Intel – Same Chip, New Package.
In the 1980s, Intel maneuvered in what was said to be one of the best corporate turnarounds as it ditches its core memory business and plunged headfirst into the then nascent CPU computing chips industry. Headed then by CEO Andy Grove, Intel took what was unimaginable and turned it into a 30-year stranglehold over the semiconductor industry. What Intel required then was a serious level of organizational culture alignment, extraordinary leadership execution, and a competitive advantage where it can leverage itself upon. Fast forward to today, Intel currently faces yet another existential crisis, one that is significantly more difficult, different, and structurally challenges Intel’s dominant position in the semiconductor industry. Yet, despite the limited visibility on the future path of the semiconductor industry, I believe that the 3 fundamental principles that I identified above would provide a foundation of which we can build our assumptions for the path Intel could take moving forward. These 3 pillars are exactly what I will be building my thesis on Intel and why I think Intel could once again revive itself into a dominant position despite the herculean challenge ahead.
History
Intel first started in 1968 by then semiconductor pioneers Gordon Moore, Robert Noyce, and Arthur Rock in the space of SRAM and DRAM memory chips. These memory chips represented most of Intel’s business up during the early years when it extracted high levels of margins and profitability supplying such chips. However, during the 1980s, accompanied by the onset of large Japanese chip companies who was supplied with cheap capital from group owned banks – “built enormous fabs, manufactured efficiently (kaizen), and came to possess unit cost advantages with which underprice Intel and gained market share.” (Scuttleblurb, 2023). These Japanese players initially were beneficial to Intel, as they helped fill product shortages during recession when Intel had to pull back investments in production capacities. However, their quality and technology soon reached parity and beyond Intel’s capabilities. In 1976, Japan had less than 30% market share. But by 1986, Japan has overtaken U.S and became the leader in supply memory chips, with more than 50% of market share.
Initially, Intel tried its best to improve its quality. It brought cost down, improve quality and performance metrics but the Japanese players were always able to fight back due to their low-cost structures. The Japanese’s sales strategy follows as such – “1. Find out Intel and AMD’s socket prices 2. Quote 10% below their price. 3. If they requote, go 10% below again. 4. Do not stop until you win”. Looking back, Intel was fighting a battle it cannot win.
Facing such odds, Intel must decide – 1. Do I double down on memory R&D spending to regain technological edge? At the point in time, Intel was spreading R&D across 3 different technologies, with the bulk of the spending already on memory chips while the remaining was invested in microprocessors. 2. Do I give up my largest revenue driver and jump on the relatively nascent microprocessor industry? In the 1980s, microprocessor demand was still low, and demand for it represented a slower-growing, smaller volume market size than memory chips. 3. Do I do nothing? Although the competitive landscape in the memory segment has toughened, Intel was still extremely profitable during up-cycles. It generated high margins, relatively strong topline growth, and was comfortable. Not surprisingly, Intel chose to do nothing.
However, by the fall of 1984, the semiconductor cycle stopped and switched. Intel began hemorrhaging cash as demand for its products dwindled while the Japanese continued to undercut Intel’s pricing. The need for a different strategy became apparent as Intel increasingly faces pressure on the two fronts. It was not only until mid-1985 when the fateful conversation between Gordon Moore and Andy Grove occur which allowed Intel to make the leap of faith to divest out of memory chips and pivot into the nascent but growing CPU segment.
Intel’s success over the past 3 decades was driven by 1. Organizational Culture Alignment, 2. Extraordinary Leadership Vision and Execution, 3. Competitive Advantage where it could leverage itself upon. Although it is hard to quantify and pin-point these qualitative factors as the key driver for Intel past and future success, I believe the observations and arguments that I put forth would be reasonable to suggest a strong level of relationship between Intel’s success and the said forth factors. I will attempt to highlight the parallels between history and the uncanny resemblance between what has occurred and what is happening. History does not repeat itself, but it does rhymes.
Economics and How the Semiconductor Industry Operates
The semiconductor industry is incredibly competitive, with the number of large public semiconductor across the value chain shrinking to about 97 by the end of 2020, as compared to more than 200 in the 2010s. The economics of this industry, one that is characterized by its high capital intensity, R&D spending scale and limited talent pool, creates a pull where the current dominant players become more dominant as it builds up scale and talent which puts immense pressure on middle and smaller players to catch up.
There are generally 2 operating playbooks in this industry, Fabless (Specialized) and IDM (Integrated), although other permutations exist such as Fab-lite. Over the past decade, the economics of this industry as described above resulted in the need for specialization, as both the intensity for R&D and capital expenditures is too high for a single company to undertake without sufficient scale. Only 3 companies exist under the IDM model today, namely Samsung, Intel, and Micron. Only Intel competes in the most competitive segment, namely microprocessors for consumer computers and data center chips.
