The big winner of chips in the next five years

Semiconductor Industry Observations

2024-05-29 09:17Posted on the official account of Anhui Semiconductor Industry Observation

We believe that the AI wave will bring profound changes, not only to the technology industry, but to society as a whole. These changes may have had as much impact on the world as the Agricultural Revolution and the Industrial Revolution, both of which had a huge impact on the economy.

While the exact course and timing of these changes is unpredictable, one thing is clear: the AI wave will not be possible without advances in hardware and software, especially silicon wafers, and a steady supply. The complexity of semiconductor design and manufacturing, coupled with rapid innovation and supply chain fragility, has created unique and challenging dynamics that, in our view, are reshaping leadership in the semiconductor industry.

Our projections show that 1) companies that provide manufacturing equipment, components, and software to build manufacturing facilities; 2) chip manufacturers; 3) The total revenue of chip and AI software design companies will approach $1 trillion this decade. Our research shows that four companies – Nvidia Corp., Taiwan Semiconductor Manufacturing Corporation, Broadcom Inc., and Qualcomm Inc. – will account for nearly half of this trillion-dollar opportunity.

In this edition of Breaking Analysis, we asked David Floyer, Analyst Emeritus at CUBE Research, to quantify and predict the dynamic semiconductor ecosystem. We compare our market share in 2010 to our market share in 2023 and offer a five-year outlook for more than a dozen of the top players in the industry. We also used some survey data from Enterprise Technology Research to analyze where the overall market is headed, our assumptions about the market and the top players, which companies we think will win and which will lose, and why.

We'll also discuss the following five items:

How sustainable is Nvidia's moat?

How does competition affect NVIDIA, including from hyperscalers, Intel Corporation, Supermicro Semiconductors, and others?

Challenges faced by two companies that design and manufacture semiconductors, Intel and Samsung Electronics Co., Ltd.

What edge opportunities mean for business and competition.

Risks in our scenarios include geopolitical, technological and energy risks.

Artificial intelligence captures budget momentum

Let's start with the macro impact of the awakening of generative AI over the past two years on information technology spending. The data below from ETR shows the 19 industries that the company tracks each quarter. The vertical axis is the rate of expenditure or net score, and the horizontal axis is the penetration rate, which is the penetration rate of the sector in the survey.

We've shown this several times before, but note what happened to ChatGPT in the October 2022 survey in the month before its release. That month, it fell below the 40% magic line and has been trending upwards ever since. As a result, other industries have been suppressed. As we reported, 42% of customers say they are stealing money from other budgets to fund AI. We know that, in general, the return on investment in enterprise AI is currently in the form of small productivity gains, and for most organizations, it is not self-funded.

The point is that AI consumes not only conversation, but also consumption momentum.

In our view, this has three significant implications for all organizations.

The first is to improve productivity across the board. We expect that in the coming years, people who adopt AI in large numbers will see an initial 20% increase in productivity and eventually a 50% increase.

The second is the quality of service. For example, a contact center representative should be able to answer any customer or prospect's questions correctly, accurately, and in a timely manner. Or customers can self-help with answers via voice on the company's website in the language of their choice.

Perhaps the most important value of AI for organizations is the potential automation of business processes. Specifically, the reduction of people in a business process streamlines the business process and the company as a whole.

Therefore, the guidance we provide to our clients is that a combination of the three is ideal. If you don't plan to increase productivity tenfold in the next five to ten years, startups and competitors are risking your business.

Nvidia's moat is wide and deep

Well, let's get straight to the point. Nvidia's momentum is phenomenal and has caught the attention of everyone in the industry. The speed of innovation in the AI ecosystem, especially Nvidia, is astounding. The chart below highlights the new era of computing we are in, catalyzed by large language models and AI breakthroughs.

This chart shows Nvidia's progress in trillions of floating-point operations since 2016. We superimpose the progressive description of Moore's Law. The comparison is striking: Nvidia achieved a 1,000x improvement in parallel/matrix computing (what Nvidia calls accelerated computing) in eight years, and Moore's Law improved 100x in ten years.

It's important to understand that in this issue, we've predicted the semiconductor industry ecosystem, and we've made trade-offs in scope. What we mean by this is that we're shaping Nvidia as a complete platform solution and a company that's building an end-to-end AI data center, what's called an AI factory. It sells this capability through partners.

