vs.
NVIDIA H20 under certain DeepSeek R1 scenarios. Combined with competitive pricing, this
supports strong cost-per-token economics, which we view as the key CSP decision metric.
Qualitatively, Cambricon benefits from deep customer integration. Based on our industry
checks, its multi-year collaboration with ByteDance enabled continuous hardware–
software co-optimization and real-world deployment validation, providing an advantage in
application-level tuning and commercialization readiness.
Taken together, we view Cambricon as a near-term leader in inference-driven deployments,
particularly where efficiency and customer-specific optimization are critical.
Iluvatar: Leveraging supply chain resilience with strong order
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 9
visibility
We believe Iluvatar is well positioned to benefit from accelerating domestic AI chip
substitution in China, supported by supply chain resilience, software compatibility, and
improving commercial traction.
Based on our industry checks, leading Chinese CSPs have placed sizable pre-orders for
Iluvatar's TianGai-150 AI chips, with shipments expected to commence in 2H26.
Importantly, Iluvatar’s diversified foundry strategy – including export-compliant
production at TSMC – offers greater capacity visibility vs. peers relying solely on domestic
fabs or non-compliant manufacture overseas, reducing supply disruption risk.
On software, Iluvatar’s GPGPU architecture offers high CUDA compatibility, lowering
migration friction. The company has helped clients to migrate LLM stacks from NVIDIA
platforms to TianGai-150. In our view, this positions Iluvatar favorably as enterprises seek
pragmatic NVIDIA alternatives.
MetaX: Positioning for scalability through software and
supply
Within the GPGPU pathway, we view MetaX as a credible domestic participant, supported
by its focus on improving CUDA ecosystem compatibility. While CUDA remains NVIDIA’s
key moat – given deep integration across compilers, libraries (., cuDNN, NCCL), and a
large developer base – it also creates structural switching costs that are not easily
replicated.
Against this backdrop, MetaX’s strategy of building a CUDA-like software stack and
compatibility layer provides a reasonable adoption pathway for domestic customers.
Based on our industry checks, the company has made steady progress in compiler
adaptation, framework compatibility (., PyTorch), and runtime optimization, although
overall ecosystem maturity and stability still lag global leaders.
In addition, MetaX adopts a pragmatic manufacturing strategy, leveraging relatively
mature nodes (., N+1/12nm) to support yield stability and supply availability. While this
may limit peak performance vs. leading-edge products, it offers a more balanced trade-off
among performance, cost, and manufacturability. Overall, we view MetaX as a company
with improving execution and scalability potential, though further validation in large-scale
commercial deployments remains a key factor to monitor.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
10
Performance and Cost: Which Domestic AI Chips Stand Out?
Inference economics more important than foundation model
training for China AI GPU market
In a prior China AI Insights note, we conducted a comprehensive comparison of domestic
AI accelerators across key performance and economic metrics, including total cost of
ownership (TCO), total processing performance (TPP), token output per second (TPS)
under DeepSeek R1 inference, and performance per watt.
1. TPS - the revenue
We believe TPS (tokens per second) is another relevant metric for China’s
inference driven market. Unlike peak FLOPS, TPS captures end to end system
performance, reflecting hardware capability (compute throughput, memory and
interconnect bandwidth) and software efficiency under real world workloads. Using
DeepSeek R1 as our benchmark and calibrating assumptions against NVIDIA’s disclosed
H200 result (5,899 TPS in Feb 2025), we find that leading domestic accelerators – such
as Huawei’s Ascend 950PR/DT and Cambricon’s MLU690 – can outperform NVIDIA H20
by 50-150% in our scenarios. This reflects improvements in compute capability and
system level optimization and compute to network balance.
Domestic vendors have made meaningful progress through memory, interconnect, and
system architecture improvements, enabling competitive inference performance despite
process node disadvantages. This reinforces our view that performance leadership is
increasingly workload dependent, with domestic chips already competitive in inference
scenarios, even as NVIDIA maintains an edge at the technology frontier.
Exhibit 12: TPS (tokens per second) analysis for China AI accelerators
5,899
1,365
1,033
777
1,536 1,464
170
956 942
2,063
1,201
538
1,071
247
1,236
327
1,341
570 625
1,521
932
1,179
226
932
-
1,000
2,000
3,000
4,000
5,000
6,000
7,000
H200 A100 H20 910B 910C 950PR MLU 370 MLU580 MLU 590 MLU 690 BW1000 C500 C600 S4000 S5000 BI-V100 BI-V200 BR166 P800 PPU R300A P800 I20 L600
NVIDIA Huawei Cambricon Hygon MetaX MooreThread Illuvatar Biren T-Head Kunlun Enflame
AI inference TPS per AI accelerator with DeepSeek R1 (Tokens/s)
Source: Company data, Morgan Stanley Research estimates (*Our assumptions include: AI LLM: DeepSeek R1; Input token: 1,024; Output token: 1,024; Number of active experts: 9 out of 257; Model size: 671GB; Model
layer: 61; Batch size: 1; Computing power: FP8, and if FP8 is not available, we will use FP16).
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 11
Exhibit 13: Our TPS result matches with NVIDIA's published
result for H200 (Feb 2025), with same token length and data
format assumptions
Source: NVIDIA
Exhibit 14: GB300 could deliver 50x of token per watt
performance vs. H200 with DeepSeek R1
Source: NVIDIA
TPS Methodology
Our TPS framework uses core hardware and workload variables that determine token
throughput in inference. Hardware inputs include: (1) effective compute throughput (FP8,
or FP16 where FP8 is unavailable), (2) memory bandwidth, (3) interconnect bandwidth,
and (4) chip utilization rate (UTR). Workload inputs include model size (671GB), layers
(61), active experts (9/257 under MoE), input/output token length (1,024/1,024), and batch
size (1).
We select DeepSeek R1 as the benchmark model given its representativeness within
China’s current LLM ecosystem and its MoE architecture. To calibrate our framework, we
benchmarked NVIDIA H200 output figures disclosed in 1Q25 with DeepSeek R1.
Specifically, in February 2025, NVIDIA's H200 achieved 5,899 TPS when running
DeepSeek R1 inference.
Exhibit 15: Our key TPS formulas
Source: Morgan Stanley Research
Limitations
若有侵权，请联系删除。公众号：趣享财经
M Foundation
12
Our TPS estimates rely primarily on hardware specifications (compute throughput,
memory bandwidth, networking bandwidth).
A critical variable in our model is chip utilization rate (UTR), which reflects how effectively
theoretical hardware throughput translates into sustained real-world performance. We
fine-tuned our UTR assumptions such that our modeled TPS approximates NVIDIA’s
published results. This calibration step helps anchor our model to observable market data
rather than relying purely on theoretical peak specifications.
It is also important to note that NVIDIA’s published results likely incorporate multiple
layers of software optimization, including kernel-level tuning, TensorRT graph
optimization, improved memory scheduling, communication library optimization, and
refined expert routing for MoE models. As such, the NVIDIA benchmark (Feb 2025 result)
we adopted could be viewed as a point-in-time reference that may improve further with
subsequent driver, framework or firmware updates.
Accordingly, while our framework provides a structured, hardware grounded basis for
comparing inference performance across vendors, realized TPS may vary depending on
software maturity and optimization, workload mix, and cluster configuration. Our analysis
assumes a fixed inference workload (input/output tokens of 1,024/1,024 and batch size of
1), which may not reflect all deployment scenarios.
Of note, when assessing NVIDIA’s next-generation platforms, such as the GB300, the
performance gap widens materially. In NVIDIA’s latest publication (link), the company
indicates that GB300 delivers up to 50x improvement in token-per-watt performance
versus H200 with DeepSeek R1, driven by significantly higher compute throughput,
enhanced networking bandwidth, next-generation HBM performance, and support for
lower-precision formats, such as FP4.
Importantly, our comparative analysis in this report is confined to products that China can
currently procure or may reasonably access in the near term. As such, our benchmarking
framework does not incorporate NVIDIA’s latest frontier platforms. In absolute terms,
NVIDIA’s most advanced systems remain meaningfully ahead of the products included in
our modeling. Accordingly, our conclusions should be interpreted within the context of
China-accessible supply rather than global technology leadership at the cutting edge.
2. TCO - the cost
TCO remains one of the most compelling advantages for domestic AI chips. On an all-in
basis – including chip acquisition, power, and supporting infrastructure – we estimate
domestic accelerators can deliver 30-60% lower TCO than NVIDIA solutions currently
accessible in China. Lower upfront chip pricing and structurally lower electricity and
infrastructure costs in China drive this edge. The advantage becomes more pronounced at
scale, particularly in inference heavy deployments where high utilization pushes
operating expenses to dominate lifecycle costs.
When we translate system performance into cost per token, the gap narrows. While
NVIDIA retains an absolute performance lead – especially at the high end – leading
domestic accelerators have reached broad cost per token parity with NVIDIA’s A100/
H20 class products and in certain configurations may outperform. In our view, this marks
a critical inflection as CSPs optimize for monetization and utilization rather than peak
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 13
silicon capability.
