AMD: The market miscalculates its AI potential and will regret it

thesis

Expectations for the growth of artificial intelligence (AI) have created some big winners for 2023 so far. Microsoft Corporation (MSFT) and NVIDIA Corporation (NVDA) are rightfully considered leaders in AI development technologies in the software and hardware space. Both should be well positioned to benefit from the growth of this new market. And their share prices and market capitalization have risen sharply year-to-date, as seen in the chart below. Notably, NVDA has become one of 5~6 companies depending on how you count – with a trillion dollars market capitalization at the time of this writing thanks to AI expectations.

Compared, Advanced Micro Devices, Inc. (Nasdaq:AMD) seem to have lagged behind the wave of artificial intelligence so far. Its market cap currently hovers around $179 billion, just a fraction of the billion-dollar valuation of NVDA and MSFT. It’s current market capitalization translates into an increase of only 41% year-to-date, in stark contrast to MSFT’s 72% and NVDA’s 186%.

Against this backdrop, my argument here is to argue that AMD’s current market cap represents an underestimation of its AI potential. As for the details later, AMD is focusing on AI products and expanding its data center product portfolio. And I see its products as a strong competitor for NVDA’s flagship AI chips in the near future.

AI chips from AMD

Currently, Nvidias A100 and H100 are the best sellers in the field of AI. Both are high-end GPUs specifically designed for AI and machine learning applications. They are the most powerful GPUs on the market and are used by many of the world’s leading companies. Their power is further augmented by the accompanying NVDA software suite. As such, there’s certainly good reason to be optimistic about NVDA’s whereabouts in the future of AI.

However, I see AMD chips (plus some from Intel as well) to be strong competitors in the near future. The table below summarizes some of the main competitive offerings and how I see things. These key competitor products include the AMD Instinct MI250X and the Intel Ponte Vecchio series. These chips are also designed for AI and machine learning applications and are all good replacements for the A100/H100 at a competitive price.

Notably, a recent MosaicML study found that MI250 can achieve 80% of the performance of A100 in large language model (LLM) benchmark tests. This is mainly due to software updates from AMD released in late 2022, as well as development support provided by the new Meta Platforms version of the open source software PyTorch, announced in March 2023. This allows chips from AMD to enjoy better support software and be more easily used in training popular AI models. And Hanlin Tang, CTO of MosaicML, pointed out that software is just the biggest roadblock for most AI professionals. Tang also expects AMD to deliver even better performance on new HPC GPUs as software tools continue to improve. With these improvements, I predict the MI250’s performance will soon be on par with the A100.

Evaluation and profitability

However, AMD is trading at a fraction of NVDA’s valuation multiples. Before I go any further, I have to clarify that AMD’s valuation is by no means low in absolute terms. With a TTM P/E of 37.8x (see table below), it’s quite expensive even by AI standards. For example, it is higher than MSFT’s 36.1x TTM P/E ratio.

However, when fully contextualized, it is one of the cheapest AI titles in relative terms. First, compared to NVDA, its current P/E ratio is highly discounted. As seen in the table above, AMD’s TTM P/E of 37.8x is nearly only one-quarter that of NVDA’s of 138x, and its FY1 P/E of 39.5x is about 27% discount to that of NVDA of 54.4x. When considering growth potentials, its FY2 and FY3 P/Es are even more discounted than NADA.

Second, adjusted for its profitability, AMD’s valuation is quite in line with the overall market, as shown in the following two graphs. The source data used in these graphs is obtained from the Dartmouth Tuck Business School database. And the charts were made following the Fama-French framework as detailed in my previous article. A brief summary of this framework is provided here for ease of reference:

The method is named after Eugene Fama and Ken French. Fama shared the Nobel Prize in Economics with Lars Peter Hansen and Robert Shiller. The Fama-French method is considered a significant improvement over the CAPM method. Past results following this method have shown that the following two factors have indeed held up across all time periods.

• The profitability factor represented by OP (operating profitability). The OP factor in period t is defined as operating income divided by the sum of book equity and minority interest for the last fiscal year ending in time period t-1.
• The valuation factor represented by the usual P/E ratio.

As seen in the first chart below, AMD’s average OP over the past few quarters is about 48% as defined in the Fama French method, above the 25% percentile of the overall stock market. Even measured by ROCE (return on invested capital), its average is around 37%, pretty close to the 25% percentile of the overall stock market. So, as seen in the second chart below, its FY1 P/E of 39.5x is also near the top 25% percentile of the overall market, in line with its profitability ranking. While its FY2 P/E of 26.8x is well below the 25% percentile, signaling potential undervaluation.

Risks and final considerations

AMD also faces some challenges. It has reported mixed financial results in recent quarters due to a number of factors. Its data center revenues were largely flat year over year as higher cloud sales were offset by lower enterprise sales. Customer segment revenues declined significantly year-on-year. These results are both caused by a variety of headwinds. And, the way I see things, the major headwinds include short-term macroeconomic uncertainty and also high inventory.

These headwinds have caused a combination of a decrease in the average selling price and also in shipment volume. For example, its customer segment reported a 26% decrease in average selling price and a 55% decrease in unit shipments in the recent quarter. Though I view both of these headwinds as only short-term and temporary factors.

Longer term, competition will remain intense in the AI ​​space. As mentioned above, both NVDA and INTC also offer AI chips. And in addition to these chip companies, software companies like Google and Meta Platforms are also developing their own in-house AI chips.

To address the long-term risk, I believe AMD chips offer several unique technological advantages to remain competitive both in the AI ​​space and in general. For general applications, its Zen architecture is a highly efficient and scalable architecture that allows AMD chips to perform well in a wide range of applications. Zen architecture is based on a number of innovative design features, including out-of-order execution, branch prediction, and speculative execution. These features allow the Zen architecture to execute instructions more efficiently and improve performance. AMD’s MI250X AI chip is based on the CDNA2 architecture, which is a newer architecture designed specifically for high-performance computing (HPC) and AI-oriented workloads. Additionally, the MI250X chip is manufactured using TSMC’s 6nm process technology, which is among the most advanced process technologies available today.

In closing, my argument is that AMD’s current market capitalization represents a substantial undervaluation. And my thesis rests on two arguments. First, my view is that the market underestimates the potential of its AI-oriented products, especially the MI250 chip and its software support, as detailed in the second section. And second, I think Advanced Micro Devices, Inc.’s valuation is still reasonable and in line with its profitability, as detailed in the third section.