Report Analysis: TSMC's AI Revenue to Double by 2027, CoWoS Capacity Remains a Bottleneck

marsbitPublished on 2026-06-25Last updated on 2026-06-25

Abstract

Morgan Stanley's report on June 23 forecasts explosive growth for TSMC's AI-related revenue. It predicts revenue will surge to $86.3 billion in 2027, a 218% increase from 2026's $27.1 billion. While GPUs (led by Nvidia) remain the primary driver, significant new demand is emerging from AMD's Venice CPUs and MI400 GPUs, and Google's TPUs. A critical bottleneck is CoWoS advanced packaging capacity. Global demand for CoWoS is projected to reach 2.694 million units in 2027, a 93% year-on-year increase. Even with TSMC's planned expansion to 200k wafers/month and non-TSMC capacity, a shortage is expected, particularly for high-end CoWoS-L used by Nvidia. This scarcity grants TSMC pricing power. Key near-term catalysts include improved ABF substrate supply, rising CPU demand, and Nvidia's Rubin Ultra production. The report highlights winners across the supply chain, including MediaTek (Google TPU partner) and ASE Group.

Author: Rita

TideResearch Insights

Morgan Stanley released an in-depth supply chain report on TSMC on June 23. The core conclusion: based on the latest supply chain survey, the forecast for global CoWoS advanced packaging demand in 2027 has been revised upwards. TSMC's AI-related revenue is projected to reach $86.3 billion in 2027, a 218% increase from $27.1 billion in 2026. Nvidia remains the core driver, but AMD's CPUs and Google's TPUs are becoming new growth engines. Most critically, even if TSMC expands its CoWoS capacity to 200,000 units per month, it may still fail to meet the global demand gap of 2.69 million units.

TSMC's AI Revenue Experiences Explosive Growth

Numbers speak loudest. Morgan Stanley predicts TSMC's AI-related revenue will hit $86.3 billion in 2027, compared to $27.1 billion in 2026, representing 218% growth. This is an exponential leap, not linear growth.

Breaking down the $86.3 billion, GPU revenue accounts for $28 billion, custom AI chip revenue for $18 billion, CoWoS advanced packaging revenue for $40 billion, and AI server CPU revenue for $0.3 billion. TSMC's role in the AI chip industry chain is not just wafer manufacturing; advanced packaging has become an equally important revenue stream.

By 2028, this number is expected to climb further to $106.6 billion. Within three years, TSMC's AI business has nearly tripled in scale. The construction of AI infrastructure is far from saturated; TSMC, as the "shovel seller," is still in the capacity ramp-up phase.

But can capacity really keep up?

Nvidia's demand for TSMC's CoWoS capacity remains the absolute main force. Morgan Stanley expects Nvidia's Rubin and Blackwell GPUs, along with its newly launched Vera CPU, will consume a total of 1.222 million CoWoS units in 2027, a 57% year-on-year increase.

What truly surprised Morgan Stanley is AMD. AMD's total CoWoS consumption is projected to surge 308%, from 120,000 units in 2026 to 530,000 units in 2027. Driving this growth is AMD's Venice CPU and MI400 series GPUs fully supporting agentic AI. The CPU market is being accelerated by AI penetration, not just the GPU market.

This shift is profound. Last year, market discussions focused on GPU dominance, but this year Morgan Stanley clearly sees data centers expanding procurement of both GPUs and CPUs. Nvidia's CPUs are also consuming capacity, while AMD's CPU consumption has surged significantly. The demand for CPU compute power on the AI inference side far exceeds expectations.

Google's TPU Quietly Becomes the Second Largest Consumer

Beyond Nvidia and AMD, Morgan Stanley particularly highlighted the demand from Google's TPUs. Google sources CoWoS advanced packaging through two channels: one is MediaTek as a design service partner, and the other is involvement from Broadcom. Both companies design and manufacture TPU chips for Google, requiring substantial CoWoS capacity.

Morgan Stanley believes that if the supply of ABF substrates improves, MediaTek's procurement volume of TPUs for Google has significant upside potential, and current forecasts might be conservative. Google's ambitions in AI chips have not yet been fully unleashed.

The CoWoS Capacity Gap Is Unbridgeable

Now, back to the most critical question: capacity.

