Edited by Brian Birnbaum.
1.0 Intel is an Asset of Civilizational Importance 2.0 What AI Means for Intel
3.0 Culture and Catching Up with TSMC
3.1 Silicon Photonics: An Innovator´s Dilemma?
3.2 Structural Business Concerns and the Tower Acquisition
3.3 Foundry Locations and Capabilities
4.0 Financials 5.0 Conclusion
1.0 Intel is an Asset of Civilizational Importance
The world needs chips like it needs water and the supply can be compromised any day now. Intel stands as the only alternative source of supply at scale.
Taiwan Semiconductors produces ~1/3 of all new chips worldwide and nearly 100% of high-end chips. Its fabs are located in Taiwan, which China is increasingly threatening to invade. Human civilization is increasingly dependent on chips and should TSMC´s fabs be compromised by China, the world´s prosperity would be at stake. Nonetheless, the situation is highly nuanced because invading Taiwan would be catastrophic for China as well. Still, the issue is worth exploring on its own since it is one of the world´s most potentially disruptive single points of failure.
As Intel moves towards becoming a foundry for others to create chips, however, it is gradually becoming a real alternative to TSMC. Given the buoyant financial nature of semiconductor foundries (TSMC yielded 47.3% of its revenue in the bottom line in FY2022), Intel has a potentially lucrative opportunity ahead and could turn out to be an asset of civilizational importance. It is not new to the foundry business and has learned plenty of lessons from its fall from grace over the last decade. However, the business is complex and requires in-depth analysis.
Before moving on with the analysis, below, I provide context which is essential to accurately framing Intel´s situation.
Firstly, it is true that memory, and especially financial memory, is very short. Back in 2014—a mere 9 years ago—Intel had a notable manufacturing lead over TSMC. However, Intel´s culture had been degrading for some time, making the following two mistakes:
Passed on making chips for the iPhone because the smartphone market at the time was small. Intel was focused on making chips for computers and the business was highly profitable. Moving into the smartphone market would have required the company to embrace lower margins and, thus, the scenario had all the elements of the Innovator´s Dilemma. Smartphones then boomed and Intel missed it, giving TSMC more volume.
Relying on DUV (deep ultraviolet light) for its patterning (etching circuits on silicon), instead of EUV (extreme ultraviolet light). The former turned out to be a nightmare to work with in the 10nm domain. TSMC decided to go ahead with EUV which ended up working well and by the time Intel launched its 10nm process, TSMC was 2 nodes ahead.
I will spare you the pain of exploring in depth what went wrong inside Intel´s culture. In short, the company turned arrogant, and innovation withered. The moral of this story is that although it now seems that TSMC has been ahead forever, the truth is that its leadership is partly the result of Intel’s missteps, in addition to the hard and smart work that makes TSMC a formidable company: in the field of semiconductors, the tides can turn rather quickly.
It is somewhat difficult to analyze Intel without touching on geopolitics. A superpower like the US should have absolute control over its chip manufacturing and it knows this now. Last month I enjoyed reading the book “Chip War” by Chris Miller , from which my main takeaway is: nations are often compelled to back up their semiconductor industries and, in fact, most of the companies that dominate the space across the globe today are the result of said support, like Samsung and TSMC. As recent evidence, see the 2022 CHIPS for America Act, which appropriated $52B in funding to the industry.
Everything across our civilization is increasingly a function of processing more information, to make better decisions. The economy has relied more and more on storing data to train AI that can make useful predictions. Semiconductors are paramount—they are a matter of national security—and Intel currently stands as the only viable institution that can allow the US to eliminate the bottleneck brought about by the dependence on TSMC and, by extension, on Taiwan.
The above is one hell of a tailwind for Intel. However, foundry revenues only account for a fraction of Intel´s revenue today, with most of the company´s revenue coming from its client group. The competitive dynamics in the latter are cloudy and given Intel´s market cap today ($132B), initiating a long position now in the pursuit of the foundry opportunity requires a considerable faith in Intel´s ability to continue designing and producing competitive compute engines.
