AMD Breaks $1 Billion In Datacenter GPU Sales In Q2

As expected, AMD has once again raised its forecast for sales of its Instinct MI300 series GPUs, and as it has broken through $1 billion in revenues for its “Antares” line of compute engines in the second quarter, it is now expecting to surpass $4.5 billion in sales of these devices for all of 2024.

While the revenue ramp for AMD’s Epyc family of CPUs has been on a steady and slightly exponential climb since the first generation “Naples” processors debuted in 2017, the Antares GPUs are without question the fastest ramping product in AMD’s history, and despite some mutterings about demand issues in Q1 of this year, we appear to be once again in a position where AMD’s supply is not big enough to meet demand.

That demand keeps GPU prices high, and we think pretty consistent with the exorbitant prices that Nvidia can command with its “Hopper” H100 and H200 GPU accelerators, and also with the even higher prices that Nvidia is expected to charge for its announced and yet still forthcoming “Blackwell” B100 and B200 datacenter GPUs.

“On the supply side, we made great progress in the second quarter,” explained AMD chief executive officer Lisa Su on a call with Wall Street analysts going over the numbers. “We ramped up supply significantly, exceeding $1 billion in the quarter. I think the team has executed really well. We continue to see line of sight to continue increasing supply as we go through the second half of the year. But I will say that the overall supply chain is tight and will remain tight through 2025. So under that backdrop, we have great partnerships across the supply chain. We’ve been building additional capacity and capability there. And so we expect to continue to ramp as we go through the year. And we’ll continue to work both supply as well as demand opportunities, and really that’s accelerating our customer adoption overall, and we’ll see how things play out as we go into the second half of this year.”

As we have said, we expect for AMD to rake in around $5 billion with GPU sales, dominated mostly by the MI300X compute engines but with a smattering of MI300A units, and we think AMD is cautiously guiding us up to that number as it gains more and bigger orders and as supply steadily improves from foundry and packaging partner Taiwan Semiconductor Manufacturing Co and HBM memory suppliers SK Hynix and Samsung.

Nvidia has a much more diverse portfolio of datacenter GPUs, and a much larger business, and as such it has a much more balanced set of revenue streams. The large compute engines like the H100 and H200 drive HPC simulation, AI training, and some AI inference workloads, but there are a lot of other devices based on more modest GPUs that end up in datacenters – and these latter devices represent the bulk of Nvidia’s shipments. Moreover, Nvidia’s business is pretty evenly split between GPU compute engines aimed at AI training and those aimed at AI inference, with a small chunk of sales driven by HPC machinery that, in many cases these days, that is also used to do AI training and inference.

AMD, explained Su, is getting a lot of traction with AI inference, which is a big problem with GenAI workloads that require eight or sixteen GPUs working in concert to reply to prompts in a timely fashion – within a blink or two of an eye.

As we wrote about earlier this week, the high performance and fat HBM memory of the MI300X makes it a better choice in some ways than an H100, which has smaller memory and therefore you need more GPUs to run the inference. With inference being somewhat embarrassingly parallel and not really dependent on large coherency domains – tens of devices, not tens of thousands are needed to do the work – then it is logical for companies to start with inference when they deploy AMD GPUs, where companies are trying to make money or cut costs. That doesn’t mean the MI300X GPUs can’t be used for training. They can be, and they will be. In any event, in the prior two quarters, about half of the GPU revenues came from HPC and half came from AI, and in the current quarter, it was mostly for AI, and mostly for inference, according to Su. We think it will balance out over time and be reflective of the market at large.

It remains to be seen if the GPU business can be profitable for AMD, but Su and Jean Hu, the company’s chief financial officer, said more than once that the GPU business would be more profitable than the company average at some point in the future. Which means it is not today.

In the second quarter, AMD posted $5.84 billion in sales, up 8.9 percent from the year ago period. Net income rose by a factor of 9.8X to $265 million, or 4.5 percent of revenues, The company burned a bit of its cash hoard to invest in current and future products, but still had $5.34 billion in cash and securities as Q2 came to a close.

The Data Center group, which sells CPUs, GPUs, DPUs, and FPGAs aimed at the datacenter as the name suggests, had $2.83 billion in sales, up 114.5 percent from the prior year and up 21.3 percent sequentially from the first quarter. Operating income was up by a factor of 5.1X to $743 million, and we think that CPUs account for more of that income than GPUs at this point. That operating income represented 26.2 percent of revenues, which is a lot better than what AMD was experiencing a year ago when profits took a big hit thanks to the server recession and high costs ramping up the Antares GPUs and future “Turin” Epyc CPUs.

The Embedded group, which is dominated by the old Xilinx FPGA business, saw sales shrink by 41 percent to $861 million, and operating income fell faster by 54.4 percent to $345 million. The Gaming group saw revenues declined by 59 percent in Q2 to $648 million, and operating income fell by an even more dramatic 65.8 percent to $77 million. Ouch.

These declines are one reason why the Data Center group comprises nearly half of AMD’s revenues these days (48.6 percent in Q2, to be precise); the other reason is that Epyc CPU sales are doing great, with steady adoption of fourth generation “Genoa” and “Bergamo” Epycs and early shipments of “Turin” CPUs to the hyperscalers and cloud builders.

While these hyperscalers and cloud builders are scarfing up these Epyc chips instead of buying what are usually – but not always – inferior X86 CPUs from Intel, it is important to note that all of the hyperscalers and big cloud builders are either designing their own CPUs and AI math engines or have them in production already. We think over the longest of hauls, half of the CPUs and XPUs in the world will be homegrown, and the other halves of CPU and XPU shipments will come from Intel, AMD, and others.

Our model shows AMD datacenter CPU sales rose by 39.8 percent to $1.67 billion, which represented 2.8 percent growth sequentially from Q1. We also think that the hyperscalers and clouds accounted for $1.19 billion of these sales, with enterprises, telcos, service providers, governments, and academic institutions comprising $485 million in sales, up 62.2 percent. Su said on the call that about of third of Epyc CPU sales in Q2 went to organizations that had never bought Epyc CPUs before. We believe datacenter GPUs comprised $1.02 billion in sales, with NICs and DPUs comprising $65 million in Q2 and datacenter FPGAs making up $75 million in Q2.

The question now is how steep of a ramp datacenter GPU sales can be on and how datacenter CPU sales will hold up. Intel and AMD both expect a recovery in server sales in the second half. That could mean $7 billion in Epyc sales and $5 billion in Instinct sales in 2024, we think. Yes, $12 billion in datacenter sales for AMD is a lot less than the $85 billion or so that Nvidia will probably rake in during the same 52 weeks of calendar 2024. But that $12 billion, should things work out as we expect, will be the most AMD has ever made in the datacenter in its history in a year. And it would also represent more money than all of AMD made in calendar 2020. Which was not that long ago, was it?

More immediately, AMD is forecasting sales of $6.7 billion, plus or minus $300 million, for the third quarter, with the Data Center and Client groups more than making up for decline in the Embedded and Client groups. That is 15 percent growth sequentially and 16 percent growth year on year, and that sequential growth is probably in the same ballpark as what Nvidia will see when it reports its next quarterly results next month.