AI Revolution in 2024: Market Impact and Investment Opportunities

How we are viewing the AI theme

Usually thematic updates come after publishing an original thesis but in this case this update comes with no former thesis. I have written a little about AI being a large macro trend and unless you live under a rock you didn’t need me to tell you this. This is largely why I never published an original thesis as there was already plenty of pieces out there providing an intro to the AI theme and I didn’t want to just put my voice out there without being able to provide a new perspective. The point of this newsletter isn’t just to hear my own voice, it’s to provide actionable unique research and I think this primer will be able to provide that.

The AI theme has gotten derailed recently as people are now coming out with doubts surrounding the theme. It largely started with Goldman Sach’s coming out with a piece talking about how little benefit there is for much benefit we’ve seen. You can read it here if you’d like as I reccomend you should. These doubtful voices have only gotten louder as we continue to trade off the highs in various names. SOXX is 15% off the highs, NVDA is 14% off the highs while many other thematic beneficiaries are also trading well off their highs including some of the mag7 stocks.

I think this is the perfect time to revisit the theme as we need to really examine if the theme has played out or if we can continue to benefit from owning AI related stocks. Let’s start by revisiting our frameworks for themes as mentioned in our Portfolio Construction Piece

1. The first stage is the emergence of a trend. Trends often emerge from significant events or are highlighted by major beneficiaries (think the aftermath of covid for fossil fuels or NVDA for AI).

2. The main beneficiaries may trend higher for a period before a major fundamental event makes the market take notice of this trend. This is usually for very specific names in the trend and it’s usually also a good idea to short the laggards before the trend really emerges for the masses.

3. This stage is the pinnacle of thematic investing. It’s when the trend is at mass exuberance and latecomers start promoting stocks that barely relate to the trend. This is the time to really be selective of both longs and shorts, eventually selling into the peak hype. This stage can last a quarter or two or a few years, timing this is crucial.

4. This is the peak. This is when the market is flooded with companies barely related to the trend at best and will likely never turn a profit. The market’s expectations have entered the realm of fantastical and have almost no basis in reality. This is also when everyone is forced to play the trend, and the way they do are through all terrible companies. It’s the riskiest phase and the time to reduce the long positioning and realize profits on short positions if they’re profitable. Then identify new shorts based on false promises of beta to the theme that have no fundamental justification. This is dangerous to do if you can’t identify the stage of the cycle. Monitor these shorts as reflexivity (Alchemy of Finance topic) will be at it’s peak and consecutive red weeks can work as a signal. Relative strength is also a hugely important signal.

I still believe that we are in stage 3 and we just need to be more selective of the companies we long. A lot of the trend has played out but it still has juice to it. Just read any of the mag7 earnings transcripts and we can quickly figure out that the huge spending is not coming to an end.

This has been reflected in earnings so far as we continue to see growth in both the top line and bottom line in various chip makers as well as related semi-cap companies.

It’s clear there needs to be a higher degree of specificity in identifying beneficiaries of current dynamics in the overarching trend of AI and the areas most likely to outperform as we see it are interconnects, inference on device, and custom silicon.

Network Interconnects

We’ll look at interconnects through two different scopes: one for within data centers and one looking at the interconnects between data centers. Let’s start with the latter first.

It’s pretty clear that companies are building a lot of data centers with plans to build more. Not only are the number of data centers increasing but also the density of them, resulting in bottlenecks. These bottlenecks can only be solved through the interconnects between these data centers. We’ve all heard about the GPU clusters, which are getting so large that they have to be distributed across different data centers in what’s called a campus. Every cluster, data center and campus is going to represent a node. These nodes and the quantity of data they’re transporting has increased and will continue to increase. This communication between nodes occurs across spans, and these spans can be wired or wireless. That’s where the opportunity lies, in the wired and wireless communication networks.

This expansion of networks that will occur is not just about current demand, but also future demand proofing as companies will need to anticipate future needs. This will require quite the undertaking as the companies responsible for building and maintaining our communications networks have significantly underinvested for both the future and now.

This underinvestment began with covid, as telco’s such as Verizon and AT&T stocked up on inventory heavily to support the growing demand of networking devices for the shift to remote work. The shortages that were experienced during the height of the pandemic lead to overstocking and then the incremental demand for those products waned as the pandemic subsided. There is still some inventory that has to get worked through judging by the balance sheet of some of these companies but new orders are starting to inflect upwards in some pockets of the market. This makes sense intuitively as the overstocking happened in LAN scale hardware but are now having to deploy in campus and metro area networks.

