The innovative Advanced Micro Devices, Inc. (NASDAQ:AMD) progress of their 2.5/3D AI chips indicates a crucial step towards a future of dominating the AI computing chip sector. I am convinced that the market has yet to fully recognize the transformative role these advanced chips will play in reshaping data centers and accelerating the AI revolution. As we nearAMD’s Q4 earnings release post-market on January 30th, my main focus is on the potential of 2.5/3D chips. With these chips gaining traction, I anticipate that the company’s revenue growth will accelerate. In my opinion, AMD stock is a strong purchase leading up to earnings.
Background: What Even Are 2.5/3D chips?
2.5/3D chips represent a significant evolution in semiconductor design, allowing for a new level of complexity and integration beyond the capabilities of traditional flat chips. These chips, which integrate multiple chiplets or dies into a single package, significantly improve performance, reduce power consumption, and optimize space. AMD’s new MI300X series, which includes products like the MI300X, consists of 3D chips and is known for their advanced memory, such as the integration of HBM3, providing up to 192 GB of total memory and 5.6 TB/sec bandwidth, a significant improvement from previous generations. The chip series was revealed on December 6th, 2023.
How 2.5/3D chips Affect AI Datacenters
In AI data centers, the impact of 2.5D and 3D chip technologies, as (for example) presented by AMD’s MI300X series is transformative. These technologies enable a denser integration of computing power in a smaller footprint, crucial for the increasingly complex and demanding AI and model training workloads. The combination of advanced memory and high bandwidth in chips like the AMD MI300X and MI300A produce more efficient and effective processing of data-intensive AI applications.
These 2.5D and 3D stacking technologies are key elements in successfully meeting the performance requirements and the high demand for AI and data center applications. The overall market for these stacking technologies, driven by High-Performance Computing (HPC) applications, is expected to demonstrate rapid growth, with the HPC market anticipated to reach $107.48 billion by 2033 (more than double from 2023). This market demands 3D chip technology.
Why is AMD Uniquely positioned to benefit?
On October 27th, 2020, AMD and Xilinx agreed to merge, resulting in AMD gaining a vast array of high-performance and adaptive computing products such as CPUs, GPUs, FPGAs, and Adaptive SoCs. This move by AMD vastly expanded their capabilities in offering comprehensive and adaptive computing solutions. Xilinx’s expertise in adaptable computing and networking enhances AMD’s product portfolio, therefore helping secure a position for AMD in the rapidly evolving high-performance computing world. AMD joined forces with Xilinx due to their expertise in 2.5/3D chips. AI chips are starting to demonstrate how this transaction is starting to pay off.
AMD’s focus on 2.5/3D chip technology, as shown in the launch of their Instinct MI300X series, uniquely positions the company to compete in the AI and HPC markets. These chips offer an integrated package of CPU and GPU functionalities, significantly enhancing performance while optimizing power consumption. The MI300X series, especially the MI300X, can boast over 150 billion transistors and deliver memory capacity and bandwidth that surpasses the NVIDIA Corporation (NVDA) H100, the current market leader. With 2.4 times the memory and 1.6 times the memory bandwidth of the H100, AMD’s MI300X is designed to offer superior AI inferencing performance. AMD is literally stacking chips vertically to surpass Nvidia (it makes 13 pieces of silicone behave as one chip). Impressive.
Claims from Nvidia arose that while testing the Nvidia hopper, AMD did not use the most recent TensorRT-LLM kernel optimizations, therefore their data was invalid. AMD responded stating that at the time of the testing, Nvidia’s most recent data was not accessible, but after further advancements made within the MI300X, AMD still outperforms Nvidia’s optimized TensorRT-LLM.
What Are Earnings Expectations?
As AMD heads into earnings on January 30th, the projections are guiding for growth as the sales of the MI300X series begin to accelerate.