Exhibit 16: AI GPU pricing: China vs. the US – China generally sells at a lower price given lower margin requirements
$0
$5,000
$10,000
$15,000
$20,000
$25,000
$30,000
$35,000
A100 H20 H200 910B 910C 950PR* MLU 370 MLU 580* MLU 590* MLU 690* C500 C600* BI-V100* BI-V200* S5000* PPU* P800*
Nvidia Huawei Ascend Cambricon MetaX Illuvatar MooreThread T-Head* Kunlun*
Market price (US$)
Source: Morgan Stanley Research (*Based on Morgan Stanley Research estimates, NVIDIA's legacy chip is according to original ticker price)
Exhibit 17: Electricity costs: China enjoys much cheaper power than other economies
98
105
75
110
68
73
44
78
110
59
72
100
55
88
100
76
115
75
82
70
103
140
98
120
110
58
(110)
(90)
(70)
(50)
(30)
(10)
10
30
50
25
45
65
85
105
125
145
165
185
China Philippines
Spot Price
Australia Spot
Price
Thailand
Regulated
Tariff
US Industrial
Tariff
Japan Spot
Price
India Spot
Price
South Korea
Spot Price
Malaysia
Datacenter
Tariff
Europe Spot
Price
Singapore
Spot Price
Dubai Cloud
Computing
Saudi Arabia
Cloud
Computing
How power prices stack up (US$/MWh)
2019 2025E ∆ in power prices
Source: Company data, Bloomberg, IEX, EMA Singapore, EPPO Thailand, Morgan Stanley Research
Exhibit 18: Domestic chips have lower TCO and comparable per token cost (AI LLM inference) vs. NVIDIA's processors for China
-
H200 A100 H20 910B 910C 950PR MLU 370 MLU580 MLU 590 MLU 690 C500 C600 S5000 BI-V100 BI-V200 PPU P800
NVIDIA Huawei Cambricon MetaX Illuvatar T-Head Kunlun
TCO of 10MW capacity (US$ mn) Per token cost (US$ cent) (Right-axis) (Lower is better)
- China AI chip TCO could be 30-60% less than Nvidia's AI
processors.
- Per token cost of top Chinese AI accelerators can match
with or surpass Nvidia's processors for China.
Source: Company data, Morgan Stanley Research estimates
3. Energy efficiency - the power consumption
若有侵权，请联系删除。公众号：趣享财经
M Foundation
14
On energy efficiency, domestic chips have largely closed the gap with NVIDIA’s A100 and
H20, although they still lag newer platforms such as H100 and H200. Meanwhile, on
performance per dollar, domestic accelerators already show clear advantages, driven by
lower pricing and improving real world performance.
Incorporating acquisition cost, domestic chips deliver stronger performance per dollar due
to materially lower pricing. On this basis, leading domestic accelerators already
outperform the A100 and narrow the gap with the H200, reinforcing their appeal for
inference heavy deployments.
Exhibit 19: Performance/watt comparison
910B 910C 950PR* MLU 370 MLU 580* MLU 590* MLU 690* C500 C600* BI-V100* BI-V200* S5000* PPU* P800*
Huawei Ascend Cambricon MetaX Illuvatar MooreThread T-Head* Kunlun*
Performance/Watt
NVDA H200
NVDA A100
Source: Company data, Morgan Stanley Research (*Based on Morgan Stanley Research estimates)
Exhibit 20: Performance/cost comparison
910B 910C 950PR* MLU 370 MLU 580* MLU 590* MLU 690* C500 C600* BI-V100* BI-V200* S5000* PPU* P800*
Huawei Ascend Cambricon MetaX Illuvatar MooreThread T-Head* Kunlun*
Performance/Cost
NVDA A100
NVDA H200
Source: Company data, Morgan Stanley Research (*Based on Morgan Stanley Research estimates)
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 15
Exhibit 21: Cost efficiency and power efficiency comparison
Brand
Moore Thread T-Head Kunlun
Type GP GPU GP GPU GP GPU GP GPU ASIC (DSA) ASIC (DSA) GP GPU ASIC (DSA) ASIC (DSA) ASIC (DSA) ASIC (DSA) GP GPU GP GPU GP GPU GP GPU GP GPU GP GPU ASIC (DSA)
Product A100 H20 H100 H200 910B 910C 950PR* MLU 370 MLU 580* MLU 590* MLU 690* C500 C600* BI-V100* BI-V200* S5000* PPU* P800*
Foundry
Die size (mm^2) 826 814 814 814 666 666 480 450 650 750 650 800 700 400 500 700 750 750
Dies per package 1 1 1 1 1 2 2 1 1 1 2 1 2 1 2 1 1 1
Gross die per wafer 62 63 63 63 80 80 116 125 82 69 82 64 75 142 142 75 69 69
Yield (%) 70% 50% 50% 50% 70% 70% 30% 70% 45% 70% 25% 70% 45% 75% 80% 70% 45% 55%
Wafer price (US$) $11,000 $18,000 $18,000 $18,000 $11,000 $11,000 $12,000 $11,000 $9,000 $11,000 $12,000 $12,000 $9,000 $11,000 $11,000 $11,000 $9,000 $12,000
Die cost per chip (US$) $253 $571 $571 $571 $196 $393 $690 $126 $244 $228 $1,171 $268 $533 $103 $194 $210 $290 $316
Memory
Memory size (GB) 80 96 80 144 64 96 128 48 96 96 144 96 144 64 96 80 96 96
Memory type HBM2e HBM3 HBM3 HBM3 HBM2e HBM3 LPDDR5 LPDDR5 HBM3 HBM3 HBM3e HBM3 HBM3e HBM2e HBM3 HBM3 HBM3 HBM3
US$/16GB $160 $200 $200 $200 $160 $200 $150 $150 $200 $200 $220 $200 $220 $160 $200 $200 $200 $200
Memory cost (US$) $800 $1,200 $1,000 $1,800 $640 $1,200 $1,200 $450 $1,200 $1,200 $1,980 $1,200 $1,980 $640 $1,200 $1,000 $1,200 $1,200
Packaging
Chips per CoWoS/ packaging 30 30 30 30 30 15 15 30 30 30 15 30 15 30 15 30 30 30
CoWoS wafer price (US$) $10,000 $10,000 $10,000 $10,000 $10,000 $10,000 $7,000 $10,000 $7,000 $10,000 $7,000 $10,000 $7,000 $10,000 $10,000 $10,000 $7,000 $7,000
Packaging and testing cost (US$) $567 $567 $567 $567 $567 $1,133 $793 $567 $397 $567 $793 $567 $793 $567 $1,133 $567 $397 $397
Chip costs
Chip manufacuring costs (US$) $1,620 $2,338 $2,138 $2,938 $1,403 $2,726 $2,683 $1,142 $1,841 $1,994 $3,944 $2,035 $3,307 $1,310 $2,527 $1,776 $1,887 $1,913
Other overheads (US$) $688 $558 $662 $262 $197 $474 $488 $143 $659 $606 $856 $251 $408 $747 $559 $601 $685 $659
Gross margin assumption (%) 80% 69% 86% 84% 60% 60% 63% 55% 65% 65% 72% 60% 60% 64% 55% 68% 76% 70%
Estimated chip cost (US$) $2,308 $2,896 $2,800 $3,200 $1,600 $3,200 $3,171 $1,286 $2,500 $2,600 $4,800 $2,286 $3,714 $2,057 $3,086 $2,377 $2,571 $2,571
Chip price to China CSP
Market price (Rmb) NA CNY 85,000 CNY 210,000 CNY 210,000 CNY 35,000 CNY 70,000 CNY 60,000 CNY 25,000 CNY 50,000 CNY 65,000 CNY 120,000 CNY 50,000 CNY 65,000 CNY 50,000 CNY 60,000 CNY 65,000 CNY 75,000 CNY 75,000
Market price (US$) $15,000 $12,143 $30,000 $30,000 $5,000 $10,000 $8,571 $3,571 $7,143 $9,286 $17,143 $7,143 $9,286 $7,143 $8,571 $9,286 $10,714 $10,714
Import cost markup (%) 30% 30% 50% 50% 20% 20% 0% 20% 0% 20% 0% 20% 0% 20% 20% 20% 0% 20%
Orginal chip price in USD $11,538 $9,341 $20,000 $20,000 $4,000 $8,000 $8,571 $2,857 $7,143 $7,429 $17,143 $5,714 $9,286 $5,714 $6,857 $7,429 $10,714 $8,571
Peformance (TFLOPS at FP16) 624 148 1,979 1,979 400 800 500 96 280 315 700 240 300 128 380 500 350 316
Peformance (TFLOPS at FP8) X 296 3,958 3,958 X X 1,000 X 560 630 1,400 X 600 X 760 1,000 700 632
Performance/Cost
Power consumption (W) 400 400 700 700 400 750 580 250 560 580 750 350 450 350 700 700 700 750
Performance/watt
Nvidia Huawei Ascend Cambricon IlluvatarMetaX
Source: Company data, Morgan Stanley Research (*Based on Morgan Stanley Research estimates)
4. GPGPU vs. ASIC product definition, does that matter?
In assessing domestic AI chip competitiveness, we view architectural choice – GPGPU vs.
ASIC (DSA) – as a fundamental strategic trade off in China’s AI ecosystem. Unlike global
peers with access to leading edge nodes and mature software, domestic vendors must
optimize across manufacturing limitations, software maturity, and evolving demand
structures.