One of the key aspects of Nvidia's moat is that it builds the entire AI system, which is then broken down and sold in chunks. So when it sells AI products, whether it's chips, networks, software, etc., it knows where the bottlenecks are and can help customers fine-tune their architecture.

In our opinion, Nvidia's moat is both deep and wide. It has an ecosystem that is driving innovation in a big way. CEO Jensen Huang announced that Blackwell has another big chip behind it — which is no surprise — and that its system and network "cadence" is once a year, and that the new system will run CUDA. Nvidia claims to be "all-in" on Ethernet, the company will continue to expand NVLink for a homogeneous AI factory, and Infiniland's roadmap will continue.

Huang's claim and bet is that the more money you spend on Nvidia, the more money you save and the more you can earn.

Furthermore:

We think Nvidia's performance gains are here to stay. This means that in five years we will see 100 trillion floating-point operations per second.

But the important thing is that Nvidia is not just a chip, but a complete AI platform. It has a dedicated graphics processing unit, CPU, networking, cooling, and software – it's a complete system software. CUDA is by far the best software in the industry. It is the key AI software platform.

NVIDIA can provide the entire AI data center. Not since 1964, when IBM introduced System 360, which revolutionized the computer industry, has there been such a revolutionary product.

In addition, we aim to pick up the pace and launch a new system every year. We think the value is so great for users, hyperscalers, and anyone who uses these technologies, and combined with the cost of creating alternatives, for us, that means that Nvidia will be the dominant vendor of AI data centers for at least the next five years.

Trillion-dollar semiconductor ecosystem

Let's take a look at the focus of this study and our five-year outlook for the ecosystem. The table below sets out our views on the development of the semiconductor industry. In the first column, we present the players in the ecosystem, including chip design houses such as Qualcomm, chipmakers such as TSMC, three leading companies that are in the chip and chip business at the same time – Intel, Samsung, and Micron Technology Inc., device manufacturers such as ASML Holding NV and Applied Materials, and software providers such as Cadence Design Systems Inc., which falls into the "other" category.

Of course, we've included Nvidia as well, and we think the company has become and will continue to be the most important player in the market. Once again, we're pushing the boundaries in terms of forecasting, predicting Nvidia's entire revenue stream, not just the chips.

For each company, we present their relevant revenues for 2010, 2023, as well as our projections for each company in 2028, along with the CAGR for the relevant time period.

Methods and key findings

We collected a range of relevant financial data for each company and combined this data with our underlying assumptions to create a top-down industry model, as we describe here. We tested this data with two external data points and added a third dimension, including: 1) corporate strategy projections based on the company's long-term financial framework; 2) the opinions of various financial analysts who make long-term forecasts for these companies; 3) Apply our own assumptions about how the market will develop.

We would like to point out that our assumptions and the resulting projections deviate significantly from the generally accepted market narrative. In particular, when you consider the basic data that is publicly available, the general consensus is that everyone is a winner, and the disruption to existing companies will be modest. But we don't think so. Instead, we predict a major shift towards matrix computing, or so-called accelerated computing; We're seeing significant spending shifts leading to market disruption, especially for the traditional x86 market.

We assessed the market and came to the following high-level conclusions:

As we expand our reach, the global semiconductor ecosystem will exceed $900 billion by 2028 and approach $1 trillion by 2030.

We forecast a CAGR of 10% from 2023 to 2028.

The market is massively shifting from general-purpose x86 to parallel AI computing architectures or matrix computing to support AI.

By 2028, four companies will account for about 40% of revenue in this forecast: Nvidia, TSMC, Broadcom, and Qualcomm.

Samsung and Intel, bucking the trend by vertically integrating design and manufacturing, face similar challenges.

Artificial intelligence computers will shorten the life cycle of computers. They're going to go mainstream, not only participating in Windows updates, but also changing the dynamics of your PC's lifespan.

Arm-based designs dominate the market and will deliver significant cost advantages to companies on the Arm curve

High-bandwidth memory, or HBM, has driven unprecedented demand for memory suppliers and created tailwinds for those companies that are able to produce them.

Let's look at the data in more detail by company, in descending order of our predictions for 2028. We'll showcase the company, our projected CAGR, and our revenue projections for 2028.

1. Nvidia: 25% CAGR and $160 billion in sales

In our view, Nvidia's essentially monopoly is somewhat similar to Wintel's duopoly in the 1990s, with core GPU dominance and AI operating systems both in the same company. We believe that as NVIDIA enters new markets, its growth rate will actually accelerate, with its revenue exceeding $160 billion by 2028.

