For most of the generative AI revolution thus far, the big original equipment manufacturers, or OEMs, have been sidelined as Nvidia and now AMD have done direct allocations of their GPU compute engines to hyperscalers, cloud builders, and other lighthouse customers. But if the second AI wave is going to build and generative AI is going to go mainstream, now Nvidia and AMD have to start allocating more GPUs to the OEMs so they can drive revenues and, perhaps, make a little money, too.
The OEMs, of course, are the traditional server makers from back in the dawn of time – like Dell, Hewlett Packard Enterprise, Cisco Systems, Lenovo, and IBM – who tend to sell machinery to small, medium, and large enterprises as well as tier two service providers, governments, and academic institutions. None of these organizations are large enough to invest in their own server designs and do not have anywhere near the volumes to command the attention of the original design manufacturers, like Quanta Computer, Inventec, WyWinn, Foxconn, and a handful of others, who make more than half of the servers in the world and are happy to do so at extremely low margins.
The key distinction between the OEMs and ODMs is that the former cater to smaller customers buying in smaller volumes, and historically they have also been able to exact higher profit margins from their customers and make it up in customer volume rather than in server volume as the ODMs have to do. Both businesses are tough, and all OEMs and all ODMs are always under intense competitive pressure. Servers are the trucks of the modern economy, and PCs and smartphones are the cars.
Some OEMs have units that behave like ODMs. Dell was the first of these, and famously did the original hyperscaler server designs packing four X86 server nodes into a 2U server chassis for the company that was called Facebook way back when. The Open Compute Project, code-named “Project Freedom,” was founded in 2009 when Facebook, now Meta Platforms, decided it would design its own serving, storing, and networking gear and go directly to component manufacturers for parts and have ODMs build its machinery, in essence creating something that could act like an OEM but be priced like an ODM.
Those designs were open sourced when the OCP was formally launched to the outside world in 2011, and it is very much a reaction to the Great Recession and the desire to cut out the OEM middlemen in the server supply chain. And that was the beginning of the end for Dell’s Data Center Solutions (DCS) custom systems unit, which was pushing over $1 billion a year in custom iron at the time and which was run by none other than Forrest Norrod, who eventually left Dell to help reinvent the datacenter compute business at AMD.
HPE also used to have a big custom server business aimed at service providers, and like Dell, it essentially walked away from this business because the margins were way to low. Lenovo and Inspur still peddling iron ODM-style to customers, and Supermicro is behaving more like an ODM than an OEM these days as it builds AI servers at scale for Meta Platforms and perhaps for Microsoft if the rumors are right.
Dell and HPE have had a hard time in the past few years after the coronavirus pandemic caused a big increase in server spending and then the waning of the pandemic and then the GenAI revolution caused a recession in spending on general purpose servers. Companies had lots of capacity laying around and they deferred server upgrades – despite the very good performance increases coming out of CPUs for the past several years – so they could invest in cloud computing for both generic workloads and AI workloads and, in a minority of cases, spend on premises AI server iron for proofs of concept.
It hasn’t helped Dell, which still has its PC business unlike HPE, that the market for client computing took a dive after a massive wave of pandemic upgrades for PCs.
In the past few quarters, Dell has finally been booking revenues for AI machinery, and its growth in the past quarter was extraordinary for sure. But it remains to be seen that Dell can boost its profits from sales of AI machinery if the general purpose server market continues to slump and price pressures are increasing. As we drill into the numbers and what Dell’s top brass said about them for the first quarter of fiscal 2025, which ended on May 3 and which Dell just reported, as far as we can tell whatever profits Dell is getting from its AI servers are being more than wiped out by a decline in profits in general purpose servers.
It is hard to say from the numbers that Dell gave out how much profits AI servers are generating for Dell and how much less profitable the general purpose server business and the storage business is now than they were a year ago, but this is clearly true if you look at the numbers. And this must be a disappointment for Dell’s managers, who have been waiting – perhaps not patiently – for GPU allocations so they can start profiting from the GenAI wave.
Let’s take it from the top and drill down.
In the May quarter, Dell had $16.13 billion in sales, up 7.3 percent and down a smidgen sequentially. Services revenues, which include tech support as well as the all-important financing when it comes to expensive AI systems, rose by 3.9 percent to $6.12 billion but was also down a couple of smidgens sequentially. Overall revenues were up 6.3 percent year on year to $22.24 million, but operating income fell by 13.9 percent to $920 million. Luckily for Dell, it had a $408 million tax benefit in its back pocket, so it was able to boost net income by 64.7 percent to $960 million, or 4.3 percent of revenues. But clearly the underlying business is not as healthy as you might have expected if you were just looking at the net income number. (Which is one of the reasons why Dell’s stock is getting hammered as we write this.)
