TSMC, Samsung, and Intel, “grandeur et decadence” in the gen-AI era

[Daniel Nenni]

Two key news items emerged this week from the semiconductor foundry industry:

• TSMC reported its 3rd quarter Q3 2024 results on the 17th of October 2024
• 
  
• Intel reported seeking a foundry alliance with Samsung on the 22nd of October 2024

TSMC has emerged as the dominant player in the open foundry market. Much effort has been made to limit the world’s dependence on Taiwan’s foundries but with limited success. In light of Yole Group’s recent research, including Overview of the semiconductor foundry industry 2024, are these announcements really that surprising?

As a leading market research firm deeply embedded in the semiconductor ecosystem, Yole Group provides comprehensive market and technology insights, including detailed teardowns. Yole Group’s offerings deliver a thorough understanding of the semiconductor industry’s current challenges, business opportunities, and supply chain developments. Explore the entire collection HERE.

Considering recent announcements from Intel and TSMC, more updates are expected. Today, Yole Group expert Pierre Cambou reviews the current market status and shares insights into its evolution, along with the strategies of leading semiconductor foundries.

Enjoy Yole Group’s latest chronicle!

Why is it important?​

Q3-2024 Results from TSMC have been excellent, and many commentators have applauded the 39% YoY increase in quarterly revenue. 2023 was in fact the bottom of the semiconductor cycle so it was a relatively low year for TSMC, which had seen its revenues decline 4% while its market share improved by 1.3% compared to 2022. The open foundry giant is now growing, supported by two underlying trends. The first is the growth of the semiconductor industry, which is expected to grow 19% YoY this year and with a 9.8% CAGR in the next five years; the second is the increasing reliance on the fabless / open-foundry business model by semiconductor device players. In this respect, the new semiconductor stellar performer Nvidia will claim the crown as the largest semiconductor device company in 2024, the first time ever for a fabless company. TSMC is benefiting from both these trends and by capturing 59% market share in 2023, it is leaving little room for competition.

Discussing competition, Intel has been reported to seek a foundry alliance with Samsung. So far, Intel is not a significant player in the open foundry market but has grown as a very large capacity player as an IDM, the 6th largest in terms of production capacity, behind TSMC, Samsung, Micron, SK Hynix, and Kioxia. It would be the second largest player by production capacity if we put aside the IDM memory players, Samsung included.

Pierre CambouPrincipal Analyst, Global Semiconductors at Yole Group
This talked about alliance would, in fact, be combining Samsung and Intel production capabilities to overtake TSMC.

The story of the recent Intel effort is peculiar. Fueled by Chips Act money, it went into a capacity build-up spree to jump five nodes in four years (7-4-3-2-1.8-1.4). At the same time, Intel’s revenues stumbled as they are now half what they were at their $79B peak in 2021. To make it short, a finance-driven culture had siloed the company into a narrow market, x86 processors for personal computing and servers, consequently lowering investments in emerging markets such as GPUs for AI, a typical scenario described by Harvard professor C. Christensen in its book “The Innovator’s Dilemma“. The New Star is indeed Nvidia, which took much of the server market that Intel was relying on for growth; however, TSMC is the direct winner of this situation and, therefore, the target of the proposed Samsung / Intel alliance.

Could this work?​

Only time will tell, but we would be very surprised if the Intel Foundry unit manufactures Nvidia chips to survive. Nvidia is already porting its designs to Samsung to lower the supply chain risks – could it expand this scheme to include Intel in its supply chain as an external manufacturing partner? Nothing is impossible at this stage. The very survival of Intel’s manufacturing base is at stake, and with semiconductors becoming such a sensitive topic for the US regulator, it is not impossible for them to push for such a solution. With one less foundry player, this would reduce competition in the advanced node range while claiming to build a more credible alternative to TSMC, thus improving competition: two opposing views, in fact, which both make sense.

About the author​

Pierre Cambou, MSc, MBA, is Principal Analyst, Global Semiconductors at Yole Group.