For the fourth quarter of 2023, AMD management has guided revenue to hit approximately $6.1 billion, representing a year-over-year growth of about 9% and sequential growth of approximately 5%, with a non-GAAP gross margin of about 51.5%. For Q4 EPS is forecasted to increase from $0.69 to $0.77, representing 11% year-over-year growth. According to AMD management, the team is estimating about $2 billion in sales for 2024 for the MI300X chip series. I think this is conservative guidance (more on this later).
Wall Street is specifically estimating sales of $6.14 billion for Q4, which is slightly above guidance.
What I am Looking For On the Earnings Call
As the earnings call approaches, my attention is especially focused on updates on the sales of AMD’s MI300X chip series. AMD claims that their revenue will stand at $2 billion, but I believe this is a severe underestimate. Meta Platforms, Inc. (META) stands as a big reason why.
One of Meta’s long-term goals is to increase their research in artificial general intelligence (AGI). The programming of their new Llama 3 model requires the use of a ton of chips. A recent announcement by Meta states that their AI training center will contain 600,000 GPUs, 350,000 of which will be H100 chips from Nvidia (based on what CEO Mark Zuckerberg said).
This leaves 250,000 “H100 equivalents” that Zuckerberg said would be part of his massive GPU cluster. Meta already stated they will be a customer of this new chipset.
Since AMD is a leading figure in this market (after Nvidia), we can assume that AMD could reasonably capture 50% of the leftover 250,000 chips needed for Meta, (125,000 units). The discrepancy here is that AMD is claiming they expect to only sell about 100,000 of their MI300X chips (2 billion in revenue in 2024 divided by an average selling price (ASP) of $20,000 is 100,000 chips implied).
As demonstrated, Llama 2’s increased training performance with the MI300X compared to Nvidia’s H100 chip gives me confidence that AMD could capture a significant portion of purchases for Meta.
Taking into account my expectations for the Meta chip purchase, AMD should easily surpass the expected $2 billion in revenue. If each M1300X chip costs roughly $20,000 and Facebook purchases 125,000, that’s already $2.5 billion in revenue before including other purchasing parties. I am looking for any updates on their Meta relationship on the earnings call.
Valuation (If My Assumptions Are True)
As mentioned above, I believe AMD is underestimating their MI300X revenue potential at $2 billion in 2024. Considering the demand for M1300X chips for Meta, these numbers alone beat AMD’s projected annual revenue. When analyzing other purchasing parties, AMD could ship out 300,000-400,000 M1300X chips in 2024 (according to analysts). If they only ship out 350,000 units on top of Meta, at an ASP of $20,000, that’s $7 billion in marginal revenue. Add Meta’s additional $2.5 billion in revenue, this sums up to a total of $7.5 billion potentially under-appreciated revenue ($9.5 billion total minus the $2 billion guided).
AMD is currently valued at 12.42 forward sales while the sector median is 2.90. For my upside potential calculation, I took the mean of these two values to 7.66. Using this averaged price-to-sales multiple along with the difference between my calculated revenue ($9.5 billion) and the expected revenue by AMD ($2 billion), this could unlock $57.45 billion in market cap appreciation potential. With this, I believe there is a share price upside potential of about 20% to AMD’s stock price (assuming this P/S multiple of 7.66 holds and there are no more shares issued).
Risk To Thesis
While I believe that AMD’s MI300X series will soon become a key player in the AI chip industry, the path is not perfect. The MI300X, while marked as “the most advanced AI accelerator in the industry” by AMD’s CEO Lisa Su, is still an expensive piece of hardware at $20,000/unit. Even some large AI companies are struggling to turn a profit on their AI investments, and if the industry goes through a downturn, these chip sales could be one of the first places impacted.
With a pivotal earnings announcement around the corner, I believe the release of the groundbreaking MI300X series chip has positioned AMD at the forefront of the high-performance computing and AI revolution. These chips promise vast improvements in areas such as performance, energy efficiency, and total cost of ownership. As the industry increasingly recognizes the transformative potential of these chips, AMD’s position is further solidified. I believe the upcoming earnings call is expected to reflect the positive impact of the MI300X series, helping solidify my bullishness. Heading to earnings, I think AMD stock is a buy.