What is the edge of being an IDM? This seems to be a key question in answering why Intel could remain relevant. After all, if there are no edge remaining an IDM, then investing in Intel now makes no sense. Over the past decade, it is hard to argue that an IDM model is still relevant when the cost of opening a new “Fab” could cost up to $5 billion and customer concentration means that players such as Apple and Nvidia hold most of the bargaining power. However, the key advantage of being an IDM is the potential for cost optimization. Intel noted in its recent earnings call that the “current foundry + fab-less in the same enterprise, gross margin could have a >20% advantage”. This argument stands to be very different from what the industry sees, which propelled the fab-less model. However, looking at Intel financials prior to 14nm where it faced manufacturing challenges, Intel did have a gross margin advantage. Furthermore, note that previously Intel was using more fab space for the same capacity output due to lower tool utilization and worse yields.
Notes: On the demand side, consumers unceasingly insist on better features, greater reliability, and higher speeds. This drives the R&D intensities which on average consumes around 15 – 20% of revenue (Large players have scale advantages at same level of relative spendings). On the supply side, fierce competition means that downstream players like manufacturers, testing, assemblies, and packaging must reach sufficient unit scale to achieve unique low-cost unit economics. This has created geographical specializations, where high technological, value-adding work are being done in in Silicon Valley (AMD/NVIDIA) while downstream technologies are usually done by companies in Taiwan and China, such as Foxconn and Flextronics.
Chart 1. Generalized Product Lifecycle
What is the difference in the economics between IDM and Contract Foundry like TSMC?
In the semiconductor industry, companies face what is called “deflationary pricing” where over time, the price of their products/inventory reduces as it becomes out-of-favor and requires heavy discounting. This makes inventory obsolescence a key risk for companies like Intel. Intel faces a general product lifecycle as shown in the chart below. Generally, there will be 1-2 years of supranormal pricing where Intel would be able to charge a high price before prices deflate. Intel generates most of its revenue and profits on the leading nodes, which is why it is crucial for Intel to re-obtain technology leadership as the loss of technology edge means the loss of its main source of revenue and profits.
In contrast, TSMC, operating as a contract foundry, derives a significant portion of profits and revenue (~ 50%) from trailing nodes as compared to leading nodes. This is because foundries at trailing nodes are already fully depreciated, therefore, whatever revenue generated from it has a high level of profit margin as compared to leading nodes.
What could the economics potentially look like for Intel IDM 2.0?
To imagine what the economics may look like for a successful Intel IDM2.0, one must first establish a few core assumptions. 1. Intel achieves technology parity or edge. 2. Intel generates sufficient interest in its IFS business. 3. Intel can synergies cost efficiencies between chip designs and manufacturing through data sharing.
· Assumption 1: Intel Achieves Technology Parity or Edge.
There are generally 2 aspects of argument for technology edge. One can either look at density or relative performance of chips. In this case, based on expert’s opinions (Chart below), it is possible for Intel to achieve performance edge (performance per watt) but not density edge. Generally, higher density reflects lower costs for chips. On the other hand, performance refers to chip capabilities/specifications. It is hard to argue or determine which is more crucial from the customer perspective as it depends on the customer’s requirements and goals. However, it is reasonable to argue that should Intel be able to achieve technology parity/edge in the relative performance of chip, this would allow it to steal market share away from Samsung or TSMC.
Chart 2. Transistor Density Chart 3. Relative Performance of Chips
· Assumption 2: Intel generates sufficient interest in its IFS business.
Building on the previous assumption, Intel achieves technology parity and edge. However, it is important to note that this is not sufficient. At the core of a foundry business, it is essentially a customer-facing business, where there needs to be strong customer support and communication. This means that a culture match must be achieved for trust to be built. For this, I view that Intel would have a significant opportunity to steal the 15% market share from Samsung given that Samsung has destroyed its foundry trust by stealing away IP from Apple and its customers. Conversely, I believe that TSMC would still hold most of the market share due to this fundamental driver. Therefore, the base case for Intel IFS market share in 2025 would be around 15%, largely driven by steals from Samsung.
· Assumption 3. Intel synergies cost efficiencies between chip designs and manufacturing.
As noted above on the core edge of being an IDM, the third assumption builds on the fact that Intel can achieve cost efficiency between its chip design segment and manufacturing. Leveraging upon assumption 1 and 2, it means that Intel Foundry Services has 1. Achieve technology parity, 2. Generated sufficient customer interest. What this translates to is a higher level of foundry utilization as compared to being a pure IDM, which would translate to better margins overall. Historically, Intel gross margin ranges between 58 – 62%, while TSMC ranges between 47 – 50%. Using the core assumptions above, we can generalize and forecast 3 reasonable cases for the future of IFS.