Global CoWoS demand is 1.394 million units in 2026, projected to jump to 2.694 million units in 2027, a 93% increase. TSMC plans to expand its CoWoS capacity to 200,000 units per month by the end of 2027. Non-TSMC capacity is also expected to expand to 80,000 units per month. Combined, global capacity will be about 280,000 units per month, equivalent to an annual capacity of 3.36 million units.

This seems sufficient on the surface, but the problem lies in that the 2.694 million units is the estimated global demand, and Morgan Stanley's survey may not have fully captured all demand signals. Furthermore, the distribution between high-end CoWoS-L and CoWoS-S is crucial. The most advanced CoWoS-L, which Nvidia requires, is extremely tight, and this happens to be TSMC's strength.

Although total capacity appears adequate, TSMC remains in a state of supply shortage at the highest end of advanced packaging. This gives TSMC pricing power and explains Morgan Stanley's Overweight rating on the company.

Emerging Demand Catalysts Persist

Morgan Stanley listed three near-term catalysts. First, improvements in ABF substrate supply, particularly the release of T-Glass capacity procured by MediaTek, will directly boost Google TPU shipments. Second, validation of emerging CPU demand; Nvidia's Vera and AMD's Venice are beginning volume shipments, continuously driving CoWoS consumption. Third, the mass production of Nvidia's next-generation Rubin Ultra product, expected to see significant shipments in the second half of the year.

These catalysts linked together mean that TSMC's CoWoS business will not lack orders in 2027; the key is whether capacity can keep up. From this perspective, TSMC's capital expenditure cycle is far from over.

New Winners in the Supply Chain

In the report, Morgan Stanley specifically highlighted several companies to watch. MediaTek was listed as a top pick because it is Google's primary design partner for TPUs, directly benefiting from AI demand growth. ASE Group and KYEC were also reiterated with Overweight ratings, serving AMD's Venice CPU supply chain and Nvidia's GPU supply chain, respectively.

TSMC itself remains the core beneficiary, but Morgan Stanley's view is that the entire AI supply chain is benefiting, not just the chip design end.

TSMC's AI revenue growth is indeed astonishing; doubling to $86.3 billion by 2027 is not a fantasy. However, this growth is premised on the capacity actually being built, and, most crucially, that advanced packaging capacity does not become a new bottleneck. Morgan Stanley believes it will not, but also clearly points out that supply chain differentiation is intensifying, and the line between winners and losers is being redrawn.

Disclaimer

This article is TideResearch's collation and interpretation of a third-party brokerage research report. The ratings, target prices, earnings forecasts, and related judgments cited herein are the views of Morgan Stanley analysts, representing only the position of their institution, not TideResearch's views, and do not constitute any investment advice.

Please note three points while reading: First, target prices are analysts' expectations for the next approximately 12 months, representing forecasts, not promises, and are subject to frequent adjustments based on performance and market conditions. Second, sell-side research reports are inherently bullish, and some covered companies may have investment banking relationships with the brokerage. Third, the value of a research report lies in its core logic and underlying assumptions, not just a specific target price. Focus on the logic, not just the price.

The market carries risks; decisions should be made independently. This article should not be used as a basis for buying or selling any securities.

Data source: Morgan Stanley Research Report (Charlie Chan et al., June 23, 2026) · Public market data.

TideResearch · 2026 June

Trending Cryptos

Related Questions

QAccording to the Morgan Stanley report, what is the predicted AI-related revenue for TSMC in 2027 and what does this represent compared to 2026?

AThe report predicts TSMC's AI-related revenue will reach $86.3 billion in 2027, representing a 218% increase from the $27.1 billion projected for 2026.

QWhat are the main drivers for TSMC's AI revenue growth in 2027 beyond Nvidia, as highlighted in the report?

ABeyond Nvidia, the main drivers are AMD's Venice CPU and MI400 series GPU for agentic AI, and Google's TPU demand facilitated through design partners like MediaTek and Broadcom.

QWhat is the critical bottleneck identified for TSMC's AI growth, and what are the projected global CoWoS capacity versus demand figures for 2027?