This overstocking wasn’t the only reason for the underinvestment however. The large telecommunication companies have really backed themselves into a cpital allocation corner over time which has depressed revenues and profits that sell them equipment. Throughout the 2010’s bond yields continued to drop, so the large companies increased dividends to attract investors. This led to declining profitability as a result of decreased investment which caused dividend growth to outpace business performance and increased leverage. This leveraging and subsequent deleveraging have put strains on capital budgets which have caused new orders to decline to recession level territory. The current capex intensitiy is comparable to the dotcom burst levels when looking at capex to sales for the industry average. We think this ratio has bottomed and companies are prepared to start building out networks.

This is starting to show up in transcripts across the industry. CSCO announced a 14% increase in new orders in the most recent quarter. Due to lead time, it won’t be reflected for another few quarters but it is still enough for the CEO to call the bottom.

“We had a strong close to fiscal ‘24, delivering $13.6 billion in revenue for the fourth quarter, coming in above the high end of our guidance range and $53.8 billion for the year, coupled with growth in annualized recurring revenue, remaining performance obligations, and subscription revenue…We saw steady demand as we closed the year with total product order growth of 14% and growth of 6% excluding Splunk, indicating that the period of inventory digestion by our customers is now largely behind us as we expected.”

We also have similar sentiment from Corning:

“In our carrier business, customers are reaching the end of their inventory drawdowns and beginning to order closer to their current deployment levels. Additionally, government efforts to bring high-speed internet to rural communities through the Broadband Equity Access and Deployment Program will contribute to growth beginning in 2025 and add significant sales over the next several years… Carrier sales were actually down 10% as customers continued to draw down their inventory. Encouragingly though, carrier network sales grew sequentially as customers began to order closer to their current deployment levels.”

Obviously the exact timing is impossible to figure out but we feel pretty good that the turn up is coming soon. The market for Telecom services will be larger in the years to come than it is today. This is in part due to the cyclical nature of the industry but will also be the result of new technology that demands advanced connectivity.

Our ability to generate data has completely passed our ability to consume it. Compute power and time are going to waste and the key to eliminating this waste to improve our communications network. Let’s take a step back and look at exactly what we need and who will benefit.

A communication network is infrastructure that moves information. The information is encoded and needs to be transported as quickly as possible. These networks can be categorized into:

• LAN - Local Area Networks, think WIFI

• WAN - Wide Area Networks, connects LANs over long distances, think the internet itself

• MAN - Metro Area Networks, the bridge between LANs and WANs, connect multiple LANs within a metro area

It will take both wired and wireless networks in tandem to create networks capable of supporting future demands, especially at MAN levels. This is due to increasing data demands, low latency requirements, distributed computing and balancing the benfits of LAN with the reach of WANs. Speed and capacity are the name of the game.

The wired network story will revolve around fiber optic cables. Electrical signals are too slow, we need the speed of light. Using fiber optic cables we will have lower latency, higher bandwidth, lower power consumption and immunity to electromagnetic interference which are all great attributes to power next gen AI. Corning is the leader in Fiber Optic Cabling and is establishing a significant presence in the AI/ML data supply chain as well. They just announced development of an interdata center network with LUMN which will consume 10% of Cornings global fiber capacity for the next two years. PRY, NEX are other suppliers of cabling while TEL and APH are producers of connectors for wired networks.

Obviously these cables are worthless unless connected to nodes that are generating optical signals. Transceivers serve to translate those signals into light that can travel across the fiber cable, receive it and transform it back. This is often where the bottleneck for transmitting signals occurs but COHR and LITE are leading th introduction of 800G and 1.6T transceivers that alleviate these issues. COHR optical tech revolutionized long-haul fiber optic communications by increasing data transmission rates and distances. CIEN is another critical supplier as their tech enables 800G transmission speeeds over single wavelengths.

The wireless network story is harder to get a grasp on as it involves 5G networking, which has uglier ways of playing it. Some names in the space are NOK, ERIC, MRVL and RBBN. MRVL will be added to our basket but for other reasons which will be seen later.

Data Center Interconnects

Within the data centers bottlenecks are still very much a big problem. Before we get into the nitty gritty, I think it’s important to get the terminology down pat before continuing. Using the help of chatGPT we get:

• Bandwidth: The wider the hallway, the more data can pass through at once. It’s the maximum rate of throughput for a given path.

• Latency: Latency is the delay between sending and receiving data.