1. Base Case – 15% market share (2025) of total foundry market, 30% EBIT Margins
2. Bear Case – 10% market share (2025) of total foundry market, 25% EBIT Margins
3. Bull Case – 20% market share (2025) of total foundry market, 35% EBIT Margins
Thesis 1. Organizational Culture Alignment (History)
When Intel first pivot into CPU, this leap of faith was not without its cost. By pivoting into CPUs, Intel is in essence, giving up its identity. This was not an easy conversation to have given that the cultural and corporate identity of Intel being a memory chip player were engrained deeply into each salesman, designer, and engineer employed at Intel. Pivoting into CPUs means admitting defeat and it requires a monumental shift in corporate identity, culture, and grit. Yet, when the decision was made to pivot and announcements were made to customers, the response and feedback from customers were much more muted than expected. Customers were surprised that Intel took so long to make this decision given that it was the most rational decision to non-invested observers and have been already arranging for secondary suppliers’ way before Intel even contemplated getting out of the memory business. This was the first positive indicator that Intel had made the right decision.
The second indicator came when top management gathered all the memory developers of Intel into an auditorium and announced that the focus of the company from now on will be in microprocessors, not memory, and that by staying in Intel, it would signal that they are no longer memory developers, but microprocessor developers bearing the flag for Intel’s mainline business. The response from the group was unexpected. According to Andy Grove, “these people (developers), like our customers, already know what was inevitable before senior management faced up to it. There was this measure of relief that they no longer had to work on something that the company wasn’t fully committed to. This group, in fact, threw themselves into microprocessor development and they have done a bang-up job ever since.” Although Intel faced tremendous difficulties moving forward, including massive layoffs and factory shutdowns, its newest product, the 386 chip, became the most successful microprocessor that ever existed. As the 1990s arrive, Intel have already invested heavily in new microprocessor designs that were meant to foster the growth of the computer industry. Through industry partnership first with IBM and then later consumer partnerships with Microsoft, Intel became the dominant supplier of CPUs and began riding the rise of the personal computing industry.
By 1992, Intel became the largest semiconductor company in the world, even larger than the Japanese memory companies with their cheap capital and low-cost structures. In the book “Only the paranoid survive” written by Intel’s CEO Andy Grove, of which is the key source material I used to write this segment of the paper, summed up perfectly the importance of organizational culture alignment in combating in what seem to be an ever changing, volatile, uncertain, risky business environment.
“One last lesson, and this is a key one: while Intel’s business changed and management was looking for clever memory strategies and arguing among themselves, trying to figure out how to fight an unwinnable war, men and women lower in the organization, unbeknownst to us, got us ready to execute the strategic turn that saved our necks and gave us a great future.
Over time, more and more of our production resources were directed to the emerging microprocessor business, not as a result of any specific strategic direction by senior management but as a result of daily decisions by middle managers: the production planners and the finance people who sat around the table at endless production allocation meetings. Bit by bit, they allocated more and more of our silicon wafer production capacities to those lines which were more profitable, like microprocessors, by taking production capacity away from the money-losing memory business.
Simply by doing their daily work, these middle managers were adjusting Intel’s strategic posture. By the time we made the decision to exit the memory business, only one out of eight silicon fabrication plants was producing memories. The exit decision had less drastic consequences as a result of the actions of our middle managers.” (Andy Grove, 1988)
Thesis 2. Extraordinary CEO Execution (History)
In hindsight, everything is much clearer. Yet, the scope for decision making is often the greatest when the relevant information is at the scantiest. By the time more data becomes available, the margin of maneuver tends to have narrowed. So, how can one prepare for such challenges and us as investors, evaluate whether an organization can make hard decisions under the most limited of options and data? History might provide some clues.
We have arrived at the part of the paper where it is likely to be the most normative, questionable, and subjective argument where it is hard to falsify the underlying factors of “Extraordinary CEO Execution”. Afterall, what attributes are considered extraordinary and what attributes are considered ordinary? Nonetheless, I will attempt to situate the insights from a business perspective, and address why these key insights would be fundamental to allow us to identify the same parallels we now see in current cases.
Firstly, the CEO must have the courage to make bold decisions while remaining realistic. In history, the pivot from memory to microprocessors appears to be a case which we can attribute to extraordinary CEO execution. However, I must qualify this argument by stating it was the middle management of Intel that even made such a pivot an option for Intel to undertake. Nonetheless, the final trigger was required by then Intel CEO, Andy Grove, to undertake.