The above is by no means out of the question and going forward, I will be breaking down the analysis into two parts. Firstly, I will analyze Intel´s compute engines at a fundamental level. Secondly, I will analyze its foundry business and will attempt to bring it all together into a coherent thesis. Lastly, bear in mind that a lot can change in the coming years. My thesis is not intended to serve as a definitive point of view, but rather as a mental model which one can employ to observe the situation going forward.
2.0 What AI Means for Intel
AI requires computation acceleration and Intel´s product roadmap is looking mediocre, or rather, aimed at the low and medium segments of the market.
Analyzing semiconductor companies can be confusing because their typically broad range of products and respective exotic names can put anyone to sleep. However, it is my view that these products almost always stem from a fundamental product strategy that companies in the space draft and commit to on average once a decade. By understanding the strategy, the playing field becomes much easier to analyze, such as the case with AMD and its chiplet platform. In what remains of this section, I will break the Intel vs AMD dilemma into a simple and effective model.
AI is the next compute platform, and semiconductors companies are committing to product strategies built specifically for these products. AI is blurring the lines that have traditionally helped us distinguish between CPUs and GPUs, which are more of a marketing than a functional construct. AI only cares about who can store the most data, train the best models, and make the most accurate inferences . This is why we now see Intel and AMD announcing CPUs with deep learning accelerators - CPUs are increasingly relevant only if they can work with deep learning models, thus meaningfully entering the domain of GPUs.
Deep learning models consist of linear algebra on the way forward (inference) and differential calculus on the way back (learning, training). You can learn the specifics in my post, Artificial Intelligence 101. The way forward consists of multiplication and addition and the way backwards consists of derivatives, all performed in matrix format. Each deep learning model requires a specific set of computations, and a chip that is especially designed for a given model will always outperform others that are more generic in nature.
On the CPU side—which is especially relevant to Intel´s competitive position with respect to its rival, AMD—latency is the primary concern. CPUs cannot match GPUs when it comes to training models, but, when equipped with accelerators, they can surpass them in terms of overall functionality when it comes to making inferences. The more specifically an accelerator matches the computational needs of a given model, the lower the latency (time required to make an inference) and, thus, the higher the customer satisfaction (and usage outcomes).
There is a cost to specificity, so the key question is, between Intel and AMD, which has the best fundamental product strategy? And further, will they overlap as much as the traditional narrative suggests?
As I explained in depth in my AMD deep dive, the company acquired Xilinx to infuse its existing product lineup with the former´s expertise in FPGAs, which are like ASICs that can reconfigure themselves on the go and thus take on a specific shape to best suit any given deep learning model, but at a marginal cost (more about this here). On the other hand, Intel announced its new Advanced Matrix Extension (AMX) on its Architecture Day in 2021, which is a chip specialized to compute matrix-based operations.
Fundamentally, FPGAs are likely to outperform Intel´s AMX since they can optimize at the circuit level and thus are able to drive latency improvements at the edge. This is likely to make AMD´s products more suitable for higher-end applications, because naturally the specificity of FPGAs will come with a cost. In turn, Intel´s AMX will likely be more suitable for lower-end applications, like devices that sit at the edge. Additionally, Intel´s Habana chips are quite impressive, but are also ASICs and are thus static—it can be misleading to assess the performance of a generic chip on a specific model, or a set of them.
Despite the static nature of Intel´s Habana chips versus FPGAs, Gaudi2 displays quite phenomenal performance versus Nvidia´s A100. Resnet50 is a convolutional neural network, used for computer vision (the stuff that enables driver-less cars, object recognition etc).
Further, I doubt the future holds a world in which processors have just one accelerator. As AI becomes more pervasive, processors will need a multitude of engines to address a growing range and volume of workloads. Thus, connectivity will be paramount, further compounding the differences at the latency level between the two companies. AMD owes much of its success over the last decade to its Infinity Fabric, which enables the company to connect chiplets efficiently. Intel is only just getting started with chiplets now and, although it is headed in the right direction, it is far behind AMD.
Today, processors are connected to one accelerator (1), but soon they will be connected to many (2) and further, FPGAs will point at other FPGAs (3), to make inferences about what shape a given FPGA should take, to best make an inference for a particular model. As this happens, the companies that are able to connect computing units marginally better than others should gain an exponential advantage, as marginal latency deltas advantages aggregate to a considerable one.