• Topology: Topology refers to the arrangement of nodes and connections.

• Protocols: Interconnects use protocols to standardize how data is packaged, sent, and received.

• Serialization and Deserialization (SerDes):  SerDes converts parallel data into serial data for transmission and then back again.

• Digital Signal Processors (DSPs) - fine tunes the data

• Switches guide data to its proper location.

• Pulse Amplitude Modulation (PAM4) technology allows the stacking of serving trays higher, delivering more data in a single trip.

• Network Interface Cards (NICs): NICs provide network connectivity for servers in an AI data center, enabling the transfer of data between servers, storage systems, and other networked devices. Their performance is critical for AI workloads that require high-speed data movement.

• Compute Express Link (CXL): CXL is a high-speed interconnect standard that enables CPUs, GPUs, and other accelerators to share memory and resources more efficiently in AI data centers. It reduces latency and improves performance for AI models that require large memory pools.

• Smart Network Interface Cards (SmartNICs): SmartNICs offload networking tasks (e.g., security, storage management, and load balancing) from the server CPU, allowing the server to focus on AI processing tasks. They improve data center efficiency by enhancing networking performance and lowering CPU usage.

• Chiplets: Chiplets are small integrated circuits that can be combined into a larger processor. In AI data centers, they enable the design of flexible and scalable processors that support high-performance computing (HPC) and AI workloads by optimizing compute and memory resources.

• Coherent Interconnects: Coherent interconnects enable seamless communication between different computing elements (e.g., CPUs, GPUs, and accelerators) in AI data centers. They allow data to be shared between processors without redundancy, enhancing performance in complex AI tasks.

• Remote Direct Memory Access (RDMA): RDMA enables high-speed data transfer between servers in an AI data center without involving the CPU. It reduces latency and increases throughput for AI applications that require large amounts of data to be transferred between nodes, improving overall performance for distributed AI workloads. These don’t really exist/work yet.

The choice of interconnect technology is very crucial as it impacts not just the speed of data transfers but also the energy efficiency and overall system performance. Interconnects are the technologies and protocols that enable different components in a computing system to communicate and transfer data. They can help alleviate the Von Neumann bottleneck and other performance limitations.

In a rack-scale NVL52 system we can estimate there are more than 5,200 individual connections, which gives us the sheer number of connections that are needed at the board, server and rack levels. I’m gonna go over some ideas but will be up to the reader to research each idea to make a decision. There are so many names in this space, I’m just going to go over a few and how they benefit.

Credo Technology is one of these companies. CRDO 0.00%↑ is a fabless designer of data center-focused interconnect devices such as line card PHY and SerDes chips that are sold into oem’s, odm’s and as a licensable IP. They also make optical data signal processor chips (DSPs) and active electrical cables are one of their main products. 75% of their sales come from AI/ML customers.

Hyperscale GPU clusters utilize a parallel computing model called SIMD (Single Instruction-Multiple Data). In this model, the same instruction is executed at the same time on multiple data-points. Inference on the other hand typically involves processing smaller amounts of data. Inference has an emphasis on speed and meticulous precision rather than scale. Each request is unique and requires accurate and timely results. This shift towards specialized inference hardware is likely to be accelerated by AI and the data it generates.

As AI continues to advance and generate more data, the physical footprint of training and inference systems may shrink, the digital footprint will expand and general compuiting workloads is likely to grow alongside with it. We need to explore the latest advancements in interconnects that are driving performance and efficiency in both domains. This is where optical interconnects and co-packaged optics. These are where MRVL 0.00%↑ and AVGO 0.00%↑ shine.

We highlighted MRVL 0.00%↑ in our Ten Trades for 2024 note and it continues to be one of my favourite picks for the eventual custom silicon boom and progression of AI because of their expertise and positioning to combine multiple elements of the semi-custom silicon value chain. They are currently working through 3nm designs and making investments at the 2nm scale after becoming a leader in 5nm process over the last several years. This is important as their chips are not discrete components but entire system on chip technology. While the adapation has been slow with all eyes on GPUs, the growth of application specific integrated circuits seems likely to drive the next wave of advancement in AI. We think this time is now and is echoed by AVGO CEO.