Secondly, the CEO must have foresight and vision to leverage on current strengths while reducing current weaknesses. The first characteristic of this was the ability to understand the importance of trust and the fact that Intel has pivoted into a consumer facing business. Looking back at the financials, why has Intel subsidized OEMs and placed such an effort in its marketing? Intel had initially invested about $250 million into the marketing campaign where it provided subsidies to OEMs who included their logo “Intel Inside” in their own products and advertisements. The belief was that this would encourage consumers to recognize Intel as a sign of quality and innovation. In the 1990s, this was an abnormal case as it was during a period where no major company had ever tried “ingredient branding” for technology components before. However, by the end of 1992, over 500 OEMs have signed onto the cooperative marketing program and 70% of OEM ads carried Intel’s logo. Much of the benefits is still in effect today as Intel still holds significant sway in the enterprise segment where purchasing considerations often tilts towards Intel as the standard for quality and reliability. This is why even though the client computing segment of Intel proves to be much stickier as compared to the data center segment of its business.
Thesis 3. Company’s Competitive Advantage and Moat. (History)
What I am interested in figuring out here is why Intel was able to become successful in pursuing the microprocessor industry. What strategic moat does Intel have to operate in an environment where not only the visibility on future demand is limited and constantly changing, but also the lack of a viable operating playbook to guide the type of investment and technological decisions Intel needs to make? Is the moat still present today?
During the 1990s, there was a battle between RISC (Arm/ RISC-V) processors against CISC (intel) processors and which architecture would be the standard in the industry. RISC processors, even today, are generally much simpler to design. On the other hand, CISC processors are much more expensive to design given the complication driven by the x86 instruction set. Therefore, for high-end workstations where the general volume is much lower, there is a structural advantage in using RISC. However, for the general average user, CISC processors were much more affordable. This is because although design costs were higher on average, Intel possess the production scale and volume. In the semiconductor industry, higher scale meant lower unit prices. This manufacturing edge meant that CISC based processors are more likely to achieve higher and faster penetration as compared to RISC based processors despite the lack in performance.
Secondly, a key driver of RISC failure was the lack of an open platform. As RISC based workstations are often premium, they are usually sold at premium prices. The problem with such a business model is that should the low-cost alternative improve themselves, which they often can and will, they often start to move up the value-chain and capturing more profits. With more profits, this creates the virtuous cycle of improvement. The economies of scale were the major competitive advantage that Intel had.
(Note: This is also the exact issue that Intel currently faces. At the core issue, Intel faces these problems; 1. Uncertain business environment, product obsolescence risks, demand uncertainty. Also, if I am arguing that the competition is shifting towards packaging technology, then why is Intel able to capitalize on this and scale? Why is intel able to make the right investment decision? Yes, competitive dynamics is not as skewed if we are comparing purely on a manufacturing prowess where TSMC dominates. But still, I need to investigate the history to see if I can draw any fundamental truth and strength that Intel posseses, be it human capital, know-hows, intellectual property, networks, which it can once again, derive its strength from as it tries to compete on the packaging front.
So, what now? (Differences and Similarities within History and Current)
Fast forward to today, Intel currently faces a crisis that is significantly more difficult and different as compared to the one imposed by the Japanese memory players. Firstly, we have changes in business environments. Intel must operate now as an entrant rather than as an incumbent. In its recent announcements and presentations, Intel announced that it is entering into the AI space as a software and platform company with an ecosystem to complement their hardware product lines. For example, it has launched several new software toolkits such as the OpenVINO (programming model), oneAPI which allows developers to deploy a single developmental code across different types of chips and architectures. However, one should be skeptical of being optimistic about such efforts to help gain ground in this highly competitive space. As Nvidia CEO mentioned in the 2Q2019 earnings call, “I don’t really know how one programming approach or a simple API can help make 7 different weird things work together. I can’t fit it in my head. Programming is not as simple as making a PowerPoint slide.” Furthermore, Intel’s existing market share in data centers is nothing if not misleading. According to a source from counterpoint, Intel continues to lead in market share for its data center CPU market, with a 71% market share in 2022. However, if we only consider Intel’s marketshare on new data center spends, Intel only manages to capture about 30% of 2023 expected data center revenues, with AMD capturing around 15%, up from 0% 5 years ago, and Nvidia capturing 57%, up from 12% 5 years ago
Chart 4. x86 Data Center Revenue
Secondly, there is no technological edge that Intel can leverage on. In fact, Intel is 1-2 generations behind competitors across product lines in data centers, consumer chips, networking equipment, AI accelerators, FGPAs and more. Its progress is also stiffened by the bottlenecks and challenges it had with making manufacturing progress. The analyst, David Kim, author of the investment blog, scuttle blurb, sums up the technological gap Intel had and faces moving forward.
“Intel supposedly has smaller transistor gates and therefore more density than peers at the same node - Intel’s 14nm is equivalent to TSMC’s 10nm, its 10nm equivalent to TSMC’s 7nm, etc. But even accounting for this, Intel is falling behind. In 2016, Intel was 3 years into 14nm when TSMC had just transitioned to 10nm; by 1h20, Intel was ramping up 10nmproduction while TSMC was already shipping 7nm and 5nm. By the time Intel starts shipping 7nm in mid-2023, TSMC will be at 3nm…so even if Intel’s 7nm is technically equivalent to TSMC at 5nm, TSMC, and by extension AMD, which fabs with TSMC, will still be running ahead.” (Scuttleblurb, 2023).