Architecturally, GPGPUs offer superior programmability and flexibility, suiting a rapidly
evolving model landscape and compatibility with mainstream frameworks such as
PyTorch. This flexibility is particularly valuable in China, where software ecosystems are
still converging and model architectures continue to iterate. However, GPGPUs are less
power and area efficient and typically require more advanced nodes to achieve
competitive performance.
In contrast, ASIC based designs enable tighter hardware–software co optimization,
allowing vendors to partially offset process disadvantages and deliver higher efficiency in
targeted workloads. This makes ASICs attractive for inference heavy or relatively stable
use cases. The trade offs include reduced generality, higher software porting costs, and
greater sensitivity to changes in model architecture.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
16
Exhibit 22: GPGPU vs. ASIC: What's the difference?
Nvidia H200 Huawei Ascend 950PR
Type GPGPU ASIC(DSA)
General computing
capability
Strong (AI, graphic, scentific
computing, etc)
Relatively weaker (Focus on AI
acceleration)
AI Matrix computation Implemented via Tensor Cores Natively implemented via 3D Cube
Ecosystem CUDA
CANN (Compatible with PyTorch,
etc.)
Main use case General-purpose AI acceleration AI LLM training and inference
Source: Morgan Stanley Research
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 17
Cambricon (): Leading in inference performance
and customer anchoring; Initiate at OW
Cambricon is a leading domestic AI chip designer, focusing on
cloud and edge AI processors with proprietary architecture and
full-stack capabilities. We view it as one of China’s core players in
AI accelerators, positioned to benefit from rising domestic AI
compute demand and ongoing localization trends under
geopolitical constraints. Its MLU series chips have achieved
meaningful deployment in major CSP workloads (., search,
advertising, recommendation), demonstrating improving product-
market fit and supporting future order visibility. The company’s
ongoing transition toward a chip-centric revenue model, combined
with progress in domestic supply chain substitution, enhances
revenue quality and scalability. In addition, continued product
iteration (., next-generation MLU roadmap) and software
ecosystem development (NeuWare) provide potential upside as
adoption deepens across large-scale AI clusters.
Earnings: Our EPS estimates for 2026 are above consensus. We
expect revenue growth to remain solid, supported by continued
CSP order ramp and product iteration, although profitability may
be influenced by supply chain dynamics and R&D intensity.
Valuation: We derive our price target using a residual income
model. Our price target implies a 110x 2026e P/E multiple and 32x
2026e P/S, which we believe is justified given the company’s 90%
revenue CAGR over 2025-28 and its positioning in China’s AI
semiconductor market.
Key downside risks include: 1) slower-than-expected AI demand,
2) customer concentration and order volatility, and 3) supply
chain and yield constraints.
Exhibit 23: Cambricon: Summary of key metrics
Source:
Company data, FactSet, Morgan Stanley Research. e = Morgan Stanley Research estimates.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
18
Cambricon: Financial Summary
Exhibit 24: Cambricon: Quarterly financials
Source: Company data, Morgan Stanley Research. e = Morgan Stanley Research estimates.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 19
Exhibit 25: Cambricon: Financial summary
Source: Company data, Morgan Stanley Research. e = Morgan Stanley Research estimates.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
20
Risk Reward - Cambricon Technology Corporation
()
Rmb1,
Key valuation assumptions underpinning our model include: an % cost of equity (derived
from a beta of , risk-free rate of %, and equity risk premium of %), a long-term
payout ratio of 26%, a medium-term growth rate of 16%, and a perpetual terminal growth
rate of %.
MS Rating
▪ Deep engagement with major CSPs with
proven product-market fit, with MLU590
widely deployed in SAD workloads,
supporting strong order visibility as AI
infrastructure scales.
▪ Domestic supply chain transition is
underway, with production shifting to SMIC;
while yields remain a challenge, gradual
improvement is expected through 2026.
▪ Aggressive product roadmap, with next-
generation MLU690 expected in 4Q26,
potentially delivering a ~ performance
uplift and sustaining technology leadership.
▪ Our price target implies 32x of 2026e P/S ,
which is slightly lower than its historical
average P/S multiple.
Consensus Rating Distribution
86% Overweight
14% Equal-weight
0% Underweight
Source: Refinitiv, Morgan Stanley Research
Risk Reward Themes
New Data Era: Positive
Secular Growth: Positive
Technology Diffusion: Positive
View descriptions of Risk Rewards Themes here
Rmb3,
61x 2026e P/S
We assume (1) >120% revenue CAGR in
2025-28 driven by stronger-than-expected
domestic generative AI infrastructure build-
out and large-scale procurement of cloud
training & inference chips; (2) sustained
share gains in China’s domestic high-
performance AI chip market; (3) gross
margin improves to over 55% in 2026 and
2027 thanks to production scale effects,
mature technology iteration and optimized
high-value product structure.
Rmb1,
32x 2026e P/S
We assume (1) 90% revenue CAGR in 2025-
28 driven by domestic generative AI
infrastructure spending from CSP clients; (2)
gross margin reaches 51% and 48% in 2026
and 2027 given lower selling prices due to
products with smaller die size.
16x 2026e P/S
We assume (1) <60% revenue CAGR in
2025-28 given slower-than-expected
domestic AI capital expenditure and LLM
commercialization progress; (2) market
share loss in China’s AI chip market, and
failure to secure continuous mass orders
from top-tier customers; and (3) gross
margin falls below 40% in 2026 and 2027
due to fierce price competition, rising wafer
fabrication costs and underutilized
production capacity.
Risk Reward – Cambricon Technology Corporation ()
Leading in inference performance and customer anchoring
PRICE TARGET
RISK REWARD CHART
Key: Historical Stock Performance Current Stock Price Price Target
Source: Refinitiv, Morgan Stanley Research
Rmb1,,,
Rmb3,,(+%)(+%)Rmb3,(+%)
Rmb1,,(+%)(+%)Rmb1,(+%)
(%)(%)(%)
APR '25 OCT '25 APR '26 APR '27
0
800
1600
2400
3200
CNY
OVERWEIGHT THESIS
BULL CASE BASE CASE BEAR CASE
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 21
Drivers 2025 2026e 2027e 2028e
Gross Profit (YoY) (%)
CSP capex expansion
Large model deployment growth
Domestic substitution tailwind
Product iteration (MLU roadmap)
Software ecosystem improvement
Policy support for AI chips
New customer penetration
Scaling of AI clusters
100% Mainland China
Source: Morgan Stanley Research Estimate
View explanation of regional hierarchies here
Stronger-than-expected AI demand
CSP order ramp-up
Accelerating localization
Capacity and yield constraints
Customer concentration risk
Slower technology iteration
Morgan Stanley EstimatesMean
Source: Refinitiv, Morgan Stanley Research
FY 2026e
Sales /
Revenue
(Rmb, mn)
12, 20,
Net income
(Rmb, mn)
3, 6,156
EPS
(Rmb)
Risk Reward – Cambricon Technology Corporation ()
KEY EARNINGS INPUTS
INVESTMENT DRIVERS
GLOBAL REVENUE EXPOSURE
RISKS TO PT/RATING
RISKS TO UPSIDE
RISKS TO DOWNSIDE
MS ESTIMATES VS. CONSENSUS
14,
4,
若有侵权，请联系删除。公众号：趣享财经
M Foundation
22
Cambricon: Investment Positives & Concerns
Positives
Deeply entrenched with major CSP customers with proven product-market fit.
According to our industry checks, Cambricon's MLU590 chips are widely deployed in
search, advertising, and recommendation (SAD) systems across major cloud service
providers. These deployments demonstrate strong performance and reliability, providing
visibility into future orders as CSPs scale AI infrastructure.
Domestic supply chain transition is progressing. While historically reliant on overseas
foundries, our industry checks confirm Cambricon began shifting wafer production to
SMIC in 2025. Although yields remain challenged on advanced nodes, we expect steady
improvement through 2026 as processes mature.
First profitable domestic AI chip company validates business model. Cambricon
achieved full-year profitability in 2025, generating net income on Rmb65bn
revenue. This milestone shows domestic AI chip companies can achieve sustainable
profitability at scale, a critical validation for the sector.
Aggressive product roadmap maintains technology leadership. According to our
industry checks, the company may launch and ship its next-generation MLU690 chip in
4Q26, delivering ~ performance improvement (as gauged by our TPS result).
Concerns
Sanctions and capacity risk: The company was added to the . Entity List in 2022,
cutting off access to TSMC’s 7nm manufacturing. It has since shifted to SMIC’s N+2
process, but yield rates for large-die chips remain low, leading to supply constraints.
SMIC yield constraints limit capacity and margin upside. Our industry checks indicate
MLU580 adopted local wafer manufacturing and is still experiencing limitations from
SMIC's capacity and yields. This highlights the risk that slower-than-expected yield
improvement and lower-than-expected capacity allocation could delay order fulfillment
and pressure gross margins.
Rising competition in AI semi market. Competitors like Huawei Ascend and Iluvatar are
gaining traction with CSP and government customers, which may compress Cambricon’s
market share, with potential price competition further squeezing margins.
