Investing Mistakes, Chapter 1001
Investing is a difficult, humbling pursuit. I am often asked about mistakes I have made. While the list of my mistakes is quite long, I thought I would write up my most recent mistake of commission, i.e. a mistake based on an action I took that resulted in a loss of capital as opposed to a mistake of omission that resulted in missing out on potential gains. I have been following semiconductors for 20 years and had never been bullish on Intel until June, 2020 when I bought the stock and made it a significant position for the first time in my career. This was a mistake.
I had owned AMD throughout 2019 and into early 2020, but was generally disappointed in their underlying fundamental performance. Their share gains in server CPUs in particular were slower than I expected based on the performance per watt per $ shown in third party benchmarks. I did a lot of work to understand potential reasons for this and kept coming back to a paper written by Urs Hölzle “Brawny cores still beat wimpy cores, most of the time” in 2015. Urs Hölzle was Google’s 8th employee and is currently a Google Fellow and its SVP of Technical Infrastructure, where he oversees the design, installation, and operation of the servers, networks, and datacenters that power Google’s services. The conclusion of this paper is that single core and single threaded performance is often the rate limiting factor even for highly parallelized workloads:
“Even though many Internet services benefit from seemingly unbounded request- and data-level parallelism, such systems aren’t above the law. As the number of parallel threads increases, reducing serialization and communication overheads can become increasingly difficult. In a limit case, the amount of inherently serial work performed on behalf of a user request by slow single-threaded cores will dominate overall execution time…although we’re enthusiastic users of multicore systems, and believe that throughput-oriented designs generally beat peak-performance-oriented designs, smaller isn’t always better. Once a chip’s single-core performance lags by more than a factor to two or so behind the higher end of current-generation commodity processors, making a business case for switching to the wimpy system becomes increasingly difficult because application programmers will see it as a significant performance regression: their single-threaded request handlers are no longer fast enough to meet latency targets. So go forth and multiply your cores, but do it in moderation, or the sea of wimpy cores will stick to your programmers’ boots like clay.”
As anyone even vaguely familiar with the x86 CPU market understands, AMD has significantly more cores than Intel and overall superior performance per watt per $, but the AMD cores are “wimpier” than the Intel cores and slower on single threaded performance. It was fascinating to me that despite being an entire node behind on manufacturing (Intel at 14 nm vs. AMD at TSM’s 7 nm, which is equivalent to Intel’s 10 nm), Intel was still ahead on single core and single threaded performance.
Beyond this single core, single threaded performance advantage, Intel has the “exorbitant privilege” of having all x86 software code effectively optimized for their own architecture. No dominant digital processor company has lost the #1 market share position over the last decade because of this privilege. What this means in practice is that software just runs faster on Intel processors than AMD processors — it is not enough for AMD to be 20% faster in benchmarks, because real world software optimizations by developers mean that software generally runs faster on Intel. AMD needs to be much faster for a sustained period of years for software to be rewritten and optimized for their own particular architecture. As a sidenote, this is why Windows on ARM did not work for PCs and ARM server CPUs have been slow to take off — any ARM CPU needs to offer dramatically higher performance per watt per $ to justify rewriting software so that it can run on ARM instead of x86. On top of the “exorbitant privilege” conferred by software developers optimizing for Intel’s own architectural flavor of x86, Intel is much better at writing their own software — compilers, frameworks — than AMD and all of this software is essential for other software to run well. Having followed the GPU market for 20 years, I can confidently state that software has never been a strength of AMD and is the main reason they lost to Nvidia. In the CPU market, Intel’s inference software frameworks are far superior to AMDs and inference is a growing workload for CPUs.
Given better single threaded, single core performance, better overall software and the “exorbitant privilege” of having most of the world’s software optimized for their architecture, I was becoming intrigued with Intel given a dramatic period of underperformance vs. other semiconductor companies. The main risk I focused on was manufacturing. Specifically, being able to successfully ramp their 10 nm++ process. Intel had led the world in semiconductor manufacturing for decades and always got to each new node before their competition. Given Moore’s law and learning curves, this gave Intel an enormous performance and cost advantage on top of the “exorbitant privilege” of having software optimized for their own particular flavor of x86. As I wrote in January, 2019, Intel lost this crown to TSM by being over 4 years late to 10 nm which led to TSM fully ramping their equivalent 7 nm node before Intel even really began to ramp 10 nm. This seismic shift profoundly advantaged all of TSM’s customers who compete with Intel: AMD, Xilinx and Nvidia.