Thirdly, if not the most important factor, Moore’s law has slowed tremendously. Moore’s law, which refers to the observation that transistors in an integrated circuit double about every 2 years, has slowed significantly as we push the limits of physics. Ironically, Moore’s law isn’t so much of a law, but rather, a tempo that was created by Intel strict adherence to its “Tick-Tock” production model where every microarchitecture change (tick) is followed by a die shrinkage of the process technology (tock). Why is this important? Because now, it is significantly more costly and difficult to play catch up. Comparing Capex Intensity of only the manufacturing segment of the semiconductor business, capex spendings as an industry has close to 3x in 10 years.
However, on the flip side, the slowing of Moore’s law is also a strategic inflection opportunity for Intel because it is shifting the playing field from competing purely on manufacturing prowess, one that TSMC has a stranglehold upon, into a much more balanced field – Advanced Packaging. But before analyzing the advanced packaging dynamics, one need to ensure that Intel has 1. Regain the organizational culture needed and 2. Re-obtain extraordinary, grovian-like Leadership Execution. As stated above, these are the 2 necessary conditions that Intel needs to have in order to compete in this new business environment. These are the pillars that Intel will have to rely upon to draw its strength from. These will also be the 2 pillars I will first expand my thesis on.
1. Driver 1. Organizational Culture Alignment (Future)
As I have repeatedly mentioned above, analysis of qualitative factors such as culture and CEO is subjective at best. However, there are “indicators” which may be signs of Intel taking the right step. For this case, till date, there is only 1 significant indicator from a cultural perspective that Intel is making the right steps – 1. Establishment of IFS with separate P&L statement, with 3rd party compatibility and, opening of IP portfolio to customers. When Intel first announced this IDM2.0 strategy, many were surprised but felt cautious about this. The common impression is that this sounds “too good to be true”. Afterall, Intel has tried previously to be a foundry but has failed. Why would Intel be able to succeed now?
Firstly, the underlying motivations of Intel becoming a foundry is drastically different. Previously, it was viewed as a secondary means of revenue and profitability driver. Now, Intel Foundry Services is viewed as a means for survival and relevancy. With this switch in perspective comes a full-fledged commitment in trying to make IFS a success. This shifts the probability of success to be not dependent on culture, but to the competitive environment and offerings of IFS.
Secondly, the idea is well-received by customers. Qualcomm’s president announced that “Qualcomm is excited about the breakthrough RibbonFet and PowerVia Technologies coming in Intel 20A. We are also pleased to have another leading-edge foundry partner that will help the U.S fabless industry to bring its products to an onshore manufacturing site.” In other words, following the same rhetoric from our historical analysis, what Qualcomm and many other customers were trying to say is “What took you so long?”. This again is another indicator of Intel taking the right steps. But more importantly so, Qualcomm’s announcement further highlight the importance of “another leading-edge foundry”. This puts weight behind the speculation and argument that fabless players would shift volumes away from TSMC into Intel should Intel start regaining technological edge.
Lastly, culture change. Anecdotal observations by ex-intel managers and technical crew note that the internal culture has shifted from being “overtly confident in their products to becoming more aware and motivated in proving that they are still indeed the ‘king’ of chips”. Though I agree that basing an investment decision on anecdotal observations and opinions of experts does not necessarily warrants conviction for an investment thesis, these are again small indicators of Intel’s underlying switch in culture. Many sell-side analyst would argue that Intel’s history as an IDM over the past 20 years has created a deep culture that would conflict against the demands of being a contract foundry. However, looking back at history once again, it took Intel 2 years to change it deeply ingrained belief of being a company that was founded on memory chips to being a company that is synonymous with microprocessors. History shows that the odds are on Intel’s side.
However, it is in my opinion that Intel still doesn’t get the foundry model. Being a foundry means putting the customer first. I get it - Intel perhaps would only want to present its best products and services to its customers so that the customer will get receive the best support and service Intel can provide. This is why the current process flow is such that it first begins with an internal node before opening up 6 - 8 months later to foundry customers once it has resolved any internal issues. However, no leading edge customer will be willing to wait 6- 8 months while you try and fix your problems. Intel needs to involve the customers at the beginning of the process, not at the end! Should Intel move towards this form of conducting business, then this would be a significant indicator that Intel would be able to successfully execute this strategy.
2. Driver 2. Competitive Moat – Advanced Packaging
As mentioned above, the slowing of Moore’s law has provided Intel with a strategic inflection opportunity, one that shifts the competitive dynamics from competing purely on manufacturing prowess, into the area of Advanced Packaging where the competitive landscape is much more balanced. In this segment, I will explore whether Intel holds any significant advantages/moat in this segment and why perhaps Intel would be able to leverage upon this technology to recover its positioning in this industry.
