What Intel calls “cloud digestion” as the cause of the massive pullback in spending in its Data Center Group is looking more and more like a case of “Epyc indigestion” for Intel, not for the hyperscalers and cloud builders. And the top brass at Intel should be thanking the heavens for their good luck that capacity for advanced fab processes has been severely constrained at the same time Intel struggled to get its server chip act together. If that had not been the case, AMD might be truly cleaning Intel’s CPU clocks.
In the first quarter ended in March, AMD was hitting on all cylinders, with its client, server, and game console businesses all up significantly year-on-year as well as up sequentially, delivering a stunning 92.9 percent revenue growth, to $3.45 billion, and an amazing 3.4X increase in net income, to $555 million. All of this good business helped AMD increase its cash hoard by 2.3X to $3.12 billion, which helps cushion the blow on that pending $35 billion acquisition of FPGA maker Xilinx. When you start using “X” instead of “percent” on numbers moving up and to the right, that is always a good sign.
We have contended for years that AMD itself was a little skittish about overcommitting on its wafer supply agreements with Taiwan Semiconductor Manufacturing Corp, particularly given its history during the Opteron server chip era with its own foundry and its spinoff into GlobalFoundries, when it took some pretty big financial hits to negotiate wafer supply agreements down as Intel was ramping up the Xeon server attack during and after the Great Recession. But. On a call with Wall Street analysts discussing the financial results, Lisu Su, AMD’s chief executive officer, said that AMD had managed to negotiate more capacity from TSMC and was able to up its revenue forecast for the year by $1.2 billion to $1.3 billion – most of which we think is going to come out of Intel’s hide in both client and server CPUs.
On the server front, it is easy to assume that all of AMD’s success has come from the rollout of the third generation “Milan” Epyc processors, but that is not the case according to Su, who said on the call that server CPU sales were “more Rome weighted in the first quarter compared to Milan, but there was good growth in both.” Heading into the second quarter, AMD is projecting “good growth” for both Rome and Milan, with Milan growth accelerating, and by the third quarter Milan will cross over and bypass Rome. It is always better to have two things to sell, and the pricing across the two generations works itself out with discounts on the older stuff and premiums on the new stuff, as necessary.
When talking specifically about money for server chips, Su said that Epyc CPU sales more than doubled year over year and grew by a “strong double digit percentage” sequentially from Q4 2020. Each Epyc launch has ramped faster than the previous one, and Su added that “2021 marks an inflection point in terms of scale, ecosystem support, and customer adoption for Epyc and Instinct processors.” She added that overall datacenter product revenues more than doubled since this time last year and represented a “high teens percentage” of overall revenue, and that datacenter revenues will grow faster than AMD overall as we progress through 2021. That means datacenter could be as high as 25 percent of sales, we think, as AMD exits 2020. That’s starting from essentially zero five years ago, and that is remarkable and, we always believed, perfectly plausible because markets support competition if competitors bring it.
In the quarter, AMD’s Computing and Graphics group posted sales of $2.1 billion, up 46 percent, with operating income of $485 million, up 85.1 percent. The Enterprise, Embedded, and Semi-Custom business was up by 3.9X to $1.35 billion on the strength of Eypc CPU sales; game console chip sales (which we think were down a tiny bit). Instinct GPU accelerator sales (which we think declined year-on-year) are reported in the Computing and Graphics group, and presumably some of the other GPU cards are added into “datacenter sales” for AMD, too.
If we take what Su said and then apply some model magic to it, we figure that total AMD datacenter sales were up 104.4 percent to $675 million in the first quarter, which represented a 9.5 percent increase sequentially from Q4 2020. We think Epyc CPU sales were up 141 percent year on year to $608 million, up 11.7 percent sequentially, and that Instinct GPU sales as well as some visualization GPUs that get lumped into the datacenter number were off 14.1 percent to $67 million.
Let’s have some fun with math now. AMD’s Epyc CPU sales were up $356 million by our math comparing Q1 2020 to Q1 2021, and Intel’s Data Center Group sales were off $1.43 billion. Some of Intel’s revenue decline is due to increasing price competition from AMD (and some from homegrown Arm servers at Amazon Web Services), and some of it is due to the fact that a CPU sale lost also takes away chipset sales and in some cases a motherboard sale, so there is a multiplicative effect. Intel can counterbalance that by tossing in network interface cards or FPGAs into a CPU and chipset sale for a bigger bundle, but at a much-reduced price. Our point is that a dollar gained by AMD is not a dollar lost to Intel – it is many dollars lost. And, as we have said before, this sets a new ceiling for pricing and that money per unit never comes back. Ever. Ask the failed RISC/Unix server businesses of Sun Microsystems and Hewlett Packard.
The AMD effect on operating income is much more dramatic. It is hard to allocate operating income to Epyc CPU sales, but the Compute and Graphics group, which is sells client CPUs and APUs as well as graphics cards, has been running at around 60-ish percent operating income. We think Instinct GPUs and Epyc CPUs are probably somewhere around 50 percent, which is Intel’s peak in the Data Center Group, so that should be on the order of $304 million in Q1 2021, up from around $126 million in Q1 2020. That’s an incremental $178 million in operating income. If the operating income is higher because AMD charges a more fair price to begin with, fine. Call it a $200 million boost in operating income related to Epyc CPUs.
Intel’s operating income in Data Center Group is down a stunning $2.22 billion year on year in Q1 2021. Every dollar in operating income that AMD has gained has cost Intel about 11X that amount in lost operating income. You have to adjust that for increasing costs on 10 nanometer and 7 nanometer processes for the Xeon line and other factors. But even so, the AMD effect could be a 5X down draft on operating income for Intel’s Data Center Group for every dollar that AMD gains. And on the pure CPU only revenues, that could be against a 2X to 3X revenue multiplier, which is the difference at list price Intel has enjoyed over Epycs for the past couple of years. That means every CPU dollar AMD gains hurts Intel by 2X to 3X, and then it when other factors are thrown in – lost NIC sales, lost FPHA sales, lost motherboard sales, lost chipset sales – that can be a 5X loss multiplier, which translates into a 5X operating income decline multiplier.
This is what we expected, and talked about, six years ago when AMD said it would step back into the server ring. And this seems to be what is happening.