Further, today´s deep learning models are static, in the sense that they do not train themselves. But they will soon be equipped with recursive reinforcement learning capabilities, whereby they will be able to tweak their parameters and even larger abstractions of the model autonomously. I believe that in this scenario, FPGAs will also gain relevance in the training side by enabling the models to evolve with less restrictions at the hardware level.
In aggregate, I believe over the next decade deep learning models will require an underlying enabling hardware platform that is as rapidly evolving as the models themselves. As far as I can see, FPGAs are best suited for the task, compared to a more static approach like Intel´s AMX. However, I have noticed a series of interesting things:
IoT seems to be picking up across the board, being the fastest growing business segment for GlobalFoundries, which also exhibits rather notable growth at TSMC. As many of you know, one of my top convictions (Blackberry, deep dive) envisions a future in which all things are connected to the internet. I suspect that technologies like Intel AMX may have have a place in this future, at the edge, where devices will likely lag in intelligence with respect to devices sitting closer to datacenters.
Since Intel´s 2015 acquisition of Altera, back then Xilinx´s main competitor, the company has not made many strides on the FPGA front. However, FPGAs are now back on Intel´s roadmap as standalone products and the firm is spending considerably more in R&D than its rival AMD: $17.53B (+15.39% YoY) and $5B (75.92%), respectively, in FY2022. With Intel´s cultural restart, which I explore in depth in the next section, it should not be too easy for AMD to run away.
Intel has a very large incumbency, with 62.8% of market share in the x86 CPU space as of Q4 2022, while AMD had 35.2%. This is a large distribution advantage, which will likely help the company counter AMD´s seemingly superior product roadmap.
In all, I get the sense that Intel´s product roadmap is currently pointing at a relatively mediocre future. However, the company´s distribution advantage together with its superior R&D allocation and reawakening culture (again, next section) make it more likely than not that it will at some point come out with relevant innovations. Further, I suspect that its current product roadmap has a place in the low to medium segments of the x86 market, which may sustain the division´s profitability going forward. Hard to tell as of yet.
3.0 Culture and Chasing TSMC
In the fab business, once you have the capital and the talent, success comes down to culture.
Pat Gelsinger spent 30 years at Intel before being pushed out by Paul Otellini in 2009, CEO at the time. If you spend time researching these two men, one stands out as a leader and the other as a manager. If you then look under the company´s hood during their respective tenures, it looks qualitatively different. From Otellini and up until Gelsinger, Intel looks manager-centric. From Gelsinger onward, it looks engineering-centric. In retrospect, the company´s poor performance over the last decade is largely attributable to its inadequate culture and the clash between Otellini and Gelsinger seems far more poignant than it may have in the past.
Intel 10 is a great case study to help understand the practical implications, since it has been plagued by notorious technical and manufacturing challenges that have caused repeated delays in its mass production. The technical challenges did not arise on their own, but rather as a result of organizational and cultural dysfunction, which started to brew at the beginning of Otellini´s tenure. His predecessor, which succeeded Andy Grove, was known for his technical expertise, his focus on quality and productivity, and his passion for education and research. According to Gelsinger and Dr. Ann Kelleher (Senior VP and GM of Technology Development at Intel), Intel 10´s difficulties stemmed from the following:
Intel took on too much risk at once, the jump from 14nm to 10nm being too big. It did not assess risk well and, further, seemed to lack contingency plans.
It made the wrong technological choice by deciding to pursue DUV which led to disastrous yields, instead of EUV which turned out to be the right pick. At the time it seemed like EUV was not feasible.
Most importantly, however, Intel did not run a parallel program for EUV, even though it actually drove the creation of it, by investing billions of $ in ASML, the sole supplier of EUV machines. The Intel 10 fabs were not designed to be flexible, so by the time it was clear that EUV was the way to go, it was not possible to install EUV machines in the fabs.