What is Advanced Packaging? Why is it important?
Historically, packaging is viewed as a downstream, labor intensive and low value-added back-end activity. As a result, most of these packaging activities were being outsourced down into Asia where there is a structural lower cost of operating due to lower labor costs. However, as Moore’s law slows, packaging technologies are now becoming structurally important and a key determinant of future depth and breadth of innovation for emerging technologies.
Advanced packaging in essence refers to the packaging techniques that are associated with a decrease in package sizes and power consumption along with an increase with the density of interconnects. In simpler terms, it generally refers to the vertical stacking of chips, called 2.5D Packaging or 3D Packaging although other forms of advanced packaging such as Flip-Chip and Fan-In packaging techniques are currently more prominently deployed. Historically, transistors sized are shrunk and increasingly fitted onto a small piece of silicon, with extremely high transistor density. With the slowing of Moore’s law, transistors size is now shrinking at a slower rate which has a dragging impact on the three-performance metrics, namely Power, Performance and Area (PPA).
In 2014, the overall packaging market value stood at $53 billion, with 20 billion of that in advanced packaging. By 2024, estimates suggest that the overall market value of packaging will be around $96 billion, with advanced packaging accounting for about $48 billion (50%). Currently, advanced packaging technologies are mainly used in mobile consumer electronics, with the System-on-a-chip packaging technique. However, more recently, electronic devices supporting telecommunications and infrastructure, as well as automotive, defense and other end-markets have started to include advanced packaging in their technology roadmaps.
Landscape of Advanced Packaging
With respect to the real definition of advanced packaging, there are only 5 companies that are involved with high volume production of logic packaging techniques. These companies are TSMC, Samsung, Intel, Amkor, and ASE. The supply chain of these packaging materials is generally commoditized, although it is experiencing tightness due to shortages. Each of these companies employs a plethora of different packaging techniques. Areas where they generally differ are in 1. Substrate Material, 2. Size, 3. RDL, and 4. Stacking usages. Amkor and ASE are also involved in other OSAT activities.
Generally, where foundry and OSAT packaging techniques differ is the use of lithography defined redistribution layers (RDL) which are much dense than what most OSATs can produce and therefore, are much capable of accommodating complex wiring. These processes are referred to as fan-out wafer level packaging (FOWLP), fan-out chip on substrate (FoCoS) which the largest OSAT, ASE, is capable of processing. Lastly, fan-out system in package (FOSiP) techniques which are mainly used in smartphones, smartwatches, communications, automotives and more.
ASE latest packaging solution, called fan-out embedded bridge (FOEB), is largely similar to TSMC’s InFO-LSI and is pitched against Intel’s EMIB. However, ASE makes many claims against Intel’s EMIB technology but experts have noted the following arguments. 1. EMIB yields are not 80% - 90% as per ASE marketing materials, its close to 100%. 2. First Generation of EMIB technology has scaling limit in terms of the number of dies, but second generation of EMIB technology does not. Intel will be releasing a product with the largest package ever, an advanced package that is 92mm by 92mm which is the largest ever seen. FOEB does have some advantages in terms of density and package bump size using lithographically defined RDLs, but Intel retains the cost advantages.
The biggest showstopper is the CoWoS-S packaging technology by TSMC. This is the highest volume 2.5D packaging platform out there by a long-shot and it is used for Nvidia’s datacenter GPUs. Generally, most compute heavy chips with high bandwidth memory (HBM), such as AI chips, use CoWoS. Experts has noted the pervasiveness of CoWoS-S. Currently, demand far surpasses supply of CoWoS-S as Nvidia has gobbled up all the supply with this. “In Q3, Nvidia’s long-term supply obligations jumped up to $6.9 billion, including $1.64 billion of prepayments and another $1.79B of prepayments in the future.”
Where is Intel on the Advanced Packaging Landscape?
There are 5 main technologies that Intel currently markets itself as the leader for advanced packaging. 1. EMIB Technology (50 – 40-microns bump pitch), 2. Foveros Technology (50 – 36 microns), 3. Foveros Omni (25 microns), 4. Foveros Direct ( <10 microns) and 5. PowerVia (Back-end Power Delivery System).
Intel first used its Foveros technology in its Lakefield hybrid CPU SoC in 2020. Essentially, instead of stacking dies on another dies which are just a congregation of dense wires, Feveros involve both dies with active elements and utilizing 3D packaging technology, it created a hybrid CPU core architecture with big performance core and small efficiency cores.
The next Foveros product is the Ponte Vecchio GPU which is out in September 2022. This product includes 47 different active chiplets that are packaged together using EMIB and Foveros connected at 36-micron bump pitch. Most of all Intel’s client computing lineup will be using 3D stacking technologies codenamed Meteor Lake, Arrow Lake, and Lunar Lake. The first data center CPU using 3D stacking technology is called Diamond Rapids followed by Granite Rapids.