High customer concentration creates revenue volatility. Given Bytedance is a key
account, its capex and allocation decisions could create significant uncertainty for order
flow and revenue. In addition the top five customers accounted for % of revenue in
2024, with Bytedance contributing %. The loss of any key customer could have a
material impact on revenue.
EDA tool restriction risk: Chip design relies on EDA tools from Cadence and Synopsys.
Further restrictions could impact product development cycles and iteration speed.
若有侵权，请联系删除。公众号：趣享财经
Foundation
-27.
tic
That said,
pacity
cast
g a 90%
that have
d has
le to
028,
s Diff.
-8%
-19%
-12%
-11%
-6%
-7%
M
Cambricon: Where We Are Vs Consensus
We are more bullish than consensus on the company's revenue trajectory in 2026
We believe Cambricon's 2026 revenue will still be partially constrained by domes
advanced logic foundry capacity, which will result in limited MLU580 shipments.
we still project 222% revenue growth for 2026. As domestic wafer production ca
constraints and advanced packaging bottlenecks are gradually alleviated, we fore
Cambricon's revenue will rise to Rmb33bn in 2027 and Rmb44bn in 2028, implyin
CAGR in 2025-28.
Turning to profitability, Cambricon is among the few domestic AI chip companies
successfully navigated the early-stage high-investment and order-scarce phase, an
now achieved sustainable profitability traction. We project net income attributab
shareholders will reach in 2026, in 2027 and in 2
implying net margins of %, % and %, respectively.
Exhibit 26: Cambricon earnings estimates
US$ mn Mse Consensus Diff. Mse Consensus Diff. Mse Consensu
Net sales 20,944 14,057 49% 33,186 24,196 37% 44,256 47,904
Gross profit 10,588 7,703 37% 16,009 13,284 21% 20,801 25,606
Operating profit 6,240 5,221 20% 11,097 9,386 18% 15,772 17,994
Pretax Income 6,356 5,111 24% 11,244 9,336 20% 15,934 17,994
Net income 6,156 5,042 22% 9,558 8,643 11% 13,544 14,361
EPS for consensus 19% 8%
Margins
Gross margin % % % % % %
Operating margin % % % % % %
Pretax margin % % % % % %
Net margin % % % % % %
Opex % % % % % % %
2026E 2027E 2028E
Source: Bloomberg, Morgan Stanley Research estimates
Morgan Stanley Research 23
若有侵权，请联系删除。公众号：趣享财经
M Foundation
24
Cambricon: Valuation
We set our 12-month price target at Rmb1,588. With Cambricon now consistently
profitable, we value the company using a residual income model, which we believe
appropriately captures intrinsic value by fully incorporating our earnings forecasts.
Key assumptions: an % cost of equity (beta , risk-free rate %, equity risk
premium %), a 26% long-term payout ratio, 16% medium-term growth, and a %
perpetual terminal growth rate. We note that due to economic burdens, China is still in a
low-interest-rate environment. Therefore, we use a relatively low 2% as the risk-free rate.
Regarding the perpetual growth rate, the three Chinese AI accelerator companies covered
in this report are all relatively young, with their establishment less than a decade ago.
Amid the rapid development of AI, we believe these firms will maintain a phase of high
growth and volume expansion over the next ten years, justifying a perpetual growth rate
of 6%. Furthermore, when compared with the AI semiconductor leader NVIDIA, the
market capitalization of companies such as Cambricon may only equate to % of
NVIDIA’s level. This also appears low, given China’s strong push to develop large language
models and the broader AI market.
Our price target implies a 110x 2026e P/E multiple and 32x 2026e P/S.
Exhibit 27: Cambricon: Residual income valuation
Rmb million 2026E 2027E 2028E 2029E 2030E 2031E 2032E 2033E 2034E 2035E 2036E 2037E
Total Equity 17,499 25,558 36,602 48,228 61,715 77,360 95,507 116,559 140,978 169,305 202,164 240,280
Net Profit 6,156 9,558 13,544 15,711 18,225 21,141 24,524 28,448 32,999 38,279 44,404 51,509
ROAE % % % % % % % % % % % %
Residual Income 3,979 6,306 9,001 10,498 11,958 13,604 15,482 17,639 20,126 22,999 26,322 30,170
Spread % % % % % % % % % % % %
Ending Equity Capital 17,499
PV of Forecast Period 96,661
PV of Continuing Value 560,292
Equity Value 674,452
No. of Shares 425
Price Target 1,588
Source: Company data, Morgan Stanley Research estimates
Bull & bear cases
Our bull case value is Rmb3,000, and assumes a stronger revenue ramp with accelerated
gross margin expansion and domestic AI chip substitution.
We assume (1) >120% revenue CAGR in 2025-28 driven by stronger-than-expected
domestic generative AI infrastructure build-out and large-scale procurement of cloud
training & inference chips; (2) sustained market share gains in China’s high-performance AI
chip market, with breakthroughs in long-term mass procurement contracts from key
internet giants and state-owned enterprise customers; (3) gross margin above 55% in
2026 and 2027 thanks to production scale effects, mature technology iteration, and an
optimized high-value product mix.
Our bear case value is Rmb800, and assumes slower revenue growth with gross margin
erosion amid intensifying competition.
We assume (1) <60% revenue CAGR in 2025-28 given slower-than-expected domestic AI
capex and LLM commercialization; (2) market share loss in China’s AI chip market and
failure to secure continuous mass orders from top-tier customers; (3) gross margin below
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 25
40% in 2026 and 2027 due to fierce price competition, rising wafer fabrication costs, and
underutilized capacity.
Peer comparison
We expect Cambricon's EPS to increase at an 87% CAGR from 2025 to 2028. We also
project revenue to deliver a 90% CAGR over the same period. We attribute the higher
earnings and revenue growth to surging capex in domestic AI computing infrastructure
and generative AI capacity build-out by major clients – including CSPs, telecom operators,
and state-owned enterprises – as well as accelerated domestic substitution of overseas
high-end AI chips amid supply chain restrictions. In our view, the higher earnings growth
justifies Cambricon's high multiples of 110x 2026e P/E and 32x 2026e P/S.
Exhibit 28: China AI semi peer comparison table
Source: Company data, FactSet, Morgan Stanley Research
Exhibit 29: Cambricon P/S multiple
0x
20x
40x
60x
80x
100x
120x
140x
21-01 21-07 22-01 22-07 23-01 23-07 24-01 24-07 25-01 25-07 26-01
Cambricon one-year forward P/S
Source: Company data, FactSet, Morgan Stanley Research estimates
若有侵权，请联系删除。公众号：趣享财经
M Foundation
26
Exhibit 30: Cambricon P/E multiple
0x
20x
40x
60x
80x
100x
120x
140x
160x
25-01 25-07 26-01
Cambricon one-year forward P/E
Source: Company data, FactSet, Morgan Stanley Research estimates
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 27
Iluvatar (): Leveraging supply chain resilience with
strong order visibility; Initiate at OW
Iluvatar is a domestic GPU designer focused on general-purpose AI
computing, with products spanning AI training and inference. We
initiate coverage with an Overweight rating, as we believe the
company is well positioned to benefit from accelerating domestic
substitution, supported by improving supply chain resilience,
secured CSP orders, and a visible path to profitability. In
particular, upcoming shipment ramp of the TianGai-150 chips from
2Q26, with a stronger ramp in 2H26, provides clear near-term
revenue visibility. The company’s diversified foundry strategy,
including access to TSMC, offers greater supply stability relative
to peers, while its high compatibility with the CUDA ecosystem
lowers customer migration barriers. In addition, a clear trajectory
toward profitability, supported by operating leverage, and a
product portfolio covering both training and inference workloads
position Iluvatar to capture broader demand across China’s AI
computing market.
Earnings: Our EPS estimates for 2026–27 are and
. We expect strong revenue growth driven by TianGai
shipment ramp and CSP demand, with margin expansion
supported by operating leverage as scale improves. (Note that
Iluvatar's IPO was in Jan 2026, so there are, as yet, no meaningful
consensus estimates.)
Valuation: We derive our price target using a residual income
model. Our price target implies 44x 2026e P/S, which we believe
is justified given the 122% revenue CAGR that we forecast for
2025-28 and an improving profitability profile.
Key downside risks include: 1) delays in TianGai shipment ramp,
2) tightening export controls impacting supply chain access, and
3) intensifying competition from both domestic peers and global
incumbents.
