Advanced Micro Devices (NASDAQ:AMD) reported Q1 2024 earnings on Tuesday, April 30. On the earnings call, AMD also provided an update about its MI300 AI accelerators, reporting an increase in the 2024 backlog from $3.5 billion to $4 billion. AMD remains poised for strong growth in data center GPUs this year, and its current guidance implies that MI300 revenues should reach about $1.6 billion/quarter (6.4 billion/year) in Q4 2024.
However, I am concerned that the slow backlog growth this quarter could potentially be a warning sign about problems in terms of competitiveness with Nvidia (NVDA). In March, Nvidia announced its next-generation Blackwell GPUs, along with expanded AI software offerings. It is possible that these announcements could have significantly shifted the competitive equation against AMD’s MI300 GPUs. If so, the slow growth in AMD’s backlog this quarter could be a sign of things to come for the rest of this generation of chips. With the latest information from AMD, this now seems like a significant risk to me, and a wait and see approach makes more sense. I am, therefore, downgrading AMD from Strong Buy to Hold.
In this article, I first go over AMD’s MI300 outlook for 2024, and then discuss my reasons for caution.
Q1 MI300 Revenue Was In Line With AMD’s Implied Guidance
During the last earnings call, AMD had reported that its 2024 MI300 backlog had grown from more than $2 billion in December to more than $3.5 billion in January. This figure has now been updated to more than $4 billion.
The important thing to note here is that these MI300 figures from AMD are not a forecast of estimated total sales in 2024. They are the backlog of firm commitments. CEO Lisa Su had been pretty clear about this during the Q4 earnings call, stating:
“the way to think about the $3.5 billion is these are customers that we’re working with who have given us firm commitments on what they need.”
After the Q4 earnings report, I had estimated that AMD’s backlog, combined with management’s commentary, implied roughly the following (committed) quarterly sales for MI300 sales in 2024 (amounting to $3.9 billion total in 2024):
- Q1: $600 million
- Q2: $800 million
- Q3: $1,100 million
- Q4: $1,400 million.
AMD reported that its total sales for the MI300 (since launch) crossed $1 billion in Q1. MI300 sales were $400 million in Q4, implying that they came in around $600 million this quarter — right in line with the company’s previous implied guidance.
Q2 MI300 Revenue Likely About $900m
AMD did not provide an exact guidance for data center GPU revenue for Q2, but we can read between the lines to arrive at an implied figure.
To start, at the midpoint, AMD is guiding a $230 million increase in total revenue. Moreover, management explained that “the data center GPU ramp… will be more than the overall company’s $200-some million ramp.”
How much more? Hard to say exactly. But I think around $300 million seems reasonable. AMD is guiding that in Q2 “gaming is [expected to be] down similar zip code like Q1” (i.e., 33% or ~$300 million). Embedded is expected to be flat. Client and data center CPU are both supposed to be up. If MI300 sales come in at $300 million, then that would leave $230 million sequential growth in client and data center CPU (i.e., ~7%).
This seems like it should be roughly in the ballpark of what to expect, although perhaps MI300 could grow a bit faster. But for now, I think $900 million in Q2 seems like a realistic expectation.
It’s worth noting here that the margin of error in AMD’s revenue guidance is plus or minus $300 million for Q2. This means that there could be some potential upside with MI300 sales, although it’s not clear how much more MI 300 can contribute because management noted that they are supply constrained in Q2.
My Updated MI300 Quarterly Projections For 2024
Using the above figures, we can estimate the following rough quarterly data center GPU projections for AMD (for a total of $4.3 billion in 2024):
- Q1: $600 million
- Q2: $900 million
- Q3: $1,200 million
- Q4: $1,600 million.
We’ll see what the shape of the ramp looks like, but this seems like a reasonable rough estimate. The figures suggest that MI300 revenue should reach a run rate of about $6.4 billion in Q4 (up from last quarter’s projection of $5.6 billion). AMD continues to make excellent progress as far as 2024 is concerned.
However, there may potentially be a problem here in terms of what AMD can accomplish in 2025.
The Potential Problem
The point of concern is that AMD’s backlog might not be growing fast enough to sustain its growth trajectory in 2025. For AMD to keep growing strongly in 2025, it would have to hit revenue of about $2.5 billion/quarter at some point. However, the backlog for MI300 only grew by (presumably) $500 million or so in Q1. This makes me wonder if, despite the strong projections for 2024, the MI300 (and its upcoming refresh) will be able to sustain AMD’s growth in 2025.
