Fabricated Knowledge Quarterly Review
This is the first edition of a quarterly summary piece to highlight my ideas and best pieces quarterly.
This is not an April Fool’s post. Sorry, serious business only in semiconductor land.
Anyways, one of the most significant pieces of feedback I get is, “What are your favorite ideas?” and “How do I understand all of the jargon in the semiconductor industry?” Well, I finally listened to the feedback, and I thought I would do a quarterly post recapping my current pieces and takes and highlighting what I currently like. Consider this a meta piece of my content.
Semiconductor Performance
This is one of the 16th-best quarters in SOXX since 2002. While it’s a pretty solid showing at a 17% quarterly return, this doesn’t compare to the incredible returns in 2020 and 2023. Now, the misleading thing is that SMH, a broader ETF, is up 30%. That’s primarily due to the Nvidia weighting. Here’s the YTD performance of a large watchlist I keep. Nvidia is the leader year to date.
Diving deeper, automotive companies (AEHR, WOLF, MBLY, IFX, ACLS, INDI, STM) and companies in troubled subsectors are the biggest losers. Meanwhile, the winners are all AI-focused, with companies like Nvidia, MU, AWE, ANET, AMD, and others leading the way.
Here’s my cycle timer of where we are in each end market. AI continues to rip while no meaningful inventory can be found. Memory is now on fire, with price increases and declining inventory, and the general data center is now modestly positive. Smartphone sales are not as strong, while consumer and networking businesses are about to grow soon. Additionally, I think Automotive and Industrial are closer to a trough.
Now for the content overview this quarter.
Important Content this Quarter
This quarter is going to be AI-focused. I think my most essential pieces reflect this reality. My favorite piece is the macro-level talk about supply and demand for memory companies and why HBM will fix almost all aspects of the supply problem for memory.
After writing this piece, Micron reported and confirmed everything I said to be accurate and better than expected. The ongoing memory saga will be interesting, and given HBM's dominance, I think we will hit meaningful new highs this cycle.
Of course, it would be a mistake not to talk about the 800-pound gorilla of semiconductors, Nvidia. I thought this year’s event was mostly expected, but I appreciated and liked some of the new product announcements. I wrote about the entire event here and reframed what “AI is.” In this framing, I think it makes a lot more sense as just one part of a more significant trend.
And I won’t be remiss in mentioning the “what inning it is” post I did. I want to highlight that while AI might be a scorching hot subject, penetration compared to the internet is much closer to 1995 than it is to 1999.
My annual outlook so far looks alright, and instead of growing pains, this looks like the first year AI is stepping into its own. AI is growing up very quickly.
Now, amid all the talk of the new King of Semiconductor companies, I would be remiss not to mention the old king, Intel. Last year, it was one of my favorite ideas, but this year, I think it’s time to move on. Intel’s process turnaround is here, but now it’s time to deal with the hard reality that it’s still primarily an x86 company.
Share shifts remain unfavorable, as they were designed out of the DGX system by the new Grace CPU chip from Nvidia. Is there a place in the new world for Intel’s CPUs? I don’t think so. However, the process technology could save the company. The problem is that it’s quite some time from now.
Anyway, I will continue with my favorite ideas and thoughts.
My Continuing Favorite Ideas and Thoughts on Stocks
I want to start by saying that I can hold multiple ideas simultaneously. I continue to be long-term bullish on many companies (predominantly secular winners like Nvidia), but I also keep some caution. I wrote about why I think that Semiconductors are due a bit of a pullback given how much momentum is in the space.
Now, I think I have some talent here, as I called it bearish on the absolute peak year to date for the space. I am unsure if this continues, but this run has become a bit exhausted by many metrics. I believe we are due a bit of a breather in both broad indices, especially in semiconductors. I would say I am tactically cautious.
Now, speaking of which, I am quite a fan of the Axcelis short. I don’t recommend shorting, but this one seems interesting if it’s your job. Chinese demand might hold up, but Axcelis assumes zero share, and that utilization doesn’t grow.
They have three huge headwinds against them: Chinese WFE, Chinese Competition, and Tool Utilization. And I think each can get much worse. Read here:
Now, I will tell anyone who listens that I am not quite a fan of Microchip. They screwed up the cycle about as drastically as it gets, and I think there’s a bit of a funny curse on this stock. I wrote about why I wouldn't say I like Microchip and the analog names here. Multiples are rising, and it’s hard to underwrite them. Now note that in the long term, I like this, but as I wrote in the factor piece, I think this trade is a bit blown out.