Exhibit 31: Iluvatar: Summary of key metrics
Source:
Company data, FactSet, Morgan Stanley Research. e = Morgan Stanley Research estimates.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
28
Iluvatar: Financial Summary
Exhibit 32: Iluvatar: Financial summary
Income Statement Cash Flow Statement
Rmb mn (Years End Dec ) 2025 2026E 2027E 2028E Rmb mn (Years End Dec ) 2025 2026E 2027E 2028E
Net sales 1, 3, 7, 11, Cashflow from Operations (1,) (1,) (1,) 1,
COGS () (1,) (3,) (5,) Net profits (1,) () 1, 2,
Gross profit 1, 3, 5, Depreciation
Operating expenses (1,) (1,) (2,) (3,) Working Capital Change () (1,) (2,) (1,)
Operating income (1,) () 1, 2, Other adjustments
Non-operating income Cashflow from Investing () () () ()
Pre-tax income (1,) () 1, 2, Capex () () () ()
Income tax Change of LT Investment ()
Minority Interest Change of ST Investment
Reported net Income (1,) () 1, 2, Other adjustments () () () ()
( m) Cashflow from financing 2, 3,
Reported EPS (Rmb) () () Increase in L/T debt
EPS for consensus (Rmb) () () Increase in S/T debt
Cash Dividend Paid
Issuance of stock 3,
Balance Sheet Other adjustments 2,
Rmb mn (Years End Dec ) 2025 2026E 2027E 2028E Exchange rate adjustment ()
Cash 1, 2, 1, 1, Net change in cash 1, 1, (1,)
Mkt Securities
AR/NR 1, 1,
Inventory 1, 3, 3,
Other 1, 2, 3, Financial Ratios
Current Assets 3, 6, 8, 10, 2025 2026E 2027E 2028E
Long-term investments Growth(%)
Fixed assets Revenue
Intangible Assets Operating profits () NA
Other assets Pretax profits () NA
Total Assets 3, 7, 9, 11, Net profits () NA
S/T borrowings EPS () () NA
AP/NP 1, Margins (%)
Other ST liabilities Gross Margin
LT debt Operating Margin () ()
Other LT liabilities Pretax Margin () ()
Total Liabilities 1, 1, 2, 2, Net Profit () ()
Common shares Return (%)
Additional capital 3, 3, 3, ROAE () ()
Retained earning 2, 1, 3, 5, ROAA () ()
Other shareholders' equity () () () () Gearing (%)
Total Equity 2, 5, 6, 9, Net Debt/Equity () () () ()
Total Liab. & Shrhldr's Equity 3, 7, 9, 11, Liabilities/Equity
Ratios (X)
E = Morgan Stanley Research Estimates Current ratio
Source: Morgan Stanley Research, Company Data Quick ratio
Others
AR/NR Turnover (days)
Inventory Turnover (days)
AP Turnover (days)
Cash Conversion (days)
Source: Company data, Morgan Stanley Research. e = Morgan Stanley Research estimates.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 29
Risk Reward - Iluvatar CoreX Semiconductor Co., Ltd.
()
HK$
Key valuation assumptions include:
An % cost of equity (derived from a beta of , risk-free rate of %, and equity risk
premium of %)
A long-term payout ratio of 33%
A medium-term growth rate of 16%
A perpetual terminal growth rate of 6%
▪ Strong order visibility from domestic
CSPs, with TianGai-150 shipments ramping
from 2Q26 and contributing >Rmb4bn
revenue over 2026–27.
▪ Diversified foundry strategy ensures
supply security, with access to TSMC
capacity mitigating risks from export
controls and domestic yield constraints.
▪ High CUDA compatibility enables low-
friction migration, with successful LLM
deployment validating product-market fit.
▪ Clear path to profitability, with breakeven
expected in 2026 and full-year profitability
in 2027 driven by operating leverage and
scaling.
▪ Our price target implies 44x 2026e P/S,
which is lower than the peer average of 75x
2026e P/S.
Risk Reward Themes
New Data Era: Positive
Technology Diffusion: Positive
View descriptions of Risk Rewards Themes here
HK$1,
80x 2026e P/S
We assume: 1) >150% revenue CAGR in
2025-28e driven by a more aggressive
domestic AI infrastructure build-out and
large-scale procurement of both training and
inference chips from Chinese cloud service
providers and state-owned enterprises. 2)
Gross margin improves to over 60% in
2026e and 2027e thanks to production
scale effects, mature technology iteration
and optimized high-value product structure.
3) The company achieves full-year
profitability in 2026, ahead of our base case.
HK$
44x 2026e P/S
We assume: 122% revenue CAGR in 2025-
28e given strong orders from CSP clients
and solid supply chain capacity. 2) Gross
margin at % in 2026e and % 2027e
due to slight price competition. 3) The
company achieve full-year profitability in
2027.
HK$
22x 2026e P/S
We assume: <80% revenue CAGR in 2025-
28e given slower-than-expected domestic AI
capital expenditure and large model
commercialization progress. 2) Nvidia
relaxes export restrictions on mid-range AI
chips to China, leading to intensified price
competition. 3) Gross margin falls below
45% in 2026e and 2027e due to fierce price
competition. 4) The company does not
achieve full-year profitability until 2028.
Risk Reward – Iluvatar CoreX Semiconductor Co., Ltd. ()
Leveraging supply chain resilience with strong order visibility
PRICE TARGET
RISK REWARD CHART
Key: Historical Stock Performance Current Stock Price Price Target
Source: Refinitiv, Morgan Stanley Research
HK$$$
HK$1,$1,(+%)(+%)HK$1,(+%)
HK$$(+%)(+%)HK$(+%)
HK$$(%)(%)HK$(%)
APR '25 OCT '25 APR '26 APR '27
0
250
500
750
1000
HKD
OVERWEIGHT THESIS
BULL CASE BASE CASE BEAR CASE
若有侵权，请联系删除。公众号：趣享财经
M Foundation
30
Drivers 2025 2026e 2027e 2028e
Operating profit (YoY) (%) () ()
CSP AI capex growth
TianGai shipment ramp
CUDA migration trend
Supply chain advantage (TSMC)
AI inference expansion
Policy support
Customer base expansion
Margin improvement trajectory
100% Mainland China
Source: Morgan Stanley Research Estimate
View explanation of regional hierarchies here
Stronger-than-expected CSP orders
Faster CUDA replacement with Iluvatar's
software
Expansion of overseas and domestic capacity
Order ramp below expectations
Escalation of sanctions
Intensifying competition
Morgan Stanley EstimatesMean
Source: Refinitiv, Morgan Stanley Research
FY 2026e
Sales /
Revenue
(Rmb, mn)
Note: There are not sufficient brokers supplying
consensus data for this metric
3,060
Net income
(Rmb, mn) Note: There are not sufficient brokers supplying
consensus data for this metric
(262)
EPS
(Rmb) Note: There are not sufficient brokers supplying
consensus data for this metric
()
Risk Reward – Iluvatar CoreX Semiconductor Co., Ltd. ()
KEY EARNINGS INPUTS
INVESTMENT DRIVERS
GLOBAL REVENUE EXPOSURE
RISKS TO PT/RATING
RISKS TO UPSIDE
RISKS TO DOWNSIDE
MS ESTIMATES VS. CONSENSUS
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 31
Iluvatar: Investment Positives & Concerns
Positives
Upcoming order inflection from domestic cloud providers. According to our industry
checks, major Chinese cloud service providers placed significant preorders for Iluvatar's
TianGai-150 AI chips, with shipments set to begin in Q2 2026 and ramp sharply in 2H26.
We estimate these orders will contribute over Rmb4bn to Iluvatar's 2026 and 2027
revenue.
Industry-leading supply chain security. Unlike domestic peers that rely exclusively on
SMIC for advanced nodes, Iluvatar maintains a diversified foundry strategy and produces
fully compliant AI chips at TSMC. This provides greater capacity certainty and eliminates
production-disruption risk from further export restrictions.
Superior software compatibility and migration capabilities. Iluvatar's GPGPU chips offer
near-seamless compatibility with NVIDIA's CUDA ecosystem, significantly reducing
customer migration costs. The company has helped clients to migrate LLM stacks from
NVIDIA platforms to TianGai-150.
Clear and imminent profitability inflection. We expect Iluvatar is on track to achieve
positive adjusted net income in Q3 2026 and full-year profitability in 2027. Strong
revenue growth with operating leverage should drive this outcome, ., R&D and SG&A as
a percentage of revenue decline further in 2026-2027.
Comprehensive product portfolio covering training and inference. Iluvatar also offers
edge AI inference products, the Zhikai series, enabling it to capture broader customer
demand and provide end-to-end AI computing solutions.
Concerns
TianGai-150 order execution and Tiangai-300 validation risk. Approximately 65% of our
2026 revenue forecast depends on successful TianGai-150 deployments. Also, future
growth depends on the Tiangai-300 product. Rigorous 3–6 month customer validation
and TSMC CoWoS packaging constraints could delay order placement and revenue
recognition.
TSMC supply chain remains vulnerable to US export controls. While Iluvatar's TSMC
partnership provides a competitive advantage, further US restrictions on advanced node
manufacturing for Chinese companies would severely impact operations.
Profitability inflection could be delayed. Aggressive R&D spending on TianXuanand
TianJi architectures and intensifying price competition could push full-year profitability
beyond 2027.
Domestic competition intensifies across segments. Major competitors – including
Huawei Ascend 950/960, Cambricon MLU580/690, and MetaX's upcoming C600 – will
directly compete for cloud and government orders. We estimate Iluvatar's top line to
reach by 2027, but aggressive peer pricing could limit upside.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
32
Iluvatar: Earnings Estimates
We believe Iluvatar's 2026 revenue will be driven by strong demand for its TianGai-150
chip, with supply constraints far less severe than for domestic peers due to diversified
foundry partnerships. We forecast revenue of in 2026, up 196% YoY. As the
TianGai-300 chip enters full-scale production in 2H26 and mass shipment in 2027, we
project revenue of in 2027, and in 2028, implying a 122% CAGR over
2025-28.