Pierre’s mission is dedicated to market & technology analyses of the semiconductor industry. At Yole Group, Pierre has authored numerous market & technology products. Acknowledged as an expert in the semiconductor industry, he is regularly interviewed and quoted by leading international media.

Previously, Pierre held several positions at Thomson TCS, which became Atmel Grenoble (France) in 2001 and e2v Semiconductors in 2006. In 2012, he founded a semiconductor startup that was later acquired by Wx Solutions.

Pierre holds an engineering degree from Université de Technologie de Compiègne (France) and a Master of Science from Virginia Tech. (VA, USA). Pierre also graduated with an MBA from Grenoble Ecole de Management (France).

Related products​

• Overview of the Semiconductor Foundry Industry 2024
• Wafer Data Monitor
• Semiconductor Trends in Mobile and Consumer 2024
• Generative AI 2024 – Impact on Processors, Memory, Advanced Packaging, and Substrates
• Semiconductor Trends in Data Centers 2024
• Overview of the Semiconductor Devices Industry 2023

Source: www.yolegroup.com

 

[samwilde]

That last chart makes it look like SMIC has less capacity than UMC. Totally bogus.

SMIC is #3 in capacity behind TSMC and Samsung. And Hua Hong will surpass UMC and GlobalFoundries to become #4 until the end of the decade.

 

[Daniel Nenni]

That last chart makes it look like SMIC has less capacity than UMC. Totally bogus.

SMIC is #3 in capacity behind TSMC and Samsung. And Hua Hong will surpass UMC and GlobalFoundries to become #4 until the end of the decade.

SMIC's capacity is definitely understated and will continue to grow. Same with Hua Hong. Say good bye to the Chinese fabless market and thank you to the US Government for making it happen.

 

[blueone]

The New Star is indeed Nvidia, which took much of the server market that Intel was relying on for growth

This is nonsense. Nvidia has not taken server market share from Intel. AMD has. Nvidia made their own market in data center AI training and inference that Intel isn't a significant factor in. Yet another self-appointed expert in semiconductors who really doesn't seem to know what he's talking about.

 

[Daniel Nenni]

This is nonsense. Nvidia has not taken server market share from Intel. AMD has. Nvidia made their own market in data center AI training and inference that Intel isn't a significant factor in. Yet another self-appointed expert in semiconductors who really doesn't seem to know what he's talking about.

He is definitely out of touch. Years ago Intel had the compute datacenter market. Intel and AMD now have FPGAs and GPUs for the server market but nothing like Nvidia, but again, it is about the ecosystem and Nvidia has a great one. There is no way for AMD or Intel to cut and paste an ecosystem that is years in the making even if they had competing silicon which they do not. Nvidia is an AI race horse, AMD and Intel are x86 plow horses.

 

[XYang2023]

He is definitely out of touch. Years ago Intel had the compute datacenter market. Intel and AMD now have FPGAs and GPUs for the server market but nothing like Nvidia, but again, it is about the ecosystem and Nvidia has a great one. There is no way for AMD or Intel to cut and paste an ecosystem that is years in the making even if they had competing silicon which they do not. Nvidia is an AI race horse, AMD and Intel are x86 plow horses.

If I would need to deploy Llama based solutions (7b, 70b, 405b, not just hosting), I would consider Gaudi systems. CUDA is not really a problem here.

 

[blueone]

He is definitely out of touch. Years ago Intel had the compute datacenter market. Intel and AMD now have FPGAs and GPUs for the server market but nothing like Nvidia, but again, it is about the ecosystem and Nvidia has a great one. There is no way for AMD or Intel to cut and paste an ecosystem that is years in the making even if they had competing silicon which they do not. Nvidia is an AI race horse, AMD and Intel are x86 plow horses.

My issue with the statement I quoted is that the term "server" is generally known in the computer industry to refer to general purpose systems, meaning they are based on CPUs, not GPUs or other AI-specific processors. IMO, any legitimate industry expert knows that.

AMD is only recently serious about datacenter GPUs. It should be interesting to see if their investment turns out to be successful against Nvidia. As you say, that will take time. I agree with you about the importance of an ecosystem and a rich software strategy, but this is one area I think Intel is probably the most capable player in the overall computer industry. The challenge for them is that Gelsinger is not serious about anything except CPUs, so the Gaudi strategy for AI is really a sideshow. The primary Intel AI strategy is CPU-based, which I believe is unlikely to be competitive in the long run.