These companies will not benefit much from AI, and estimates are already coming down, yet shares won’t. Multiples cannot go up forever, and I think that on a relative basis, I will continue to like this relative short to SOXX.
Read more in my Microchip section in this post.
Last but not least, I continue to love the subsystem trade, especially MKSI. They are worth a quarter or two of outperformance relative to the semicap players. There’s an inventory dynamic that should show up in street numbers by the end of the year as Semicap ramps for a significant 2025 year.
What always happens is when the inventory builds, subsystems do well, and then gross margin and operating leverage kick in. MKSI will see cycle highs in share price, and that’s 50% higher from today.
Before I go, I want to take a victory lap and “close” an old post, SiC as a dog. In that piece, I said that shorting AEHR made a lot of sense, and today, shares are ~12 dollars. That’s quite the return, but I would probably close any remaining short.
Last, given my stampede of positive Nvidia content, I continue to like Nvidia in the long run. I’ll have a post about this coming out soon.
Jargon Learning Quarter
Since I plan to do this quarterly and just finished OFC, we will go over some of the essential jargon for networking specifically.
LPO (Linear Pluggable Optic): This method uses SerDes in the host device to skip a DSP in the pluggable transceiver. DSPs are great for error correction and interoperability but come with greater power. Given that power has become a huge constraint, LPO has become a proposed technology to delete the DSP and save on system power.
This technology might never take off, but it is an interesting alternative to Marvell’s dominance in the pluggable DSP business.
LRO (Linear Receive Optics): This was all the talk this year, and I believe it will effectively be the solution to lower the DSP tax in the data center. The goal here is to remove one of the DSPs in a pluggable module, with a linear receive and a DSP transmit in the line.
That means from a server box, the pluggable transmitting the signal will have a DSP to error correct and clean up the signal, but the receiving end will be DSP-less. This is pretty much only available because of how good the Broadcom Tomahawk 5 SerDes is. It can deal with a lot of noise and is probably the best SerDes in the game today.
PAM4 (Pulse Amplitude Modulation 4-Level): This is a type of signal modulation that is different from NRZ (Non-Return-to-Zero). This can be used over both Copper and Optical links, and has 4 voltage levels, and thus increased bandwidth over NRZ.
PAM4 is primarily used for shorter transmission links and can be used up to 10 kilometers. There’s a high signal-to-noise ratio, so there’s a bit of a penalty for longer-reach links. PAM4 is the workhorse of inside-the-center transmission, while Coherent optics are prevalent over longer-reach links.
Today, the fastest single link of PAM4 modulation is up to 224 Gbps. Optical transceivers are often multiple links connected, such as 8 x 100 Gbps for an 800G transceiver.
Coherent Optics: Coherent transmission uses the amplitude, phase, and polarization of light to transmit information across an optical fiber link. Each of these attributes has 16 distinct symbol states. You often see a coherent link in a dot plot with 16 dots, and this is 16 QAM (Quadrature Amplitude Modulation) with 16 data states.
This is the graphical representation of each link in a constellation.
Coherent uses light instead of electricity, can be used over longer-reach optics, and can be amplified for over 100km+ links. Because of the 16 channels, Coherent can have more bandwidth over a single link but has a much higher complexity penalty.
Passive vs. Active: Passive relies on passive optical components to distribute the signal without active processing (DSP) or additional electrical power. This is a lower cost and doesn’t have as much error correction. This often denotes whether or not to include a DSP in the link.
Active transmission uses some powered device to regenerate, retime, or error-correct the signal. This includes either a timer or a DSP, allowing for more signal error and longer distances. Passive copper cable is likely the single cheapest method of transmission, while an active optical cable over long reaches is the most expensive method—the more content of active devices, the more power and cost for the materials.
If you considered this post helpful or want to read the paywalled posts, consider subscribing or sharing. That helps me out massively! Thank you, and have a great day.
Lot of good calls that you made over last few months boss. Great job.
Love this piece. 👍🏼👍🏼👍🏼👍🏼👍🏼❤️❤️