Turning to profitability, we expect positive net income in 2H26 and full-year profitability in
2027. We project net income attributable to shareholders of -Rmb262mn in 2026,
in 2027, and in 2028, implying net margins of %, %, and
%, respectively.
Exhibit 33: Iluvatar: Earnings estimates
(Rmb mn) 1H25 2H25 1H26E 2H26E 1H27E 2H27E 1H28E 2H28E 2025 2026E 2027E 2028E
Total revenues 2, 3, 4, 5, 6, 1, 3, 7, 11,
H/H Change % % % % % % % %
Y/Y Change % % % % % % % % % % % %
Cost of Sales () () () () (1,) (2,) (2,) (3,) () (1,) (3,) (5,)
Percent of Revenues 50% 44% 50% 48% 49% 49% 49% 49% 46% 49% 49% 49%
Gross Profit 1, 1, 2, 2, 3, 1, 3, 5,
Gross Margin % % % % % % % % % % % %
Incremental Margin % % % % % % % % % % % %
Total Opex () () () () (1,) (1,) (1,) (1,) (1,) (1,) (2,) (3,)
Percent of Revenues % % % % % % % % % % % %
R&D () () () () () () () () () (1,) (1,) (1,)
Percent of Revenues % % % % % % % % % % % %
General & Adm Exp. () () () () () () () () () () () ()
Percent of Revenues % % % % % % % % % % % %
Selling Expenses () () () () () () () () () () () ()
Percent of Revenues % % % % % % % % % % % %
Operating Income () () () 1, 1, 1, (1,) () 1, 2,
Operating Margin % % % % % % % % % % % %
Total Non-operating Income (loss)
Profit Before Taxes () () () 1, 1, 1, (1,) () 1, 2,
Percent of Revenues % % % % % % % % % % % %
Taxes
Tax Rate % % % % % % % % % % % %
Total Net Income to Parent () () () 1, 1, 1, (1,) () 1, 2,
Percent of Revenues % % % % % % % % % % % %
EPS for consensus (Rmb) () () () () ()
Source: Company data, Morgan Stanley Research estimates
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 33
Iluvatar: Valuation
We set our price target at HK$600. Given Iluvatar’s near profitability and high-growth
phase, we value the company with a residual income model, which we believe
appropriately balances near-term growth momentum with long-term intrinsic value.
Key valuation assumptions include:
• An % cost of equity (derived from a beta of , risk-free rate of %, and
equity risk premium of %)
• A long-term payout ratio of 33%
• A medium-term growth rate of 16%
• A perpetual terminal growth rate of 6%
Our price target implies a 44x 2026e P/S multiple.
Exhibit 34: Iluvatar: Residual income model
Rmb million 2026E 2027E 2028E 2029E 2030E 2031E 2032E 2033E 2034E 2035E 2036E 2037E
Total Equity 5,432 6,881 9,451 11,463 13,796 16,504 19,644 23,287 27,513 32,414 38,101 44,697
Net Profit (262) 1,449 2,570 2,981 3,458 4,011 4,652 5,397 6,260 7,262 8,424 9,772
ROAE NM % % % % % % % % % % %
Residual Income 828 1,594 1,909 2,187 2,507 2,878 3,308 3,807 4,384 5,054 5,831
Spread % % % % % % % % % % %
Ending Equity Capital 5,432
PV of Forecast Period 17,004
PV of Continuing Value 111,790
Equity Value 134,226
No. of Shares 254
HKD/RMB
Price Target (HK$) 600
Source: Company data, Morgan Stanley Research estimates
Peer comparison
We expect Iluvatar’s revenue to grow at a 122% CAGR from 2025 to 2028, quite high
among domestic peers. We attribute this superior earnings and revenue growth to:
• Iluvatar’s diversified supply chain strategy, allowing use of both domestic and
international foundries, resulting in significantly fewer capacity constraints than
competitors
• Leading software stack compatibility and Prefill and Decode separation
technology, significantly reducing customer migration costs and improving
inference efficiency.
In our view, superior top-line momentum, a leading position in China’s domestic general-
purpose GPU segment, and improving revenue visibility from long-term framework
adaptation and customer lock-in support a premium multiple of 44x 2026e P/S.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
34
Exhibit 35: China AI semi peer comparison table
Source: Company data, FactSet, Morgan Stanley Research
Exhibit 36: Iluvatar P/S multiple
0x
5x
10x
15x
20x
25x
26-01
Iluvatar one-year forward P/S
Source: Company data, FactSet, Morgan Stanley Research
Bull & bear cases
Our bull case value is HK$1,100, and assumes accelerated domestic substitution with
faster-than-expected product performance and market share gains
We assume:
• A >150% revenue CAGR in 2025-28, driven by more aggressive domestic AI
infrastructure build-out and large-scale procurement of training and inference
chips from Chinese cloud service providers and state-owned enterprises.
• Gross margin rises to over 60% in 2026 and 2027, driven by scale, mature
technology iteration, and an optimized high-value product mix.
• Full-year profitability in 2026, ahead of 2027 in our base case.
Our bear case value is HK$300, and assumes slower revenue growth with gross margin
compression amid intensifying competition and delayed product launches.
We assume:
• A <80% revenue CAGR in 2025-28 given slower-than-expected domestic AI capital
expenditure and large model commercialization.
• Nvidia relaxes export restrictions on mid-range AI chips to China, intensifying price
competition.
• Gross margin falls below 45% in 2026 and 2027 due to fierce price competition,
rising wafer fabrication costs, and underutilized capacity.
• Full-year profitability not achieved until 2028, one year later than our base case.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 35
MetaX (): Positioning for scalability through
software and supply; Initiate at EW
MetaX Integrated Circuits is a domestic GPU designer focused on
high-performance training chips, with a growing presence in
China’s AI computing market. We initiate coverage with an Equal-
weight rating, as we see balanced risk-reward driven by solid
execution in software compatibility and supply chain, alongside
uncertainties in scale-up and customer diversification. The
company’s proprietary MXMACA platform offers high
compatibility with NVIDIA’s CUDA ecosystem, lowering switching
costs and supporting gradual customer adoption. In addition,
MetaX has secured relatively stable demand from government
and enterprise customers, which tends to carry higher margins
than CSP orders. Its production on SMIC’s N+1 process provides
supply chain visibility and yield stability. That said, further
validation in large-scale deployments and broader customer
penetration remain key to sustaining growth.
Earnings: Our EPS estimates for 2026–27 are and
. We expect revenue growth to be supported by
continued C500 shipments and initial C600 ramp, with margin
trends dependent on product mix and scale effects. (Note that
MetaX's IPO was in Dec 2025, so there are, as yet, no meaningful
consensus estimates.)
Valuation: We derive our price target using a residual income
model. Our price target implies a 75x 2026e P/S multiple, which
we believe reflects the 66% revenue CAGR that we project for
2025–28, balanced against execution risks and competitive
pressures.
Key downside risks include: 1) slower-than-expected C600
commercialization, 2) customer concentration and demand
variability, and 3) intensifying competition in the domestic GPU
market.
Exhibit 37: MetaX: Summary of key metrics
Source:
Company data, FactSet, Morgan Stanley Research. e = Morgan Stanley Research estimates.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
36
MetaX: Financial Summary
Exhibit 38: MetaX: Financial summary
Source: Company data, Morgan Stanley Research. e = Morgan Stanley Research estimates.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 37
Risk Reward - MetaX Integrated Circuits ()
Key valuation assumptions include: an % cost of equity (derived from a beta of , risk-
free rate of %, and equity risk premium of %), a long-term target payout ratio of 50%,
a medium-term CAGR of 18% through 2038e, and a perpetual terminal growth rate of 6%.
▪ Strong CUDA compatibility via MXMACA,
lowering migration barriers and supporting
faster adoption among existing NVIDIA
users.
▪ Stable government and enterprise orders
provide revenue visibility, with higher-margin
projects supporting near-term gross margin
profile.
▪ Mature domestic supply chain with SMIC
N+1 production, ensuring capacity
availability and relatively stable yields
versus peers.
▪ Our target price implies 75x 2026e P/S,
which is higher than the peer average.
Risk Reward Themes
New Data Era: Positive
Technology Diffusion: Positive
View descriptions of Risk Rewards Themes here
Rmb1,
149x 2026e P/S
We assume (1) >100% revenue CAGR in
2025-28e on the back of faster-than-
expected mass production and delivery of
flagship GPGPUs, as well as booming
demand for domestic general computing and
AI training chips; (2) significant market share
expansion in China’s domestic AI semi
market, with full adaptation to mainstream
AI frameworks and soverign customers; (3)
gross margin improves to over 60% in
2026e.
75x 2026e P/S
We assume (1) 66% revenue CAGR in 2025-
28e on the back of mass production and
delivery of C600 and C700 AI GPU; (2)
Market share expansion in China’s domestic
AI semi market, with rising orders from
soverign customers and entry into core
CSPs; (3) gross margin reach 54% in 2026e.
38x 2026e P/S
We assume (1) <40% revenue CAGR in
2025-28e given delayed mass production
schedule of flagship GPU products, slower-
than-expected customer validation and
weaker domestic AI computing capital
expenditure. (2) Share loss in China’s AI semi
market amid intensifying domestic and
overseas competition. (3) Gross margin falls
below 45% in 2026e due to high R&D, rising
wafer and packaging costs, and fierce price
competition in the market.