Other stacking technologies such as Foveros Omni and Foveros Direct differs in some technical execution but generally, the key difference is the reduction in overall bump pitch size.
Does Intel have a chance? Where is the moat Intel can leverage upon?
The key story of divergence here is the bet on Thermocompression bonding (TCB), one of which Intel has made it as the key technology driving all their packaging needs, and one that TSMC has not followed suit at all. This is essence, is an exact replica of what happened over the last decade with respect to pursuing the usage of EUV lithographic technologies. As well all know, ASML’s EUV technology was what helped advanced TSMC above Intel. Except today, the positions are flipped. TSMC has chosen to optimize and focus on its previous packaging technologies while Intel is investing in technological drivers. Will history repeat itself but except, the positions are flipped? I have no idea. But what this does provide is an opportunity, one where the odds are balanced rather than skewed.
TCB technology is currently used in all forms of HBM memories. To understand TCB technology, we must first understand the drawbacks of flip chip packaging which are described above. The key issue with flip back technology is related to something called coefficient of thermal expansion (CTE). As the entire package consists of many different materials, being heated up in a reflow oven causes the materials to expand at different rates. This is like baking a pie. The pie crust does not expand or heat up at the same rate as the filling. If it’s heated too fast, you risk burning the crust.
As the chip expands and cools, the difference in CTE will cause warpage which creates issues that can cause early failures or worse electrical performances.
Some examples of failures include AMD’s Fiji based GPUs. Although TSMC has partially resolved the issue by figuring out how to do a specific type of packaging technology called “Interposer-based packaging”, it has not been solved completely.
TCB solves this issue completely. In the appendix, I have placed an analysis from an expert who discusses this technique. But for the purpose of this paper, I will skip forward the technical explanations and move straight to the Pros and Cons. The Pros are: 1. It solved all the problems that are faced in manufacturing the latest high-end HBM chips that are used in AI training etc. Cons. It’s expensive. It costs about 2.5x more than the flip-chip packaging technique and can handle about 10x less volume.
Each TCB tool costs around $1.25Mil, and Intel owns around 300 of it. With the new $7 billion Malaysia Packaging Technology, it will own 2x the amount it currently own. Intel currently uses TCB tools in many of its non-advanced packaging applications. An Intel engineer has said that given the quality, reliability advantage and the yield gains, it far outweighs the miniscule, amortized costs of the tool per unit packaged in its high margin products.
These TCB tools give a lot of flexibility. It can use the same tool for standard packages, 2.5D packages, advanced 3D packages. Most importantly, the capabilities of TCB really start to shine as Intel moves towards Foveros Omni. Also, TSMC, Samsung and many others will not be able to do this sort of packaging unless it invests heavily into TCB. However, the key moat that Intel has is that it has been co-developing TCB tools for over a decade, just like how they had been co-developing and investing in EUV technologies previously. The key strength of Intel has always been its ability to involve itself with future generation technologies. Yes, Intel has made the wrong move to not move to EUV and decided to continue pushing its predecessor technology, but the same could be said for TSMC in the case of advanced packaging. As analyst David Kim best puts it “You can see why it is so easy to be fatalistic about an incumbent’s ability to resist disruption.” TSMC faces the same problem as Intel previous did – Should it risk everything on a new nascent technology that may not work out? Or should it continue optimizing its previous technology which seems to still have headroom to grow? Afterall, why stop what has not been broken? But as history suggests, TSMC may have made it first mistake which could potentially allow Intel to comeback.
3. Driver 3. Extraordinary CEO Execution (Future)
Intel has recruited ex-Intel engineer, Pat Gelsinger, who was the company’s CTO in 2001. Pat Gelsinger was later the CEO of VMWare from 2012 – 2021 but has rejoined back to Intel in 2021. In his introductory video, he spoke that he had “full support of the board members to pursue his agenda, and that if anyone expressed any concern, he would not take the job”.
In my opinion, Pat Gelsinger brings back the grovian-like culture in Intel. In fact, he spoke about this specific grovian-like culture in his opening conference where he revealed the “Intel Unleashed: IDM2.0” agenda. In his book, “The Juggling Act, Bringing Balance To Your Faith, Family and Work”, provides an inside look into Pat’s beliefs, outlook, and personality. But generally, the rhetoric around Pat Gelsinger return to Intel is noted generally positive, although some has felt that recruiting an engineer back to lead Intel may not be as charismatic of a leader as some would hope for.
Yet, one of the key characteristic that was mentioned by Andy Grove in his book in terms of what he considers as a sign of extraordinary leadership is the sign of humility, as humility breeds openness and foster communication. I have no proof that Pat Gelsinger will be the extraordinary CEO that can lead Intel out of this strategic inflection point. But at the same time, Pat Gelsinger is the one that elicits my confidence given the recent steps he has taken.