Intel also stopped doing OKRs (objectives and key results). OKRs can be traced back to Peter Drucker´s invention of MBO (management by objectives) in 1954, which Andrew Grove then transformed when he joined Intel in 1968 to the framework that is so widely known today. Incidentally, John Doerr later joined Intel in 1974 where he learned OKR and then transferred the knowledge to Google, to which the company credits much of its success.
If you go further back in time before Intel10, it also seems that suppliers felt generally patronized by Intel. Intel thought it was smarter than everyone else and largely disconnected itself from the ecosystem. This general sense of complacency seems to have set the stage for the difficulties experienced in Intel10.
It is clear that, over the last decade, Intel has followed the blueprint that almost always summarizes the fall of great companies, like General Electric for example. When they seem unassailable, companies of this sort are usually their own worst enemies and fall because people inside start to stifle innovation in favor of egocentric corporate ladder-climbing, with accountability and information flow typically stagnating. So, regarding the foundry, what has changed at Intel? Gelsinger has taken meaningful steps to fix all of the above, starting with restoring the company´s culture and then assigning Dr. Ann Kelleher to deliver the new processes as soon as possible and allocating plenty of capital to the task.
But Dr. Kelleher has a monstrous challenged ahead. Semi fabs are a game of variability management, in which the slightest deviation in one of the steps can render the final product useless. With EUV, this complexity is exponentially higher and understanding why sheds some light on the relative undesirability of entering a catch-up rat race with TSMC. Feel free kip ahead to the next paragraph if you are not interested in the details:
The EUV light source, which is generated by firing a high-power laser at a stream of tin droplets to create a plasma that emits EUV photons. The light source has to be stable, powerful and efficient to produce enough EUV light for high-volume manufacturing.
The EUV optics, which consist of multilayer mirrors that reflect EUV light at a very shallow angle. The optics have to be extremely smooth, flat, and defect-free to avoid scattering or absorbing EUV light. The optics also have to be protected from contamination and degradation by using vacuum chambers and pellicles.
The EUV masks, which are the templates that carry the circuit patterns to be projected onto the silicon wafers. The masks have to be made of special materials that can withstand high-energy EUV photons and have high reflectivity and contrast. The masks also have to be aligned and exposed with nanometer accuracy to ensure high-resolution patterning.
The EUV resists, which are the photosensitive materials that react to EUV light and form the desired features on the silicon wafers. The resists have to be sensitive enough to capture fine details with low exposure doses, but also robust enough to resist line-edge roughness, pattern collapse and other defects.
The EUV metrology and inspection, which are the tools and methods that measure and monitor the quality and performance of the EUV process and products. The metrology and inspection have to be able to detect and correct any errors or variations in the EUV light source, optics, masks, resists and wafers with high speed and accuracy.
Intel is promising to catch up by 2025, but even the world´s top experts remain perplexed at the aspirational nature of such a goal. In FY2022, TSMC spent $5.3B in R&D, up from $4.5B in FY2021, while Intel has not specified how much of its $15B+ R&D budget it is assigning to foundries.
TSMC´s greatness stems from its engineering-centric culture. Now that Intel is allocating the resources and is fixing its culture, catching up is not out of the question. It can definitely give TSMC a run for its money. Perhaps the point is not to determine whether Intel will catch up to TSMC by 2025, but whether it will at some point over the next decade and thus, whether owning a part of the (foundry) business is a sensible thing for a semiconductor investor to do. As I mentioned, a differential rat race is not something that I like to participate in, but I see a potential wildcard for Intel in the horizon: silicon photonics.
3.1 Silicon Photonics: An Innovator´s Dilemma?
A new and emerging technology, silicon photonics, presents Intel an opportunity to get ahead of TSMC in a less arduous manner.
Silicon photonics could be the technology that enables Intel to get ahead of TSMC over the next decade or two, without necessarily engaging in a tight rat race. Silicon photonics consists of using light instead of electrons, to move information around in chips. The advantage of using light is that it is faster and generates less heat and is theoretically a more efficient medium to deal with the ever exponentiating volumes of information that the future is going to bring.