Importantly, we're covering more than just chips in this forecast. Specifically, we assume Nvidia's overall platform and portfolio revenue; And let's say Nvidia continues to execute its portfolio at a fast pace.

Nvidia's performance is outstanding. It bets on super-sized chips and invests in GPUs, CPUs, networks, and software to provide complete solutions and complete data centers that can be disaggregated. Our assumptions and beliefs are that Nvidia will maintain this cadence for at least the next five years.

2. TSMC: CAGR of 14%, revenue of $135 billion

TSMC has become the go-to manufacturer of advanced chips. We expect TSMC to almost double in size over the next five years. Our core assumption is that economies of scale will provide TSMC with a significant strategic advantage, and that it will maintain its position as the world's number one foundry.

It is worth noting TSMC's investment approach. The company has just announced the A16, 1.6nm process, with the goal of launching in 2026. We believe this will be a significant milestone in its manufacturing industry, including nanosheet transistors and backside power supply. TSMC calls this the super power rail. We believe these innovations are industry-leading, and the company's proven track record in execution and high-volume delivery will keep it ahead of the curve.

3. Broadcom Semiconductor: compound annual growth rate of 10%, revenue of $58 billion

Next on the list is Broadcom, whose semiconductor revenue we only include. So, we believe that even though the company's CAGR slowed to 10%, it was actually because the company was very small in 2010. In our model, the majority of Broadcom's revenue in 2023 is spread out under the "Other" category.

Broadcom has achieved excellence through acquisitions and engineering. While it's a major silicon and AI chip IP supplier (through its custom silicon division) to Google LLC, Meta Platforms Inc., and ByteDance Co., Ltd. (which we think is ByteDance), it doesn't compete head-to-head with Nvidia in the GPU space. We expect Broadcom's semiconductor business to grow at a CAGR of 10% over the next five years, bringing the division 1.6 times its current size. It solves the really difficult problem of connecting all the GPUs, CPUs, NPUs, accelerators, and high-bandwidth memory together. It has a unique advantage in the market to continue to win. Broadcom is present in virtually every sector, including consumer, enterprise, mobile, cloud, and edge.

We believe that Broadcom is a well-positioned and well-run company. Its focus on the web is crucial. All types of high-speed networks are absolutely necessary for AI processing, and it is deeply ingrained in this market. In particular, it has established good relationships with major internet companies that will be leaders in artificial intelligence. As a result, Broadcom has early foresight of the most critical market trends.

4. Qualcomm: CAGR of 9% and sales of $55 billion

Qualcomm is in a very strong position in the mobile space and now in the AI PC space. We've seen that Microsoft's recently launched Windows AI PC stack has given it a huge boost. In terms of growth, Qualcomm's trajectory is similar to that of Broadcom. Essentially, Microsoft is following in Apple's footsteps a few years ago with Windows Copilot 11, which would be a big benefit for Qualcomm, which provides the core chip for AI PCs. This is bad news for x86-based PCs.

Microsoft has announced full support for Qualcomm-based Arm-based PCs. Now that Dell, Lenovo, and others are also announcing Arm-based PCs, all of a sudden, you're seeing a whole bunch of these moves that are directly catching up with Intel's PC install base with performance gains and 24-hour battery life.

So as you can see, our projections suggest that by 2028, the top four companies Nvidia, TSMC, Broadcom, and Qualcomm will account for about 45% of the $900-plus plus market.

5. Intel: CAGR of 2% and sales of $54 billion

We predict that by 2028, Intel's foundry revenue will account for about $22 billion of its $54 billion business. So, unlike many, we don't predict that Intel will grow during this period. We don't think the decline in x86 was offset by the increase in foundry revenue. Combined with our assumption that AMD continues to gain share of the x86 market, we expect Intel data center and client revenue to decline.

Intel is lagging behind in supporting AI PCs, and we expect its 14A process to be delayed by 12 to 18 months, which is a big bet for Intel. It combines surround gate technology (called RibbonFET) and backside power delivery (Intel calls it Power Via). The company wants to be the first to use high NA EUV technology, which is very bold when combined with other innovations, but it can also be delayed. So let's assume that 14A will get pushed.

Intel needs these three innovations to succeed and stand out in the industry. We think it's very likely that Intel will implement both of them, but even then the risks are significant. We assume that Intel's 14A will be in mass production and achieve high yield in 2028 (best-case) or 2029 (possibly), and possibly even 2030 (worst-case) with high yield.