Monocausal explanations are dangerous, but in my opinion Intel lost their manufacturing lead to TSM because of arrogance. While Intel effectively worked with all semi-cap equipment co’s at some level over the last 20 years, they tended to spend disproportionately more with a variety of #2 semi-cap equipment companies where Intel was the overwhelmingly largest customer — this gave Intel pricing power and effective control over the company. Unfortunately, the #1 semi-cap equipment companies generally steadily increased their lead over the #2 players as their largest customers (TSM, Samsung) benefitted from the enormous ramp in Smartphone volumes which also shifted the learning curve advantage from Intel to TSM and Samsung. This led to more revenue, more resources and a greater technological lead for the #1 player that TSM and Samsung relied upon relative to the #2 player that Intel generally relied upon (please note that this is a very broad generalization and I am aware of the exceptions). Nowhere was this more apparent than in lithography, which is arguably the most important part of semiconductor manufacturing. ASML’s technological gains in lithography relative to Nikon, Intel’s preferred lithography supplier, over the last 20 years have been astonishing. The manufacturing race basically came down to Intel and Nikon vs. the rest of the semiconductor industry and ASML. Intel’s belief that they could win with second tier suppliers like Nikon was arrogant.
This disadvantage in terms of ecosystem scale and technology was exacerbated by poor decision making at 10 nm. Specifically, Intel decided not to insert EUV at 10 nm and instead rely heavily on multipatterning. This was effectively a message to the world: “We are so good that we do not need EUV. We can do this with DUV, which no one else can.” On top of this, Intel tried to scale aggressively at 10 nm. Moore’s law is actually not constant. The scaling at each node is different. Intel tried to scale transistor density by 2.7x at 10 nm rather than the more common 2x to 2.4x without using EUV. I have tried to to think of a good analogy for this and the best one I can come up with is that Intel attempted the world’s highest pole vault in a torrential rainstorm while blind folded with no mat on the other side.
The result was disastrous. Intel’s 10 nm was effectively more than 4 years late. However, I thought it was encouraging that Intel admitted all of their mistakes at 10 nm and for 7 nm, Intel committed to working more with the external ecosystem of suppliers, inserting EUV and scaling more conservatively. They might never again have a manufacturing lead, but they might be able to get to parity with TSM and despite being behind on manufacturing technology they had not lost significant CPU revenue share to AMD given their superior single threaded/single core performance, better software and the “exorbitant privilege” of having 3rd party software optimized for their own flavor of x86. So I was really focused on whether or not the optimized 10 nm++ process would work for their Tigerlake laptop CPUs in late 2020 and their Sapphire Rapids server CPUs in 2021 as this would bring Intel from 2 to 3ish years behind TSM to 1 year behind with the hope of reaching parity when Intel ramped 7 nm. I didn’t focus on 7 nm as I thought Intel would be inserting a much more mature EUV machine than TSM was able to work with when they inserted EUV along with the more conservative scaling targets. As anyone familiar with the story knows, this was my critical error. Regardless, I became comfortable that 10 nm++ would work and that Intel would be able to ramp Tigerlake and Sapphire Rapids.
Beyond this, Intel was an optically cheap stock with a reasonable dividend yield and I was becoming uncomfortable with some of the more extreme valuations in technology. Intel traded roughly in-line with Oracle and Cisco, but had actually grown revenue and GP $ faster over the last 5ish years. Intel also had recently passed an important milestone with data center accounting for a majority of both revenue and GP $. The faster growth of data center and its steady increase in Intel’s revenue mix meant that Intel’s overall growth should accelerate all else equal. Intel also played into my excitement about the higher computational intensity of AI relative to software written by humans — while less positive for CPUs relative to GPUs and memory — AI was still powerful for CPUs. Capital allocation was improving with the divestiture of the modem business (exorbitant privilege worked against them) and acquisition of Habana, which I was quite familiar with and very bullish on. It worried me a little that the Chairman came from the Med Tech industry and the CEO was a former CFO, but I actually thought the new CEO was making good decisions overall.