The problem with this technology is fundamentally that silicon absorbs wavelengths shorter than 1.1 μm, which excites the silicon’s electrons (because they absorb the energy from the light), causing them to create noise and ultimately efficiency issues, which lead to heat dissipation and variability in the circuit´s behavior. We ideally want to pass photons with wavelengths much smaller than that in a chip; the shorter the wavelength, the more information a given photon can carry, and, thus, the more applications with the potential to be unlocked.
To me, it seems the technology is not impossible to bring to the mainstream via sufficient iteration, and Intel has been a leader in the space for quite some time. Intel has been researching and developing silicon photonics since 2004 and has launched a range of products that offer 100G, 200G and 400G per-second optical transceivers for data center and 5G applications. (A transceiver turns light signals into electric signals and vice versa, and it makes a lot of sense for hyperscalers to use in datacenters to move information from one server to another in the form of light).
Meanwhile, the size and success of TSMC’s current business presents questions as to the viability of competing with Intel in photonics—although it seems to be working on some kind of collaboration with Nvidia. As the tech matures and goes mainstream, it could be that Intel builds a lead that is then hard for TSMC, or any other fab, to catch up with. Intel already produces its silicon photonics transceivers in its own 300-millimeter silicon wafer fabrication facilities. This vertical integration may well compound into something meaningful. Though some time away, it’s definitely worth watching going forward.
3.2 Structural Business Concerns and the Tower Acquisition
There are some things that seem inherently wrong with Intel´s current approach to doing business.
Beyond Intel getting back on the innovation bandwagon and the emerging silicon photonics window, there are some fundamental and obvious issues with the business:
Intel ´s foundries would serve direct competitors, like Nvidia and AMD. This creates a conflict of interest and not for the first time in the company´s history. Amazon has found a quasi-legal manner to step over the line with its collaborators given the incentives, for instance. What stops Intel from moving in this direction down the line?
Now that Intel will manufacture its own chips and those of other companies, what will its best engineers want to work on? Designing its proprietary chips or building the fabs out? Per Gelsinger´s comments, it seems that many of the 21,300 employees Intel has hired in the last couple of years are particularly motivated to solve the manufacturing bottleneck described earlier. Can Intel compete in chip design when there are companies uniquely focused on the task, for instance? Gelsinger talks of a flywheel forming between the chip design and manufacturing divisions, but there is also the risk for a bifurcation of attention or even worse, the proliferation of unhealthy competition.
One of the reasons Intel´s foundry efforts failed previously is that it treated its customers as “second-class citizens.” Some analysts and industry observers have criticized Intel’s foundry services in the past for being too focused on Intel’s own needs and priorities, rather than those of its customers. For example, some have argued that Intel’s foundry services were mainly used as a way to fill its excess capacity and generate additional revenue , rather than as a strategic business unit that could attract and retain customers with competitive offerings and services. Some have also suggested that Intel’s foundry services were not willing or able to accommodate the diverse and complex demands of its customers, such as different architectures, IP cores, packaging technologies and design tools.
Intel seems to be repeating the above mistake per some articles. It seems that Intel 4 is mainly designed for Intel’s own products, such as Meteor Lake CPUs and Ponte Vecchio GPUs, while Intel 3 will be the first new foundry offering for external customers, along with the existing Intel 16 node. Intel 4 is expected to be ready for high-volume manufacturing in the second half of 2023, while Intel 3 is expected to be ready in the first half of 2024: will their customers tolerate being last again?
Regarding point number four, I am not entirely convinced the situation is black and white. For instance, Apple seems to get preferential access from TSMC, in the form of prioritizing the orders of the former, for instance, but this does not deter other customers from partnering with TSMC. One fundamental question is whether the industry has changed enough so that chip designers have far more tolerance with their manufacturing partners for want of any practical alternative. AMD only spun GlobalFoundries off in 2009, initiating its fabless journey less than fifteen years ago.
Secondly, Intel signed a definitive agreement in February 2022 to acquire Tower Semiconductor, for $5.4B all cash. The transaction has not closed yet as of April 2023, but upon it closing I expect Intel to gain the following things from the acquisition:
A customer-first mentality on the foundry side.