Risk Reward – MetaX Integrated Circuits ()
Positioning for scalability through software and supply
PRICE TARGET
RISK REWARD CHART
Key: Historical Stock Performance Current Stock Price Price Target
Source: Refinitiv, Morgan Stanley Research
Rmb1,,(+%)(+%)Rmb1,(+%)
(+%)(+%)(+%)
(%)(%)(%)
APR '25 OCT '25 APR '26 APR '27
0
400
800
1200
1600
CNY
EQUAL-WEIGHT THESIS
BULL CASE BASE CASE BEAR CASE
若有侵权，请联系删除。公众号：趣享财经
M Foundation
38
Drivers 2025 2026e 2027e 2028e
Operating profit (YoY) (%) () () ()
CUDA migration trend
Government AI capex
C500 shipment ramp and C600
commercialization
Domestic substitution tailwind
Customer expansion to CSP clients
Yield and cost improvement
100% Mainland China
Source: Morgan Stanley Research Estimate
View explanation of regional hierarchies here
Faster CUDA migration
Stronger government and CSP orders
Better-than-expected C600 ramp
Product concentration risk
Weaker-than-expected demand
Intensifying competition
Morgan Stanley EstimatesMean
Source: Refinitiv, Morgan Stanley Research
FY 2026e
Sales /
Revenue
(Rmb, mn)
Note: There are not sufficient brokers supplying
consensus data for this metric
4,
Net income
(Rmb, mn) Note: There are not sufficient brokers supplying
consensus data for this metric
22
EPS
(Rmb) Note: There are not sufficient brokers supplying
consensus data for this metric
Risk Reward – MetaX Integrated Circuits ()
KEY EARNINGS INPUTS
INVESTMENT DRIVERS
GLOBAL REVENUE EXPOSURE
RISKS TO PT/RATING
RISKS TO UPSIDE
RISKS TO DOWNSIDE
MS ESTIMATES VS. CONSENSUS
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 39
MetaX: Investment Positives & Concerns
Positives
Industry-leading software compatibility with NVIDIA CUDA. MetaX's proprietary
MXMACA software stack offers near-seamless CUDA compatibility, supporting over
6,000 CUDA APIs and enabling most applications to migrate to MetaX hardware. This
could significantly reduce switching costs and accelerates adoption.
Stable government and enterprise orders. With a strong government investment
background, MetaX has secured contracts with multiple national and provincial intelligent
computing centers, delivering higher gross margins (over 60%) vs. cloud service provider
orders.
Fully operational domestic supply chain with reasonable yields. Unlike peers still
transitioning to domestic manufacturing, MetaX has ramped production on SMIC's N+1
process with sufficient yields, providing significant capacity advantages and supply chain
security.
Focused strategy on training chips creates differentiation. While most domestic players
compete across training and inference, MetaX has concentrated resources on high-
performance training chips, enabling faster technology iteration and better product-
market fit.
Concerns
Slow CSP penetration limits revenue upside. MetaX has secured major orders from top
Chinese CSPs, such as Bytedance, yet it lags peers. Failure to gain meaningful share in this
largest, fastest-growing segment would cap long-term growth.
Process node trade-off undermines core performance. MetaX’s flagship GPUs use
SMIC’s N+1 process, a choice that delivers stable mass production and high yield, but
carries structural limitations in transistor density and power efficiency vs. more advanced
industry-standard nodes.
Over-reliance on SMIC creates supply chain concentration. The company now depends
almost entirely on SMIC's N+1 process. Any disruption to SMIC's operations or capacity
constraints could materially impact order fulfillment.
Competitive pressure weighs on market share. China’s domestic GPGPU market has
entered a hyper-competitive phase, with over 10 mass-production vendors triggering
significant price-based competition at the mid-to-low-end, while rising expectations of a
marginal relaxation of US GPU export controls to China pose additional downside risk.
若有侵权，请联系删除。公众号：趣享财经
Foundation
ture
GPU
nversion
the
t the
wed by
2027E 2028E
6, 7,
% %
2, 3,
47% 47%
3, 3,
% %
% %
2, 3,
% %
1, 2,
% %
% %
% %
% %
1,
% %
% %
1,
% %
M
MetaX: Earnings Estimates
We think revenue momentum will stay strong in 2026-28, driven by C600 and fu
product mass production, growing orders for the company’s high-performance GP
products, full-stack adaptation to mainstream AI frameworks, and faster order co
from core clients, including cloud service providers and sovereign platforms amid
ongoing domestic expansion cycle.
Our 2026/27/28 revenue growth forecasts are 145%/54%/21%, and we also expec
company to turn profitable for the first time in 2026, with EPS of , follo
EPS in 2027.
Exhibit 39: MetaX: Earnings estimates
(Rmb mn) 1Q25 2Q25 3Q25 4Q25 1Q26E 2Q26E 3Q26E 4Q26E 2025 2026E
Total revenues 1, 1, 1, 4,
Q/Q Change % % % % % % % %
Y/Y Change % % % % % % % %
Cost of Sales 1,
Percent of Revenues 45% 45% 45% 46% 46% 46% 46% 46% 45% 46%
Gross Profit 2,
Gross Margin % % % % % % % % % %
Incremental Margin NM % NM % % % % % % %
Total Opex 1, 2,
Percent of Revenues % % % % % % % % % %
R&D 1, 1,
Percent of Revenues % % % % % % % % % %
General & Adm Exp.
Percent of Revenues % % % % % % % % % %
Selling Expenses
Percent of Revenues % % % % % % % % % %
Operating Income () () () () () () () ()
Operating Margin % % % % % % % % % %
Total Non-operating Income (loss) () () ()
Profit Before Taxes () () () () () ()
Percent of Revenues % % % % % % % % % %
Taxes
Tax Rate % % % % % % % % % %
Total Net Income to Parent () () () () () ()
Percent of Revenues % % % % % % % % % %
EPS for consensus (Rmb) () () () () () ()
Source: Company data, Morgan Stanley Research estimates
40
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 41
MetaX: Valuation
Our price target is Rmb758, with an Equal-weight rating. As MetaX is moving from a pre-
profit development stage to sustained, scalable profitability, we also value the company
using a Residual Income Model. Unlike relative multiples with limited applicability at
profitability inflection, we believe this framework best captures intrinsic long-term value,
fully incorporating our earnings forecasts through breakeven and the subsequent profit
ramp-up amid the booming domestic substitution cycle for high-performance GPGPU
chips.
Key assumptions: a % cost of equity (beta , risk-free rate %, equity risk premium
%), a long-term target payout ratio of 50%, a medium-term CAGR of 18% through
2038, and a perpetual terminal growth rate of 6%.
Our price target implies a 75x 2026e P/S multiple, supported by accelerating mass
commercialization, full-stack software ecosystem competitiveness, and improving long-
term revenue visibility as it scales through the profitability inflection. As the stock trades
over 70x 2026e P/S now, we think MetaX is less attractive than Cambricon and Iluvatar.
Exhibit 40: MetaX: Residual income valuation
Rmb million 2026E 2027E 2028E 2029E 2030E 2031E 2032E 2033E 2034E 2035E 2036E 2037E
Total Equity 13,173 13,706 14,706 15,297 15,994 16,816 17,786 18,930 20,281 21,875 23,755 25,975
Net Profit 22 533 1,001 1,181 1,393 1,644 1,940 2,289 2,701 3,188 3,761 4,438
ROAE % % % % % % % % % % % %
Residual Income (833) (334) 75 202 368 563 793 1,062 1,378 1,749 2,184 2,696
Spread % % % % % % % % % % % %
Ending Equity Capital 13,173
PV of Forecast Period 3,998
PV of Continuing Value 285,924
Equity Value 303,095
No. of Shares 400
Price Target 758
Source: Company data, Morgan Stanley Research estimates
Bull & bear cases
Our bull case value is Rmb1,500, and assumes robust mass production delivery with rapid
customer penetration and gross margin expansion.
We assume (1) >100% revenue CAGR in 2025-28 on faster-than-expected mass production
and delivery of flagship GPGPUs, plus booming demand for domestic general computing
and AI training chips; (2) significant market share expansion in China’s domestic GPGPU
market, with full adaptation to mainstream AI frameworks and successful entry into core
CSPs and sovereign customers; (3) gross margin above 60% in 2026, driven by mass
production scale effects, optimized supply chain management, and an upgraded high-value
product mix.
Our bear case value is Rmb380, and assumes lower product commercialization with
revenue miss and gross margin contraction.
We assume (1) <40% revenue CAGR in 2025-28 given delayed mass production of flagship
GPU products, slower-than-expected customer validation, and weaker domestic AI
computing capex; (2) share loss in China’s general-purpose GPU market amid intensifying
domestic and overseas competition, failure to complete full adaptation to mainstream AI
frameworks, and inability to secure mass procurement orders from top-tier customers;
若有侵权，请联系删除。公众号：趣享财经
M Foundation
42
and (3) gross margin below 40% in 2026 due to high R&D, rising wafer and packaging
costs, and fierce price competition.
