In 2023, as $NVIDIA (NVDA.US)$ the top competitor to NVIDIA, the unveiling of the MI300 chip was met with high market expectations. $Advanced Micro Devices (AMD.US)$However, just two years later, the narrative in the AI chip industry remains one of 'the strong getting stronger.'
With the rise of ASICs, Advanced Micro Devices (AMD) has become an increasingly less prominent challenger to NVIDIA. The company's stock price experienced a roller-coaster ride over the past two years but has rebounded in the last three months, returning to its previous highs—what has reinvigorated AMD?
Is AMD on the rise again?
Since the launch of the MI300 series, AMD's stock price has seen significant volatility, reflecting the market's complex expectations for its chip business.
Market expectations for AMD are largely pinned on the growth in shipments of data center AI acceleration chips. More accurately, an increase in AI chip shipments is the most direct catalyst for AMD's stock price.
In December 2023, AMD first introduced the MI300 series, directly targeting NVIDIA's H100 series. The memory capacity of the MI300 is 2.4 times that of NVIDIA's H100, and its memory bandwidth is 1.6 times that of the H100.
With this more cost-effective chip, AMD is considered to have the best chance of competing with NVIDIA for a share of the AI acceleration chip market.
The high expectations for MI300 shipments were reflected in the stock price, which rose by nearly 75% from the fourth quarter of 2023 to the first quarter of 2024.
Despite the MI300 series driving significant quarterly growth in the data center business, with sales exceeding $1 billion, Advanced Micro Devices' (AMD) sales expectations for the MI300 have fallen short, undermining market confidence. The constraint on shipment volumes is not due to a lack of market demand but rather to supply chain capabilities.
NVIDIA secured up to 60% of Taiwan Semiconductor's CoWos packaging capacity with its larger chip shipments, and Samsung's memory production capacity was also prioritized for NVIDIA. These capacity constraints severely impacted the shipping schedule, especially when major customers were urgently demanding chips.
Subsequent financial performance over three consecutive quarters failed to reverse the downward trend in AMD's stock price.
In October last year, AMD introduced the MI325X, which was intended to be a competitor to NVIDIA's H200. It is expected to enter production in the fourth quarter of 2024 and begin shipping in the first quarter of 2025.
However, due to the lag in overall iteration speed, the MI325X will face competition from NVIDIA's B200 in the second quarter of this year in the accelerated chip market.
In terms of hardware performance, the NVIDIA B200 has a significant advantage in the number of transistors, memory bandwidth, and FP8 peak performance. Overall, the MI325X still lags behind the B200 by a considerable margin.
The large-scale shipment of the MI325X coincided with the second quarter of the B200's release. The demand for stockpiling more powerful chips led most suppliers to choose the HGXB200 over the MI325X, resulting in poor sales for the latter.
As a result, in its competition with NVIDIA, AMD's long-touted product cost-effectiveness has not been particularly advantageous. In fact, at the end of last year's fourth quarter, the sudden rise of the customized chip route created another AI narrative competitor for NVIDIA, while AMD has almost become irrelevant.
This year, AMD's stock price rebounded from a low of $76.48 in early April to over $170, more than doubling in value.
The rise in the company's stock price is primarily due to the launch of the MI350 series, expectations of relaxed export restrictions, and an increase in server CPU market share.
On one hand, according to institutional data, the company has managed to increase its market share in the server market from 2% to 36.5% over eight years. With the combination of 'CPU+GPU', the company's server market share is expected to continue to grow steadily.
On the other hand, in mid-July, with the U.S. government relaxing NVIDIA's export restrictions on H20 AI chips to China, Advanced Micro Devices (AMD) also confirmed that it will soon be selling MI308 AI chips in China. Additionally, the company disclosed a loss of $800 million in the second quarter, which can be seen as the end of negative news.
Advanced Micro Devices (AMD) stated that, excluding inventory write-downs, the gross margin (43%) would increase to 54%.
The market is most looking forward to the future performance of the GPU product line.
The MI350 series, released in June, uses Taiwan Semiconductor's 3nm process, integrating 185 billion transistors and equipped with 288GB HBM3E (with a bandwidth of 8TB/s). The FP4/FP6 computing power reaches 20 PFLOPS, which is four times that of the MI300X.
At the AMD press conference, the CEO of OpenAI personally committed to using AMD's latest GPUs.
If one can draw a parallel to 'Stargate,' the multi-billion dollar computing power order between OpenAI and Oracle, it is evident that, given the current demand for computing power far exceeding expectations, even NVIDIA may not be able to meet the procurement needs of major cloud service providers and large AI model players. Moreover, downstream players are still exploring more cost-effective chip solutions.
The MI400 series, set to launch next year, will feature a Helios rack system integrating 72 chips, with the goal of directly targeting NVIDIA's DGX SuperPod ecosystem.
As U.S. companies have recently reported their earnings, Morgan Stanley estimates that major U.S. technology firms will spend nearly $400 billion on AI infrastructure this year, primarily to build their artificial intelligence capabilities. This amount exceeds the EU's defense spending last year.
The demand for infrastructure by technology companies is enormous, forcing them to invest substantial sums in constructing new data centers.
If the focus is solely on competing with NVIDIA, there may be no hope of reversing the situation in the short term.