PDK (process design kit) expertise, which Tower has developed by successfully interfacing with customers in the consumer, automotive, mobile, infrastructure, medical, aerospace, and defense industries. PDKs make it easier for customers and fabs to interact and get the job done. They are used by fabs to provide their customers with the necessary data and tools to design integrated circuits (ICs) that are compatible with their manufacturing processes. PDKs typically contain information such as design rules, device models, layout cells, simulation parameters, verification scripts, and documentation. PDKs enable chip designers to use various electronic design automation (EDA) software platforms to create and verify their designs before sending them off—in this case, to Intel—for fabrication. PDKs are essential for ensuring the quality, performance, and yield of the ICs produced by the fabs.
One great test for Intel going forward is to what extent it is able to integrate the acquisition without smothering the above. Certainly, without a customer-first mentality its foundry services are likely ill-fated. And yet, Intel does not have the expertise to interface with customers across a broad range of industries like Tower does. The acquisition needs to come with this capability.
3.3 Foundry Locations and Capabilities
Despite the various cons in the Intel thesis, the below sounds much better than a fab at a stone´s throw away from the CCP.
I will use this section to list the location and capabilities of Intel´s foundries and briefly comment on its pivot to chiplets. Intel’s fabs are capable of producing processors and other chips using advanced process technologies ranging from 22nm to 7nm, as well as packaging technologies such as EMIB and Foveros. Some of Intel’s major fab locations and their capabilities are:
Arizona, USA: Fab 12 (22nm), Fab 32 (14nm), Fab 42 (7nm).
California, USA: D1D/D1X (14nm/10nm/7nm development), Fab 68 (65nm).
New Mexico, USA: Fab 11X (14nm).
Oregon, USA: D1C/D1D (14nm/10nm/7nm development), Fab D1X(14nm/10nm/7nm), Fab D1X Mod2 (7nm).
Ireland: Fab 24 (14nm), Fab 14 (22nm).
Israel: Fab 28 (10nm), Fab 10 (22nm).
China: Fab 68 (65nm).
The above list does not include the new foundries that Intel is building as part of its IDM 2.0 strategy to expand its manufacturing capacity and foundry services. Intel has announced plans to invest over $20 billion to build two new fabs in Arizona, which are expected to be operational by 2024. Intel has also announced plans to invest over €33 billion for R&D and manufacturing in Europe, spanning France, Germany, Ireland, Italy, Poland and Spain. The yet-to-close Tower Semiconductor acquisition will add eight more fabs in Israel, the U.S. and Japan to its portfolio.
According to Intel’s process technology roadmap, the company plans to release its products using the following nodes in the following time frames:
Intel 4: Manufacturing-ready in the second half of 2022 for products shipping in 2023, including Meteor Lake for client and Granite Rapids for the data center. Intel 4 delivers an approximate 20% increase in transistor performance per watt compared to Intel 7.
Intel 3: Manufacturing-ready in the second half of 2023 for products shipping in 2024. Intel 3 delivers a further 18% performance per watt improvement compared to Intel 4.
Intel 20A: Manufacturing-ready in the second half of 2024 for products shipping in 2025. Intel 20A ushers in the angstrom era with two breakthrough technologies: RibbonFET, Intel’s first new transistor architecture since FinFET in 2011, and PowerVia, an industry-first new backside power delivery method.
Intel 18A: Manufacturing-ready in early 2025 for products shipping later that year. Intel 18A will be another angstrom node with refinements and optimizations to RibbonFET that will deliver another major jump in transistor performance.
***Note that the numbers used in the names of process nodes do not actually accurately describe the underlying technical specifications. They are marketing labels.
On the Q4 2022 ER conference call, management mentioned some accounting changes that will increase the “estimated useful life of certain production machinery and equipment from five years to eight years.” Management expects the IFS deal pipeline to “extend the life of manufacturing nodes beyond what was practical within IDM 1.0”, as they move towards a disaggregated CPU architecture that allows them to leverage older and newer process nodes alike: a.k.a. chiplets. Better late than never and for a deep dive on chiplets, feel free to read the first section of my AMD deep dive. Doing so will give you a fundamental understanding of what is going on the in the semiconductor industry, which will be valuable to you going forward.
4.0 Financial Considerations and Valuation
Intel is priced for reasonable execution going forward, while the asymmetry would arise from getting its products for free and its foundries at a reasonable price.