AThe critical bottleneck is advanced CoWoS packaging capacity. The report projects global CoWoS demand to reach 2.694 million units in 2027, while total global capacity (TSMC and non-TSMC) is expected to be approximately 2.8 million units per month, or 3.36 million units annually. The supply for the most advanced CoWoS-L technology remains tight.

QWhich companies are listed in the report as new winners in the AI supply chain and why?

AMediaTek is listed as the top pick due to its role as Google's primary design service partner for TPUs. ASE Technology and KYEC are also reiterated as Overweight due to their roles in AMD's Venice CPU supply chain and Nvidia's GPU testing, respectively.

QWhat are the three near-term catalysts for CoWoS demand mentioned by Morgan Stanley?

AThe three catalysts are: 1) Improved supply of ABF substrates, particularly the release of T-Glass capacity for MediaTek, which would boost Google TPU shipments. 2) Validation of emerging CPU demand from Nvidia's Vera and AMD's Venice. 3) The mass production of Nvidia's next-generation Rubin Ultra GPU in the second half of the year.

Related Reads

IPO Imminent, OpenAI Faces Major Personnel Upheaval

OpenAI, preparing for a potential IPO, is experiencing significant leadership turmoil. In mid-August 2026, longtime "GPU geek" Scott Gray quietly left, and within three days, Chief Operating Officer Brad Lightcap (8-year veteran) and Chief Revenue Officer Denise Dresser (8-month tenure) departed. This follows a broader exodus of at least 10 senior executives in 2026, including heads of product, safety, and ethics. Analysts view this as a strategic "surgery" to transform from a research lab into a sales-driven enterprise company before going public. Revenue now tilts toward enterprise clients, surpassing consumer income sooner than expected, with annualized revenue reaching $40 billion. The new CRO, Dali Rajic, is a veteran enterprise sales leader. Concurrently, OpenAI has disbanded independent safety teams like "Preparedness," which assessed catastrophic risks, integrating their functions into core research. Critics warn this removes dedicated "brakes" on AI development. The leadership vacuum raises questions about who is the clear second-in-command after former apps CEO Fidji Simo moved to an advisory role. Co-founder Greg Brockman appears to be consolidating power. As OpenAI races against rival Anthropic ($47B annualized revenue), it faces the dual challenge of commercial execution while managing the departure of foundational technical talent and ensuring responsible AI development remains a priority.

marsbit37m ago

IPO Imminent, OpenAI Faces Major Personnel Upheaval

marsbit37m ago

AI Can 'Have Moods Too'! New Research from USTC: Confusion and Anxiety Make AI Work Better

The article discusses research from the University of Science and Technology of China and Oxford, revealing that allowing AI to recognize and act upon simulated "internal emotions" can significantly improve its performance. The study demonstrates a coherent pairing between specific emotional states in AI agents and their subsequent skill choices. For instance, an agent feeling curious and desirous will search for products, while one feeling confused and tense will rephrase queries. This mirrors human decision-making influenced by emotions. Statistical validation showed a 76.5% semantic consistency in these pairings. Crucially, the research challenges the traditional view of AI errors as flaws to be eliminated. It found that "bad" emotions like confusion, tension, or frustration serve as useful metacognitive signals, indicating a mismatch between the current strategy and the environment. By responding to these signals, AI can proactively adjust before a failure occurs. This is particularly effective in complex tasks prone to failure. For example, in tasks like "heating an item" and "picking up two items," success rates surged from 9.6% to 56.9% and 4.4% to 31.3%, respectively, when using the emotion-driven skill selection method (EMOTION2SKILL). The AI's "nervous" state about a closed microwave, for instance, prompted it to check and open it first, preventing failure. The article also mentions related work from Tianjin University, which embeds emotional prediction into world models (Large Emotional World Model, LEWM), significantly improving prediction accuracy in human-centric environments. Removing emotional data was found to degrade performance even in unrelated logical reasoning tasks. These studies build on earlier findings, like those from Anthropic, that identifiable emotional representations exist within large language models (LLMs). The focus is shifting from philosophical debate about AI emotion to practically harnessing these internal states as functional signals to enhance AI robustness and capability.

marsbit1h ago

AI Can 'Have Moods Too'! New Research from USTC: Confusion and Anxiety Make AI Work Better

marsbit1h ago

Trading

Spot

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of S (S) are presented below.

活动图片