However, given the size of the market, AMD also wants to capitalize on the opportunities.
Nevertheless, after a 70% increase in its stock price in a single quarter, is AMD's rapid growth already within reach?
AMD is not NVIDIA.
In fact, AMD's progress in AI can only be considered relatively mediocre.
During Advanced Micro Devices' (AMD) earnings call on Tuesday, analysts sought to understand whether the MI350 could quickly capture the robust demand for AI infrastructure. However, the company's vague response led to a decline in AMD's stock price in after-hours trading.
Analysts repeatedly inquired about the detailed revenue outlook for AMD's AIGPU business in the second half of the year. However, CEO Lisa Su did not provide a clear forecast as she had done last year.
She merely stated that the adoption of the MI355 is faster than expected and has garnered significant interest from customers, with initial shipments already made in June. As for how much this series can contribute to sales in the third quarter, the biggest constraint may be production capacity.
In terms of customer visibility, AMD is working closely with its clients. The advantage of the MI350 is that it can be directly deployed into existing data centers, and we are actively collaborating with these clients. For the MI400 series, which involves full rack-scale design and implementation, we are currently working with major clients to ensure that when the Helios racks are launched in 2026, they will be fully compatible with their data center builds.
Despite growth in the data center business, AMD's Instinct AIGPU remains overshadowed by NVIDIA in the market, struggling to challenge NVIDIA's dominance in terms of performance, ecosystem development, and market share.
Although AMD's financial performance was impressive, its financial metrics still fell short of those of its main competitor.
AMD's data center business grew by 14% year-over-year, while NVIDIA's grew by 154%;
AMD's overall revenue increased by 32% year-over-year, while NVIDIA's grew by 122%;
The gross margin declined to 40% (54% after accounting for specific inventory issues in the quarter), while NVIDIA's gross margin stands at approximately 75%.
AMD has accelerated the release of its AI chips from a two-year cycle to an annual one, but it still lags one generation behind NVIDIA's most advanced chip products. According to AMD's plan, the MI350 series will compete with the B200 in a new round of competition.
The AMD Instinct MI355X, based on the CDNA4 architecture, integrates 1024 matrix cores across 256 compute units, offering FP4/FP6 computational power of up to 20 PFLOPS.
In terms of inference, the MI355X outperforms the B200 by about 20-30% in vLLM/SGLang benchmark tests for DeepSeek R1 and Llama 370B. For training, the MI355X can achieve 2.5-3.5 times the performance of the MI300X when pre-training Llama 370B (FP8), and it is on par with the B200/GB200.
AMD's advantage lies in its cost-effectiveness. Reports suggest that even if the MI350 is priced 70% higher at $25,000, it remains cheaper than NVIDIA's B200, attracting cost-sensitive clients such as Oracle and IBM.
However, for customers leasing GPUs on a short-term basis (within six months), NVIDIA GPUs consistently offer better value for money.
This is because there are over a hundred Neocloud providers in the NVIDIA ecosystem offering short-term rental services. The ample supply creates a competitive market, driving down costs.
In contrast, only a few suppliers offer short-term rental services for AMD GPUs. This scarcity leads to a tight market and artificially inflated prices, weakening overall cost competitiveness. This explains why, apart from large cloud service providers, the adoption rate of AMD chips is not high.
On the other hand, the gap in software ecosystems remains a critical weakness. According to SemiAnalysis, NVIDIA's CUDA ecosystem covers over 90% of AI developers, while AMD's ROCm7, although supporting PyTorch, TensorFlow, and vLLM frameworks, still lacks in-depth optimization and community support.
In practical use, the company's software and testing ecosystem are significantly behind NVIDIA's CUDA. Advanced Micro Devices' software is plagued with bugs, and engineers need to set numerous environment variables and adjust parameters to make it work, making it much less user-friendly compared to NVIDIA.
As one of the most advanced chips on the market, the entire production capacity of NVIDIA's Blackwell GPU for the next 12 months was already fully booked last October. This means that new buyers placing orders at that time would have to wait until the end of this year to receive their shipments.
The robust downstream demand, combined with the relatively weaker supply chain capabilities of its competitors, has led to a market competition landscape where the strong continue to dominate. AMD's market share in the second quarter may continue to decline.
Epilogue
Investing in AMD currently hinges on the potential of its new MI350 series graphics cards to drive sales growth. Its valuation depends on the height of its future revenue ceiling, which is crucially determined by how much of the AI acceleration chip market it can capture.
Since the first quarter of 2023, AMD's market share in the data center GPU sector has been growing at a steady pace.
Although NVIDIA's Blackwell series products dominated the market in the first two quarters of this year, with the launch of AMD's MI355X later this year and ongoing software optimization, AMD is expected to regain some market share by the end of this year or early next year.
However, AMD has only recently begun to make significant efforts and is unlikely to compete effectively with NVIDIA in the short term.
This wave of computing power infrastructure development means a higher revenue ceiling for AMD. In its competition with NVIDIA, AMD positions itself as an additional option in the market.
CEO Lisa Su has previously stated that the AI chip market is large enough to accommodate multiple players, and AMD does not need to defeat NVIDIA to be successful. (End of article)
Always stepping on landmines with your investments?Futubull AI is now live!Precise answers, comprehensive insights, and the ability to seize key opportunities!
Editor/joryn