Inference on Device

It seems clear to us that eventually AI will transition to the consumer side rather just be about business or data centers. For this to happen we will need inference at the edge. It’s likely that this will look more like a personal assistant than anything else. Sort of like a more advanced Alexa but one that you can keep in your pocket or your wrist. However, Alex makes people uncofortable because that information ends up on a cloud which is more vulnerable and that information can then be used by other people and companies. Instead, we need to think more along the lines of how companies will get people to adopt this technology. People will have to be able to trust their device which means that AI capabilities and inference will need to be done on the device.

The most obvious and biggest beneficiary to this is apple due to their superiority in protecting their customer’s privacy. This inference on device will spurr on the next big upgrade cycle. Why will it be so significant? Let’s use chatGPT to go over what Apple announced:

• Apple Intelligence: Apple unveiled its new personal intelligence system called Apple Intelligence, which integrates powerful generative AI models into iPhone, iPad, and Mac devices. This system combines on-device processing with cloud-based capabilities to deliver personalized and context-aware intelligence.

• On-device AI processing: A cornerstone of Apple Intelligence is on-device processing, which allows many AI models to run entirely on the device, ensuring privacy and security.

• Private Cloud Compute: For more complex AI tasks requiring additional processing power, Apple introduced Private Cloud Compute. This technology extends the privacy and security of Apple devices into the cloud, allowing for more advanced AI capabilities while maintaining user privacy.

  • AI-powered features: Apple announced several AI-enhanced features across its operating systems, including:

  • Image generation and text summarization in native applications

  • Enhanced Siri with better app control and understanding of user input

  • AI-powered photo searches, object removal, and transcriptions in the Photos app

  • Email summarization and response generation

  • Custom emoji creation (Genmoji) and AI image generation (Image Playground)

• ChatGPT integration: Apple is integrating ChatGPT access into iOS 18, iPadOS 18, and macOS Sequoia, allowing users to leverage its capabilities within the Apple ecosystem.

• Privacy focus: Apple emphasized its commitment to privacy in AI, with features like IP address obscuring and no request storage when accessing ChatGPT.

This looks to me that they are plannng a device that will be purpose built for edge AI. This will reduce some of the things that have made replacement cycles less impressive recently.

AI models of the future won’t be just about inference but will require training on the fly, and fine-tuning their local version of more complex yet smaller sub-models which is essentially creating a model of model. These efficient but narrow-minded layers will alleviate concerns over memory and storage requirements but brings forward the problem of latency. We need to think about latency exponetially and take it on by adapting to collocation more. This is already being done to a degree but is limited by scarce resources. The answer to this isn’t just bigger models but we need specialization and parallelization to work together to turn LLMs into practical uses. For this to happen, logic silicon will need to be paced physically alongside low-latency memory, lightning fast storage and a hint of acceleration.

This is perfect for on device inference and like we said earlier, Apple is the one set to win. Apple has the vertical integration and custom silicon designs that put them in a unique position to use this model of models approach. If they aren’t and someone else does, then at least the eco system will still benefit which is why we’ll include some other names in the basket. This will require significant hardware upgrades including more powerful SoC’s, increased DRAM, enhanced microphones, advanced cooling systems and improved cameras. This shift is likely to result in larger bill of material costs for manufacturers.

Who will benefit?

• Memory producers like MU 0.00%↑ , SK hynix as AI smart phones will require more DRAM

• AMKR 0.00%↑ is considered to have some of the smallest packaging technologies in the world. Every nanometer counts when trying to fit a series of components onto a chip and into a device such as an iPhone, so we see significant investment in this space as likely.

• AVGO 0.00%↑ will benefit through it’s ASICs and SoC’s that are critical for inference tasks. They also provide advanced wifi, ethernet, bluetooth solutions along with being heavily invested in 5G technologies.

• WDC 0.00%↑ provides high performance storage

• QRVO 0.00%↑ provides radio frequency solutions and components such as transceivers, 5G technology etc.

So let’s summarize our findings by creating a cohesive basket that we will implement. We’re going to implement an equal weighting across the basket except for MRVL 0.00%↑ where we will use a 25% overweighting due to our bullishness on custom silicon.

This basket is getting put on with a 20% weight of our theoretical $100K portfolio. Right now the rest is held in cash until we have more primers out. Any updates in the future will be published in the chat. I’m not sure how I will handle future random option/futures trades as koyfin doesn’t allow for this. I’m thinking that we will alter the cash balance while tracking in an excel sheet. If you enjoyed this primer, then I am happy to say we have more lined up on healthcare, the uranium trade and others along with regular weekly notes. I’ll have an update out on the direction of the substack soon! Subscribe so you don’t miss a thing and I’d appreciate you sharing this with other people!