Supplemental Thesis and Considerations
Geopolitics Relationship and Considerations
Geopolitical tensions between United States and China are best described as a double edge sword for Intel. Why? Because the geopolitical uncertainty brings the large disparity and concentration of foundries in Asia to light and highlights the importance of Intel positioning in the global semiconductor landscape. However, at the same time, it could create a key barrier for Intel’s path to recovery, namely by blocking acquisitions such as Tower Semiconductors by Intel, just like how China has previously blocked Qualcomm from acquiring NXP Semiconductors in 2018. Generally, according to semiconductor geopolitical analyst, Craig Addison, who wrote the book “ Silicon Shield: Taiwan's Protection Against Chinese Attack”, described that global M&A deals often require the approval of various regulatory agencies such as the Federal Trade Commission in the US as well as the State Administration for Market Regulation (SAMR). SAMR killed the Qualcomm/NXP acquisition by not issuing regulatory approval for the deal. Note that Qualcomm was reliant on Chinese market for major revenues, and this is similar to the case for Intel/Tower Semiconductors, where both Intel and Tower Semiconductor derives 20-30% of revenues in Asia/China.
Why is Tower Semiconductor a potential catalyst to Intel’s path to recovery?
Firstly, we need to understand what Tower Semiconductor does, and what kind of deals Tower Semiconductor pursue. Tower Semiconductor focuses on lagging nodes, where recently the semiconductor shortage has been mainly concentrated on. By acquiring Tower Semiconductor, Intel is killing 2, if not 3 birds with 1 stone – 1. It is bolstering the US semiconductor supply chain, hence fitting the agenda set by politicians and 2. It is benefiting from high margin services as trailing edge fabs are mostly depreciated, 3. It is expanding its product portfolio by entering in RF CMOs, Industrial sensors, RF mobile and RF Infrastructure. Secondly, one of the key factors of what makes Tower Semiconductor a strategic asset is the type of deals it historically pursues. Tower semiconductor deals are mainly made in the form of joint ventures. For example, Tower Semiconductor is the fab partner for Panasonic, which is Tower’s largest customer at 25% of sales. There are around 5 other customers each accounting for 4 – 11% of sales. Joint Ventures by nature mean that the stickiness is generally higher due to the level of integration the two parties has set up. By acquiring Tower Semiconductor, even though at a premium of close to 50%, what Intel is essentially paying for is not just the 3 factors that I mentioned above, but an acquisition of a highly sticky, integrated, and major customer portfolio. This is a strategic asset where Intel can upsell its IFS services and provide both trailing and leading node solutions, a full-service shop which allows it to compete against TSMC.
Valuation
Intel valuation today stands at 142bil, close to a 50% decline from its peak in 2019. On a relative basis, Intel stands reasonably “cheap” – 10x normalized P/E, 8x normalized EV/EBIT, 1.4x P/B. Furthermore, from a replacement value perspective, its market book value is at least 20 - 30% above its accounting book value. Therefore, this means that the downside for Intel is quite limited, maybe a maximum 15- 20% downside from a replacement value perspective. If we assume a average cycle client computing segment (which is stickier due to reasons stated above), lower than average data center revenues, and the 3 IFS scenarios, there seems to be a high level of risk-reward.
Yet, I do not recommend Intel as an investment case yet because there are few major indicators that needs to fall in place before any forecast would be of a high likelihood. These major indicators are 1. Intel must adapt its IFS foundry system and allow customers to participate at internal nodes instead of 6 – 8 months later. 2. There must not be any significant change in upper management, especially if Intel fails to implement the ambitious 4 nodes in 4 years transition plan – more specifically, Pat Gelsinger cannot be removed from Intel. 3. Intel start delivering chips manufactured by TSMC without much issues and demand recovers. A bonus indicator would be the completion of Tower Semiconductor acquisition, of which would be a synergistic impact on the financial forecasts ahead.
Notice that I did not recommend using Intel regaining technology edge as a key indicator for looking at Intel for an investment opportunity. Technology is inherently volatile and ever changing. What is more important is to identify structural, underlying changes in an organization structure and competitive dynamics. To me, that would be the strongest investment thesis for Intel.
NOTE: Several thanks to external sources. To write up this paper, I have used several sources online, and I might have missed out some. But most notably, these contributors includes 1. David Kim for Scuttleblurb , 2. Dylan Patel - SemiAnalysis (read his advanced packaging stack), Semiwiki, Chip Wars, Only the Paranoid Survive, , loads of news articles, Intel’s reports, commentaries on reddit, twitter, and many many more writers whom I did not correctly attribute the credits here. This will be something that I will take note and give credit where its due moving forward as I write my papers.
Thank you for reading my first official blog post!
Ryan Ang.