The product roadmap looks second-tier to me, so I would frame a hypothetical investment in Intel as follows: get the products for “free” and the foundries at a reasonable price. The closer we get to that, the more Intel will actually be a “WW3 hedge” (though of course hope large scale conflict between the US and China never actually erupts). Intel is currently valued as follows, as of the 14th of April 2023:
P/S ratio of 2.08.
P/E ratio of 16.47, which is just under the Nasdaq´s average of 17.91 and SP500´s average of 18.97.
As you can see above, most of Intel´s profitability stems from its products. Thus the valuation is only justified if one deems it likely that the quality of the products or their price to quality ratio will improve. I believe I have a blind-spot in terms of assessing the lower- and mid-segments of the chip market and thus what traction Intel´s current roadmap may achieve. Further, Intel has a very large R&D budget. Assuming its talent isn’t distributed unevenly, with larger lumps found in the foundries, renewed focus on culture is likely to yield positive surprises going forward.
The theoretical framework that I outlined in Section 2.0 encompasses the Client Computing, Data Center and AI, Network and Edge and Accelerated Computing Systems and Graphics. These all consist of energizing the most electrons per second at the lowest cost, and accelerated optimization of latency is paramount to all the products contained in these segments. Naturally, depending on your view of the framework, Intel may already be an attractive hedge, but I believe it is not quite there yet.
A situation of this sort may take years to evolve favorably in the investment sense. The conflict may take what remains of this decade to fully scale up, and in that period of time, Intel´s product lineup may or may not prove to gain traction. I believe that, over the coming years, as AMD´s FPGA mesh comes online, we will gain much visibility on the matter. In the meantime, I more than welcome Intel continuing with its foundry efforts.
Intel´s revenue has been falling recently, but it seems largely attributable to the overall macro decline. Throughout FY2022 consumers have cut back their spending on digital devices across the board and organizations have similarly slowed or halted their spending on their digitization efforts. I believe that it is tempting to blame Intel for this decline, but it seems rather early to do so just yet.
The company is not getting any leaner with Gelsinger stepping up as CEO in February 2021, but I welcome the fact that much of the increase in OPEX comes from a rising R&D spend. As mentioned near the beginning, all the above is merely a mental model to start thinking about Intel, because the situation is highly dynamic, given the nature of the industry. As Intel continues to allocate more capital to R&D and work on its culture’s return to engineering and customer-centricity, its fate can change. They are late, but this is a marathon, not a sprint.
One danger to the thesis is the rather dire balance sheet, which has been worsening steadily over the past decade. While cash from operations and free cash-flow have remained buoyant throughout this time, they are both taking a hit with the decline in the top line and the uptick in capex. Much more spending is to come as Intel continues to build out its foundry infrastructure. In this context its product lineup looks less like a bonus for someone buying the stock for the foundries, and more like a necessary cash cow as Intel makes its turnaround.
I have enjoyed gaining an architectural overview of what Intel has planned for its products, but unless I am mistaken, their roadmap is not exactly avant garde. If AI is the next computing platform, delivering hardware that is able to morph and adapt to its needs at a marginal cost and after deployment is vital.
However, with its enormous R&D budget and its now increasingly healthy culture, it looks to me like a budding innovation ground. It has the potential to combine its otherwise mediocre lineup of products with ones that satisfy the above demands. Further, silicon photonics presents itself as an opportunity for Intel to get ahead of the pack in the coming decade.
For me to get involved with Intel, I need a much higher degree of asymmetry than the current valuation grants. However, this depends on your view of their current product lineup. As I have mentioned, the situation can vary dramatically in a relatively short period of time: both geopolitically and technologically.
In the meantime, I remain on the sidelines, eager to see how the mental model that I have described in section 2.0 plays out. I continue to hold my AMD position (long since 2014) based largely on this model. I believe the world is going to change very fast over the next decade and that, yet again, computing is going to take an even more prominent role in our lives.
Until next time!
⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.
You can also reach me at:
amazing research, I dont' see any competitive advantage in the cultural approach of Intel so far..
This is very well researched. Amazing stuff!