None of this mattered. Intel delayed their 7 nm program by over a year and may need to use TSM for more of their manufacturing. This is an indicator of the magnitude of the problem, as once a company stops trying to compete at the leading edge they can never get it back. It is still unclear what happened at 7 nm. Rumor is that Jim Keller, one of the greatest semiconductor architects of the modern era, went to CEO Swan with an ultimatum in June that either he or Murthy Renduchintala had to leave. If true, given that Swan chose Murthy — a terrible error — then it is almost certain that Swan did not know about the 7 nm delay when he made this decision in early June. The unprepared tone of the second quarter call certainly suggested that the 7 nm delay was a very recent revelation to the CEO and CFO. It does seem as if Intel attempted a new transistor design called GAAFET at 7 nm, which would have been extremely difficult to do concurrently with EUV insertion. If true, this means that Intel repeated their mistakes from 10 nm at 7 nm rather than learning from them. These successive mistakes at 10 nm and 7 nm are strange as one of the greatest nodes in Intel history was 22 nm which had conservative scaling targets, ramped quickly and had fantastic yields. Better to be on time with great yields with less aggressive scaling than late with terrible yields with more aggressive scaling. Strange that Intel did not learn from all of these experiences.
The 7 nm delay means that software will likely begin to be optimized for AMD and that Intel may lose their “exorbitant privilege.” And as an IDM, Intel has enormous fixed costs and fabs to fill which creates significant risk if market share really begins to shift. To add salt to the wound, AMD then finally put up a great quarter in Q2 with clear evidence of server CPU share gains. It is always the bullet you don’t see that gets you. I just did not focus on 7 nm. This was not an “unknown unknown,” which often lead to mistakes of commission. This was simply a variable I did not focus on due to an incorrect assumption that I did not adequately diligence. None of this is meant to be an excuse, only an explanation and perhaps some learning out loud. As one of my mentors is fond of saying, “There are only two things in investing, numbers and excuses. And if you don’t have the first, no one cares about the second.” Investing is ever humbling.
However, the pain I feel about this mistake of commission is trivial relative to the pain I feel about my mistakes of omission. The stocks I did not buy in March 2020 cause me much more pain and represent much more significant errors in my mind. In, “I, Claudius,” when Claudius writes of the death of Germanicus, he says “of my own sorrow, I do not trust myself to write.” That is how I feel about several stocks that I carefully considered buying in March of 2020 and did not. A 1% position in almost any of them would have generated 3–6x more profit than was lost on Intel. Mistakes of omission are always more painful than mistakes of commission for a growth investor. Investing is ever humbling.
I am not a golfer, but there is a phrase often used by golfers that I think applies to investing: “play in the present.” A golfer can evidently be so consumed and frustrated by a mistake that they made earlier in the round that it impairs their play for the rest of the round. I think the same goes for investing and it is important for investors to “play in the present” rather than being consumed by their mistakes. One must learn what there is to learn and then move on. In the case of Intel, I am not sure that there are generalizable learnings — I am an avid user of checklists, premortems, designated skeptics, falsifiable hypotheses and certainly knew the stock and industry well. Perhaps the generalizable learning is to simply double check every assumption, but this is something I have learned many times and a mistake I will likely make again.
And speaking of mistakes, as an American, I think it would be an enormous mistake for America not to do everything possible to help Intel stay at the leading edge with its fabs in America and Israel. If Intel were to outsource manufacturing to TSM, Taiwan would be the most geopolitically important country in history. It would be no understatement to say that “Taiwan would be the new Arrakis” per Epsilon Theory. Modern semiconductor manufacturing is at least as important to the economy as oil was in the 1970s — at least in the case of oil, it was available all over the world albeit at higher prices than in the Middle East. If the overwhelming majority of leading edge semiconductor manufacturing is concentrated in Taiwan with the rest in South Korea, then the geopolitical implications are significant as there is no cost curve with leading edge semiconductor manufacturing — either you can do it or you cannot. It would be as if the Middle East was the only place in the world with oil rather than simply the region with the lowest costs. And it is strange that that the Chairman, CEO and CFO of Intel have zero technical expertise in semiconductor design or manufacturing at arguably the most critical time in the history of Intel.