Sure, artificial intelligence (AI) stocks have seen a boom in their stock prices, but unlike many past bubbles, the hype around AI seems to be the real deal -- at least for the most part. Bubbles tend to form when stock prices go up a lot without a requisite jump in revenue and profits, but Nvidia's (NVDA 6.18%) blowout guidance given in May 24 shows the AI trend is translating into real revenue and profits for well-positioned companies.
Of course, where there's strong profits, competition will follow. That's why June 13 will be a pivotal day for Advanced Micro Devices (AMD 2.37%), which is looking to challenge Nvidia's AI dominance. Yet no matter who wins this race, a third tech company apart from these two is bound to benefit either way.
AMD's MI300 to take on Nvidia's H100
June 13 is the day when AMD will hold its AI Technology and Data Center presentation, in which it will unveil its AI platform and strategy. Many think it's likely AMD will reveal details of its new MI300 accelerated processing unit (APU), which will take on Nvidia's H100 systems in AI servers.
While the event is Tuesday, some specs have already been rumored to the tech press. Tech website Tom's Hardware derived information on the MI300 from a presentation on the Department of Energy's El Capitan Supercomputer, which will use the MI300, and a recent presentation from AMD Chief Technology Officer Mark Papermaster. According to Tom's, the MI300 will feature 13 "chiplets," with some stacked vertically in a 3D structure, combining the company's Zen 4 EPYC CPUs, CDNA 3 GPUs, and eight stacks of high-bandwidth memory (HMB). Overall, the chip system contains 146 billion transistors.
By comparison, the Nvidia H100 contains 80 billion transistors, but remember, that is only for the single GPU. The MI300 will probably be up against Nvidia's own Grace-Hopper superchip this summer and beyond, which incorporates not only Nvidia GPUs but also its new Grace CPU tailor-made for AI processing.
Image source: Getty Images.
AMD has an opportunity, if it can seize it
Prospective customers will no doubt be very eager to hear more details around speed, latency, and capacity of the MI300. According to a Digitimes article on June 8, data center customers are clamoring for an alternative to Nvidia, which currently dominates the AI GPU market and can therefore charge exorbitant prices.
With the advent of OpenAI's ChatGPT last November, the capabilities of AI seemed to break new ground, setting up an arms race among leading companies across a host of industries for these new AI capabilities.
According to Trendforce, the AI server market should see strong 38.4% growth this year, despite a projected decline in the overall server market. Over the medium term, Trendforce sees AI server shipments growing at a 22% average growth rate between 2022 and 2026, going from 9% of overall server shipments this year to 15% of shipments by the end of that time period. Yet given the high prices of Nvidia H100-based servers today, the percentage of AI server revenue is probably much higher than the mere shipment numbers.
Currently, Nvidia dominates this high-profit, high-growth portion of the market, with between 60% and 70% market share in AI servers. Another 20% comes from custom application-specific integrated chips (ASICs) designed by the cloud server providers themselves, such as Amazon's Inferentia and Trainium chips or Alphabet's tensor processing units (TPUs).
Data-center operators would love to see a viable third-party competitor to Nvidia, which could help drive down the pricing for these now need-to-have AI chips. So if AMD can come through, it would be a huge plus for AMD and a challenge for Nvidia. Although even an oligopolistic industry with two or three or players would probably yield good profitability for each.
Some analysts are catching on; last month, Morgan Stanley analyst Joseph Moore adjusted his AI revenue outlook for AMD, saying that it could be "multiples" higher than first believed.
And one company that will win no matter what: TSMC
No matter how much of an inroad AMD's MI300 makes against Nvidia's H100, foundry Taiwan Semiconductor Manufacturing (TSM 1.26%) makes both chips on its most advanced 5nm process. So for defensive investors, Taiwan Semiconductor may still be the better way to play the trend, as it currently makes the overwhelming majority of leading-edge chips. Recently, TSMC announced it would be raising its prices to customers between 3% and 6% next year for its advanced wafers, and it's likely to run into little resistance.
No wonder Taiwan Semiconductor caught Warren Buffett's eye last year. Although Buffett wound up selling out of his position because of Taiwan Semi's proximity to China, assuming you're OK with this geopolitical risk and are sufficiently diversified, TSMC may be a good way to benefit from this burgeoning AI chip war as well.