Now, I want to be clear that we’re just looking at a slowdown in backlog growth for a single quarter. With one data point, there isn’t a lot to go on, and this quarter’s slow backlog growth could potentially have a (relatively) innocuous explanation. Possibilities include:
- Lumpiness in orders and chance could have resulted in slow backlog growth this quarter. It is still May, so there is time for more orders to come in for H2 2024/H1 2025, and perhaps AMD will see a quick change in backlog. Management did report more than 100 customer engagements (up from “dozens” reported in Q4), so it is possible that significant new orders could flow in soon. But, of course, engagements aren’t orders, and it is unclear how many of these engagements may result in sufficiently large orders to meaningfully move the needle.
- Timing considerations among customers could have led to few purchase decisions being finalized among customers this quarter. Perhaps more customers are looking at 2025 for deploying MI300 systems. Management hasn’t mentioned anything about 2025 orders yet, so we will have to wait and see.
However, there is also a less innocuous possibility: AMD may be having trouble competing with Nvidia. When the MI300X came out, it looked pretty competitive (hardware-wise) — MI300X chips benchmarked similarly to the H100 in training and better in inference. AMD’s backlog grew quickly to $2 billion in December 2023 and $3.5 billion at the end of January 2024.
However, the competitive landscape has evolved. Nvidia held its GTC keynote in March, where it revealed its next-generation Blackwell GPUs, slated to launch later this year. Nvidia is pricing Blackwell GPUs pretty competitively at $30,000-$40,000 (considering the higher silicon content and memory). Moreover, Blackwell GPUs could see a significant boost in performance and performance/watt.
Nvidia also showcased updated networking solutions, along with an impressive slew of new and updated software including pre-trained models that customers can fine tune, and offerings for various use cases like robotics, simulation, and forecasting.
Given all these announcements from Nvidia, it would not be surprising if some customers may opt for Blackwell over MI300. It’s hard to say, but in theory it seems quite possible that Nvidia’s impressive new offerings could shift the balance against AMD. The timing of AMD’s slow backlog growth seems a bit ominous in this context. And so, although it is currently difficult to say for sure either way, I do wonder if customers may have become more inclined toward Nvidia now than they were before all the new announcements.
It, therefore, seems possible to me that AMD could be facing a significant headwind due to competition from Nvidia. If this results in AMD not being able to grow its order book fast enough, then it is possible that it may not be able to sustain its data center GPU growth trajectory in 2025 (or, at least, not to the extent that investors currently expect). It is difficult to say confidently with the information we have, but I think there is a real risk here.
After all, AMD has never had much success in data center GPU before (in essence, due to Nvidia’s better offerings). There is a competitive gap here, and Nvidia’s announcement of impressive next-generation products can’t be helpful for AMD’s competitiveness.
AMD May Be Too Risky Presently
Given my heightened concerns about AMD’s competitiveness, I am neutral on AMD and downgrading it from a strong buy to a hold. There is a possibility that AMD could disappoint in data center GPUs compared to the market’s expectations, in which case the stock price could stagnate or decline. In my opinion, there is significant risk here until we get further information.
Of course, datacenter GPU is only one (until now very small) part of AMD’s overall business. As far as datacenter CPUs are concerned, management stated that they believe that they gained market share in Q1. With Zen 5 CPUs on the horizon in H2, it is possible that AMD’s share gains could continue in both datacenter and client CPU. I therefore don’t want to suggest complete gloom and doom for AMD.
Still, it seems fair to assume that a significant part of AMD’s current valuation is based on expectations about data center GPU, so a disappointment on this front could still be bad for investment returns in the short term.
I will revisit AMD in future quarters to see if my concerns about MI300 are well-founded or if there is an innocuous explanation for the slow backlog growth after all. Moreover, despite my current concerns, I still think that Advanced Micro Devices, Inc. is a very impressive chipmaker with strong leadership and execution, so it is worth reconsidering as the year proceeds (especially if shares get cheaper). But I don’t like the high degree of uncertainty presently, and think there are other compelling AI plays that seem less risky in the short term.