In our opinion, the key to understanding Intel is that it has lost its sales lead. In our view, Apple and TSMC already take the lead, and Arm-based phones and PCs provide them with a significant learning curve moat, to Intel's detriment.

However, if Intel can successfully achieve 14A mass production in 2026 and launch a 10A 1nm node by the end of 2027 as planned, then our predictions will be incorrect and Intel will be in a much better position than we predicted.

6. ASML: CAGR of 7% and sales of $41 billion

ASML has a unique differentiator that will remain unmatched. Essentially, we believe that ASML will continue to maintain a monopoly position and be able to set prices at will.

7. SK hynix: CAGR of 10% and revenue of USD 40 billion

High-bandwidth memory has become the new enabler of AI. It is in short supply, which will drive SK hynix's growth. SK hynix's growth is actually accelerating, with revenue set to grow from $25 billion today to $40 billion by 2028. High-speed memory is very important, and the company has several options in this area.

8. Samsung Semiconductor: -1% CAGR, $38 billion

We think Samsung will have a hard time making its advanced processes work. We think it will continue to face challenges, which we believe will limit production and put it in a cost bind. We think Samsung's market capitalization is essentially the same as today's $40 billion.

Intel has said it plans to become the world's second-largest foundry by 2030. Considering Samsung's predicament, we believe this is the right target for Intel. The question is just whether Intel can achieve this. So, in that sense, catching up with Samsung is the right move.

9. AMD: 10% average annual growth rate and $37 billion in sales

CEO Lisa Su has achieved extraordinary things in this company. A key turning point was when AMD abandoned the fab, despite the famous saying of co-founder Jerry Sanders, "Real men have fabs." "This doesn't really prove that for AMD in the long run. It took a few years for the company to get back on track, but its persistence paid off.

AMD is still tightly tied to x86. By 2028, the end of our forecast period, AMD still has 45% of its revenue coming from x86, putting downward pressure on a large portion of the company's total available market. The good news is that our assumptions call for AMD to continue to take share from Intel while making progress on AI hardware.

Of course, Intel will fight desperately for share of x86 data centers, but we are more optimistic about AMD's prospects as a chip designer. It's not tethered to the foundry, and although the pressure from x86 is negative, we believe AMD will continue to hold the share. It's just faster to get to market, and the quality of the product is really good. For example, Oracle has just adopted AMD-based silicon for its new Exadata systems, which is a huge win for AMD.

10. Applied Materials: CAGR of 6% and sales of $35 billion

We believe Applied Materials will continue to play a role. It is in a very advantageous position. It's more competitive than ASML, but we think it's doing pretty well in this regard, growing from $27 billion in 2023 to $35 billion, at a CAGR of 6%. We basically forecast ASML, SK Hynix, Samsung, AMD, and AMAT to all have market caps between $35 billion and $40 billion.

11. Apple Semiconductor Value: CAGR of 12%, $33 billion

Essentially, what we're doing here is simulating the value contribution of Apple hardware to silicon wafers and making some assumptions about its value contribution in the value chain. Based on our assumptions, we found that Apple will grow at a CAGR of 15% from 2010 to 2023, compared to 12% from 2023 to 2028. Let's assume a wafer contribution of $33 billion.

There are reports that Apple will sell silicon wafers as a supplier. No such assumptions are made in our data. Still, Apple's entry into the business of designing its own chips is significant. It started with the A-series in smartphones and is now of course the M-series in its latest laptops and iPads. It was the first company to introduce a neural processing unit in both the iPhone and the PC. Now, as the competition for AI PCs heats up, it has to make a big move.

Apple has quietly led the wave of AI-powered PCs. Years ago, Apple introduced a large chip on the iPhone and integrated the CPU, NPU, and GPU on the same chip. It has a large shared SRAM, which is architecturally a leading example and is a great fit for AI. Apple has a proven track record when it comes to chips, for example, its M-series, M1, M2, M3, and now M4 are all evolving.

We believe Apple is leading the way in designing the silicon architecture needed for AI and think it will quickly respond to the Qualcomm AI PC trend. In our view, Apple is the main reason Microsoft is pushing to support Arm-based designs, as it is under pressure from Apple.