 

[XYang2023]

My issue with the statement I quoted is that the term "server" is generally known in the computer industry to refer to general purpose systems, meaning they are based on CPUs, not GPUs or other AI-specific processors. IMO, any legitimate industry expert knows that.

AMD is only recently serious about datacenter GPUs. It should be interesting to see if their investment turns out to be successful against Nvidia. As you say, that will take time. I agree with you about the importance of an ecosystem and a rich software strategy, but this is one area I think Intel is probably the most capable player in the overall computer industry. The challenge for them is that Gelsinger is not serious about anything except CPUs, so the Gaudi strategy for AI is really a sideshow. The primary Intel AI strategy is CPU-based, which I believe is unlikely to be competitive in the long run.

I don't think that is just a sideshow. With AI, you need to use some form of accelerators, and Gaudi is what they currently have for server clients. Falcon Shores, expected in 2025 with up to 1500W support, will also support oneAPI. This means that one piece of software will be able to scale from edge devices (AI PC) to data centers.

Here is one use-case demo:

 

[blueone]

I don't think that is just a sideshow. With AI, you need to use some form of accelerators, and Gaudi is what they currently have for server clients. Falcon Shores, expected in 2025 with up to 1500W support, will also support oneAPI. This means that one piece of software will be able to scale from edge devices (AI PC) to data centers.

I'm familiar with OneAPI, and I think it's a brilliant piece of work. But it's not going to unseat CUDA from its leadership position. For the time being, OneAPI's primary market is a sophisticated and standardized way of programming FPGAs in C++. That can be a substantial value-add for ASIC companies looking to quickly prototype their ASIC designs, or produce FPGA-based products with very reasonable development costs. I haven't heard much outside of Intel about OneAPI being used for AI products.

 

[XYang2023]

I'm familiar with OneAPI, and I think it's a brilliant piece of work. But it's not going to unseat CUDA from its leadership position. For the time being, OneAPI's primary market is a sophisticated and standardized way of programming FPGAs in C++. That can be a substantial value-add for ASIC companies looking to quickly prototype their ASIC designs, or produce FPGA-based products with very reasonable development costs. I haven't heard much outside of Intel about OneAPI being used for AI products.

I think it just takes time. Cost is an important consideration for buyers, and what Nvidia is currently experiencing is not sustainable. It leaves no margin for the rest of the value chain.

 

[tooLongInEDA]

Is this real ?

He (Pierre Cambou, Yole Group) claims that Samsung's "open foundry strategy has paid off" and that "is has now become a credible alternative to TSMC".

Isn't that the exact opposite of most observations here over the past couple of years ?

Then the foundry capacity proportional area chart shows Samsung's foundry capacity as almost equal to TSMC's. Yet TSMC's actual revenues are around 5X those of Samsung. Does this mean that Samsung is operating at only 20% capacity ?

I'd be interested to know who acknowledges him as an expert in the semiconductor industry, because some of this stuff seems pretty dubious.

 

[nghanayem]

Is this real ?

He (Pierre Cambou, Yole Group) claims that Samsung's "open foundry strategy has paid off" and that "is has now become a credible alternative to TSMC".

Isn't that the exact opposite of most observations here over the past couple of years ?

Then the foundry capacity proportional area chart shows Samsung's foundry capacity as almost equal to TSMC's. Yet TSMC's actual revenues are around 5X those of Samsung. Does this mean that Samsung is operating at only 20% capacity ?

I'd be interested to know who acknowledges him as an expert in the semiconductor industry, because some of this stuff seems pretty dubious.

He also calls SK, Kioxia, Micron, and other IDMs as "foundries". So I assume he is counting memory. Stupidity aside, I highly doubt TSMC has more wafer starts than Samsung (if you include memory). Samsung has a larger share of memory than TSMC has in logic and there are more memory chips/CMOS image sensors made per year than logic chips.