Who's winning in AI infrastructure: GPUs
Portfolio: Beachman's picks in the buildout of AI data centers
A few days ago, we kicked off our hunt for the next winners in AI.
In the Beachman community, we are tracking actual numbers and trends…not predictions and pontifications. We pay more attention to who is spending on AI, who is using AI, who is making money on AI. We use our AI tracker to monitor about 25-30 stocks in terms of their AI-specific products, revenues, forward estimates and competition. We are looking past NVDA’s GPUs to find the next set of investment opportunities that could work in 2025. Now might be the time to buy those stocks.
Today, we will continue our quest by focusing on AI infrastructure. Let’s get into it…
Table of contents
Main components of AI infrastructure
Types of AI GPUs
Comparing the top GPU contenders
Beachman’s picks and buy points
Conclusion
Main components of AI infrastructure
In my AI primer post from earlier this Summer, I described AI infrastructure as the bottom most layer of the AI stack that includes hardware and software components, installed in the data center, that are used to develop, train, test, power and run the AI models (LLM) and applications.
The main components in AI infrastructure are:
GPUs (graphics processing unit) - trains and runs the LLMs. Sometimes referred to as AI accelerators. Currently a $200B market with profit margins as high as 55-60%.
CPUs (central processing unit) - supports the GPUs to run more efficiently.
Memory - helps the GPU store and quickly access frequently-used data. A $25B market with margins in the 30-40% range.
Storage - holds large amounts of business data used for AI training and inference.
Networking - manages connections between GPUs, GPUs & CPUs, server racks, GPUs & memory, GPUs & storage and even between AI data centers. A $25B industry with 50-60% margins.
Server racks - modular cabinets that house all the above components. A growing commodity like business with very low margins of 10-15%.
Power - electrical systems used to power the data center.
Most of this list is self explanatory. Remember, I promised not to make this complicated.
The simple way to understand the infrastructure stack is that the GPUs are the new semiconductor chips that enable AI to work. They process large amounts of data and complete AI calculations and predictions.
Every other piece of infrastructure is in a supporting role.
All this infrastructure, including GPUs and storage and power and networking…all of them need highly specialized software to run efficiently….software that comes with the hardware component, software that has to be configured during the install and then has to be monitored while the data center is running.
So far this year, AI infrastructure stocks have provided some of the highest investor returns.
AI data centers are currently about a $250B market and expected to grow to $500B+ by 2027 and to $1T+ by 2030. The runway is potentially quite long and we are in an early innings…
Types of AI GPUs
There are two types of GPUs available today:
Merchant AI GPUs - Generic GPUs that can be used by any customer for a wide variety of use cases. These are produced by only two companies today, Nvidia (NVDA) and Advanced Micro Devices (AMD).
Merchant GPUs are currently about 85% of the total AI GPU market. They can be mass-produced at very high margins once volume production is in full swing. Since they are programmable for different use cases, they can be applied in all sorts of industrial, business and consumer environments.
Custom AI GPUs - Custom GPUs are designed and manufactured to customer specifications for their specialized internal usage. Today, Broadcom (AVGO) and Marvell (MRVL) are the largest producers of custom GPUs. A few other companies like AMD, Intel and Qualcomm dabble in this space.
Custom AI GPUs are also known as ASICs (application specific integrated circuit) or TPUs (tensor processing unit). The TPU was co-developed by Google and Broadcom. Google uses TPUs in its AI public cloud to train and run its products like Google Search, Gmail, Google Translate, Google Photos, and YouTube. They are specially designed for Google applications in Google’s data centers. Amazon, Microsoft and Meta all have their own custom GPUs for their respective applications.
ASICs are expected to grow to about 25% of the total market by 2028. Since they are co-designed and co-developed by the producer and the customer, ASICs typically have lower profit margins.
Comparing the top GPU contenders
As of today, there are only 4 fabless producers leading in the AI GPU and ASIC space: NVDA, AMD, AVGO and MRVL. These 4 companies have, so far, cornered the customer orders, the bookings, the sales, cash flows and profits. Therefore, from an investor perspective, considering a 2-3 year timeframe, it makes sense to focus on these four stocks.
There are other emerging contenders…the AI space is moving fast. Venture funding is pouring in as everyone wants to find the next Nvidia. For this post today, we will stick to what is in front of us and investable.
Here is a comparison table for NVDA, AMD, AVGO and MRVL.
NOTE: The Beachman score is published quarterly and calculated from a company’s past performance and future prospects with a max of 31 points. A stock with enterprise value / gross profit below 15 is undervalued. (EV/GP) / forward growth rate less than 25 is undervalued. Calendar years represented above.
As we look over the metrics above, here is what stands out to me:
NVDA is clearly leading the pack. They will sell over $100B of AI stuff in 2024 and grow sales to $134B in 2025.
AMD is slated to ramp up significantly in 2025+ as their GPU roadmap gets baked in and sales of AI PCs start ramping up.
MRVL, too, has some impressive AI revenue growth coming its way in 2025.
By most measures, all 4 stocks are quite overvalued…no surprises there. However, Morningstar considers MRVL to be a good value buy at its current price.
AVGO has the worst fundamentals of the lot - lopsided balance sheet, lowest shareholder value, lower trending margins and cash flow margins.
Broadcom will see about $12B in AI revenues this calendar year, followed by $15B in AI sales next year.
The table above gives you a sense of the business performance and prospective metrics that I consider when evaluating a stock for my portfolio, especially while comparing companies in the same industry.
Beachman’s picks and buy points
Given all the information above…