12. Micron: CAGR of 14% and sales of $31 billion

We believe that Micron can accelerate its growth rate, driven by high-bandwidth memory. Similar to SK hynix, Micron HBM's demand far outstrips supply. Micron performed very well. We expect its CAGR to accelerate to 14%, with revenue nearly doubling to $31 billion by 2028, from $16 billion in 2023. Micron doesn't just design chips, it's been a successful manufacturer for years.

13. Chips from cloud vendors – AWS, Google, Microsoft, Meta, Alibaba, ByteDance: 15% CAGR, $12 billion

We put hyperscale cloud providers in one category. Hyperscalers design their own chips and work with commercial vendors such as Broadcom. For example, our forecast for hyperscalers does not include Broadcom's custom chip contribution. We don't have double counting here.

We believe that hyperscale general-purpose, training, and inference chips will be used for cost-sensitive applications, such as edge inference. We assume that they will not keep pace with Nvidia at the high end, but they will get their fair share. Let's assume that AWS Graviton will account for about 20% of AWS workloads in 2023. Inferentia and Trainium make up a smaller percentage of AI workloads in 2023, as do their Google and Microsoft counterparts. Let's assume that hyperscalers make a significant contribution, but we don't think they will be the dominant factor in disrupting NVIDIA.

Hyperscalers are introducing Nvidia IP. They really have to adopt Nvidia because they can't make a similar platform on their own. We expect prices to be cheaper in the next five years, so they will continue to be big customers for Nvidia.

14. Other silicon ecosystem players: 4% CAGR and $175 billion in sales

Others include long-tail suppliers along the value chain. These include Chinese companies such as Texas Instruments, GlobalFoundries, YMTC, CXMT, startups such as Cerebras, and many more.

Let's assume that the AI PC market in general follows Apple's trend from x86 to Arm. We predict that x86 accounted for about 13% of market revenue in 2010 and will decline to 11% by 2023. It is expected to reach 5% by 2028.

Here's the visuals we just experienced. To save time, let's just say that the two companies that buck the trend among the leaders are Intel and Samsung. Micron has different businesses and has uniquely found a merger model. Artificial intelligence is bringing new investments to markets that investors have always considered risky.

AI PCs will shorten the life cycle

Here's our prediction for PCs since 2009. PC sales peaked in 2011, and that's exactly where Intel began to fall from its peak, though most people didn't realize it. David Floyer made a prediction in 2013. Here are the key points:

The consumption of iPhones has driven innovation in AI PCs.

Specifically, the first real inference came in 2017, when Apple used a neural processing unit (NPU) for facial recognition, an innovation that led to the emergence of the first NPUs in laptops and early examples of AI PCs.

Although PC sales have picked up during the pandemic, they have been declining. But we believe AI PCs are going to be a game-changer.

Microsoft has just reset its Windows AI stack around Arm – WinArm – to emulate Apple's move. PC manufacturers such as Dell Technologies and Hewlett-Packard are adopting it, and Qualcomm is seizing the moment. You can see in the green section below what we think this means for Arm-powered AI PCs, and what will happen to x86 PCs – which follows Apple's line. While not that severe, it is basically a manageable recession market.

While we predict a surge in PC sales, it's important to understand that this doesn't mean PC chipmakers like Intel will return to dominance. Currently, the market landscape is heavily influenced by Arm-based chips, which produce ten times more wafers than x86. Companies such as Nvidia, Apple, and Tesla recognized this shift early on and leveraged Wright's law to gain significant cost and time-to-market advantages in Arm-based chip design and manufacturing.

This shift highlights the growing value of ARM technology in reducing design costs and the challenges of x86. The market dynamics have fundamentally changed, and Arm's advancements have made it a dominant force, fundamentally changing the competitive landscape.

Final thoughts on key topics

Let us conclude with a few key issues that have not yet been addressed.

The future of AI and its market dynamics are rapidly evolving, with significant implications for key players and emerging technologies. Our analysis highlights the key trends and predictions that will shape the AI landscape over the next decade, with a focus on edge AI inference, energy demand, geopolitical risks, and potential shifts in semiconductor manufacturing.

We believe that the AI market is about to undergo a major transformation, with edge AI inference becoming the dominant workload. Energy innovation and geopolitical stability are essential to sustain this growth. Although NVIDIA is currently leading the way, the competitive landscape remains volatile, with potential changes that could be brought about by technological advancements and market disruptions. Our analysis highlights the need to keep a close eye on these developments as they will shape the future of AI.

How do you see the market developing over the next five years? What are your thoughts on our assumptions and predictions?

View original image 204K

• 

• 

• 

• 

•