Peer comparison
We expect MetaX's revenue to increase at a 66% CAGR from 2025 to 2028. We attribute
the superior top-line growth to booming AI computing capex from core clients, including
national supercomputing centers and potentially CSPs, as well as accelerated mass
production, full-stack software ecosystem adaptation, and customer validation of flagship
GPU products amid ongoing domestic penetration. In our view, although MetaX could
experience faster revenue expansion given core customer penetration, MetaX's premium
P/S multiple at 75x 2026e looks too high for us vs. peers.
Exhibit 41: China AI semi peer comparison table
Source: Company data, FactSet, Morgan Stanley Research
Exhibit 42: MetaX P/S multiple
40x
45x
50x
55x
60x
65x
70x
75x
25-12 26-01 26-01 26-01 26-02 26-02 26-03 26-03 26-04 26-04
MetaX one-year forward P/S
Source: Company data, FactSet, Morgan Stanley Research
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 43
Baidu & Alibaba
Covered by Gary Yu & Joanne Lau
For detailed analysis, please refer to our deep-dive report China's AI Path: Owning the Full
AI Stack via In-house Chips.
T-Head (Alibaba)
T Head Semiconductor (Pingtouge) was founded in September 2018 as a wholly owned
chip business under Alibaba. It originated as a spin off from Alibaba’s DAMO Academy,
with a strategic mandate to build custom ASIC products and support China’s
semiconductor self sufficiency. Alibaba is reportedly preparing T Head for a potential
IPO, and it plans to restructure T-Head into a standalone business with partial employee
ownership ahead of commencing any listing process.
T Head produces both GPUs and CPUs, with CPU products remaining a core part of the
portfolio.
Exhibit 43: T-head key products
Product 真武 810E (Zhenwu
810E PPU)
含光 800 (Hanguang
800)
倚天 710 (Yitian 710) 镇岳 510 (Zhenyue 510) 玄铁 C910 (Xuantie
C910)
玄铁 C920 (Xuantie
C920)
羽阵 611 (Yuzhen 611 /
TH6101)
Category / Use Data-center AI
accelerator (training +
inference)
Data-center AI inference
chip
Cloud server CPU Enterprise SSD
controller
RISC-V CPU IP core
(high-perf)
RISC-V CPU IP core w/
vector
UHF RFID tag chip
Compute N/A Peak compute disclosed
as 820 TOPS
N/A N/A Supports FP16 / BF16 /
FP32 / FP64 +
INT8/16/32/64
Vector unit supports
FP16/FP32/FP64 +
INT8/16/32/64
N/A
Memory / Storage 96GB HBM2e Not publicly disclosed on
the product page
Uses system DRAM;
supports 8-channel
DDR5
Supports TLC/QLC
NAND (1xxL/2xxL),
NVMe features
N/A (IP core) N/A (IP core) Memory disclosed: 128-
bit EPC, 96-bit TID (incl.
48-bit SN), 32-bit
access/kill shared
I/O / Interconnect Inter-chip bandwidth
700GB/s
Not publicly disclosed 96-lane PCIe PCIe Gen5 x4, NVMe
, ZNS
N/A (IP core) N/A (IP core) RFID air interface;
protocol compliant with
EPC Global G2 V2 /
ISO/IEC 18000-6C
Other key disclosed specsReported to be
positioned as
comparable to NVIDIA
H20
12nm, 17B transistors Armv9 compatible, 60B
transistors,
SPECint2017 score
disclosed
Supports enterprise
features like multi-
stream / atomic write /
IO virtualization; E2E
data protection; hot-plug
Vector units designed
per RVV
RV64GCV; high-perf
core with vector
computing ability
Sensitivity: read -24
dBm / write -20 dBm;
operating temp -40°C to
+85°C
Source: Company data, Morgan stanley Research
Valuation
In its latest results call, management indicated >60% of T-head revenue is from external
cloud customers, and has delivered >470k units so far. Management did not rule out a
potential spin-off opportunity, which could further unlock SOTP value, but no timeline
was noted. On both the demand and supply side, support remains strong: AliCloud
continues to drive enormous training and inference needs, and increasing production
capacity supply.
We value T-head at a valuation range of US$28-83bn, based on 12-24x F28e P/S, with
reference to key peers Cambricon and Iluvatar, applied an estimated revenue range of
Rmb16-24bn (CPU + GPU). This translates to US$8-24/share in BABA’s SOTP assuming a
30% holdco discount.
We revised midrange SOTP to US$250 (from US$245), T-head at US$15/share (from US
$22), implying T-head market cap of US$52bn, based on 18x F28e EV/sales:
• This is based on Rmb20bn F28e revenue (the midpoint of our estimated F28e
revenue range of Rmb$16-24bn).
• Multiple is at a 10% discount to key A-share peer Cambricon's 20x 2027e EV/sales
given A/H share discount and similar to smaller H-share peer Iluvatar’s 18x 2027e.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
44
We also applied a 30% holdco discount which is consistent with broad Internet
peers.
Exhibit 44: T-head market cap (US$ mn)
51,799 16,000 18,000 20,000 22,000 24,000
24x 55,252 62,158 69,065 75,971 82,878
21x 48,345 54,388 60,432 66,475 72,518
18x 41,439 46,619 51,799 56,978 62,158
15x 34,532 38,849 43,165 47,482 51,799
12x 27,626 31,079 34,532 37,986 41,439
P
/S
(
x
)
T-head F2028 revenue (Rmbm)
Source: Company data, Morgan Stanley Research estimates
Exhibit 45: T-head market cap to BABA (US$ mn)
36,259 16,000 18,000 20,000 22,000 24,000
24x 38,676 43,511 48,345 53,180 58,014
21x 33,842 38,072 42,302 46,532 50,763
18x 29,007 32,633 36,259 39,885 43,511
15x 24,173 27,194 30,216 33,237 36,259
12x 19,338 21,755 24,173 26,590 29,007
T-head F2028 revenue (Rmbm)
P
/S
(
x
)
Source: Company data, Morgan Stanley Research estimates
Exhibit 46: T-head market cap to BABA/share (US$)
15 16,000 18,000 20,000 22,000 24,000
24x
21x
18x
15x
12x
T-head F2028 revenue (Rmbm)
P
/S
(
x
)
Source: Company data, Morgan Stanley Research estimates
Exhibit 47: BABA mid point SOTP
US$ 90
8x F27e
EV/EBITA
US$ 100
F28e EV/S
US$ 15
F27e
EV/S
US$ 15
18x F28 EV/S
-
US$ 14
@30% discount
US$ 16
@30% discount
Mid-point US250/share SOTP (US/share)
E-commerce
Cloud
AIDC
T-head
Net cash/(debt)
Investment value
Source: Company data, Morgan Stanley Research estimates
Kunlunxin (Baidu)
Kunlunxin is Baidu’s semiconductor subsidiary. Baidu began chip development in 2011, and
in April 2021 spun the project out as a separate entity led by its chief chip architect,
Ouyang Jian. Kunlunxin is CUDA compatible, enabling developers to migrate from NVIDIA
with minimal changes. Baidu has a strategic partnership with Samsung, which
manufactured the first Kunlun chip in 2018 using 14nm. We expect both Samsung and
SMIC to remain key suppliers as domestic demand grows.
Kunlunxin focuses on producing ASICs and it is CUDA-compatible.
Exhibit 48: Kunlunxin key products
Source: Company data, Morgan Stanley Research
Valuation
Baidu proposed the spin-off of KLX in December 2025 while we expect on track IPO
progress with an expected listing timeline by late 2Q to early 3Q. The proposed listing is
part of management's plan to unlock shareholder value.
We value Kunlunxin at a range of US$20-62bn, assuming 2027 revenue range of Rmb12-
18bn, using P/S multiples range of 12-24x with reference to key peers Cambricon and
Iluvatar.
若有侵权，请联系删除。公众号：趣享财经
M Foundation
Morgan Stanley Research 45
We revised midrange SOTP to US$225 (from US$215), KLX at US$47/share (from US$45),
implying market cap of US$, based on 18x 2027e EV/sales:
• This is based on Rmb15bn 2027e revenue (the midpoint of our estimated 2027
revenue range of Rmb$12-18bn).
• Multiple is at a 10% discount to key A-share peer Cambricon's 20x 2027e EV/sales
given A/H share discount and similar to smaller H-share peer Iluvatar’s 18x 2027e.
We also applied a 30% holdco discount which is consistent with broad Internet
peers.
Exhibit 49: Kunlunxin market cap (USbn)
38,849 12,000 13,500 15,000 16,500 18,000
24x 41,439 46,619 51,799 56,978 62,158
21x 36,259 40,791 45,324 49,856 54,388
18x 31,079 34,964 38,849 42,734 46,619
15x 25,899 29,137 32,374 35,612 38,849
12x 20,719 23,309 25,899 28,489 31,079
Kunlunxin 2027 revenue (Rmbm)
P
/S
(
x
)
Source: Company data, Morgan Stanley Research estimates
Exhibit 50: Kunlunxin market cap to Baidu (USbn)
16,167 12,000 13,500 15,000 16,500 18,000
24x 17,245 19,401 21,556 23,712 25,867
21x 15,089 16,975 18,862 20,748 22,634
18x 12,934 14,550 16,167 17,784 19,401
15x 10,778 12,125 13,473 14,820 16,167
12x 8,622