Thanks to booking $1.7 billion in sales for what Dell calls “AI Optimized Servers,” which means servers loaded up with GPUs aimed at running AI workloads, Dell posted a 42.5 increase in sales for its Servers and Networking division, hitting $5.47 billion. Storage sales were up a tenth of a point to $3.76 billion. Add it up and the Infrastructure Solutions Group had $9.23 billion in revenues in fiscal Q1, up 21.5 percent. Operating income for ISG was only $736 million, however, down a half point from the year ago period.
Dell has talked about the AI server business “being margin rate diluted, but margin dollar accretive,” which is a clever way of saying that incremental business is not as profitable as other parts of the Dell systems business. When pressed on the AI profitability issue on a call with Wall Street analysts, here is what Jeff Clarke, Dell’s vice chairman and chief operating officer, had to say:
“The performance has been largely driven by enterprise, large accounts, large bids, acquisition – and those are competitive environments. Again, we look at the long-term view of them, and winning new customers in the datacenter serves us well over the course of a customer’s lifecycle with us.
And then specifically in AI, we’re not the price leader in AI. The engineering and the delta we have in the capabilities of our product time and time again, we are not the low-cost provider, the low price in the AI deals that I’m involved in, which are most of them. We are getting a premium for our engineering in the marketplace. We’re getting an advantage, but these large deals when you’re talking clusters of tens of thousands, hundreds of thousands of GPUs, are our competitive environment, but we’re not the one running the price down.
We are, again, getting a premium for the value that we have generated or created into our products. And then conversely, I would tell you in the smaller deals, our margins are substantially better as you would expect. So in enterprise deployments that are smaller deals, our margin rates are significantly better than they are in the very, very large opportunities. And as enterprise continues to deploy, I think that bodes well for us over the long-term.”
Dell and its OEM rivals must be hoping with all of their hearts that small and midsized companies will want to deploy pretrained AI models on premises with a modest collection of GPU servers for security and sovereignty reasons. Because that is how the OEMs make money – by selling high tech at a premium to smaller organizations who will pay that premium because they really have no choice.
If you do the math, AI server sales at Dell were up by a factor of 10X year on year and 2.1X sequentially, which is pretty good. Dell has a $3.8 billion order backlog and a sales pipeline for AI servers that is multiples of that backlog (what kind of multiple was not specified). That backlog is up 11.5X compared to this time last year, we estimate, and is up 31 percent sequentially according to Dell.
But no one can tell how margins are doing in general purpose servers versus storage versus AI servers. Dell does not break these figures out.
What we can tell you is that sales of general purpose servers are still way lower than even in the post-pandemic era, as you can see above, it is only now pulling out of a recession in general purpose servers, which had sales of $3.77 billion, up 2.7 percent year on year. When you do some guesstimating based on AI server sales that Dell has given out for the past three quarters and do reasonable trends backwards for the fairly limited AI server sales that Dell had back into fiscal 2022 and fiscal 2023, you will see that Dell had four straight quarters of severe double digit declines, defining its fiscal 2024, for general purpose servers. This was as AI servers started to kick in last summer, when supplies of Nvidia “Hopper” H100 GPU accelerators started to loosen up a little. If you add up all of those quarters in fiscal 2024, general purpose servers accounted for $15.82 billion, down 22 percent from fiscal 2023.
That, ladies and gentleman, is a server recession, and a pretty bad one at that.
Dell has to wait out the AI revolution unlike HPE, which had massive exascale cluster deals for HPC centers that are building large hybrid AI/HPC machines on AMD Epyc CPUs and Instinct GPUs, which filled in HPE’s server recession gap a bit. Dell has always been a strong Intel partner and is similarly a strong Nvidia partner, and it has had to be patient with modest PowerEdge machines with two or four Nvidia GPUs (matched to spread its GPU allocations across the largest possible number of machines) and wait for Hopper GPUs to be more available to sell its high-end, HGX-based eight GPU servers.
There is a reason why Michael Dell was in the front row at the GTC 2024 conference and got a shout out from the keynote by Nvidia co-founder and chief executive officer Jensen Huang, and there is a reason why Huang went to the Dell booth and was chanting the company’s name to fire everyone up. HPE is AMD’s preferred GPU partner among the OEMs right now. And given the tight allocations that AMD has for its MI300X and MI300A Instinct GPUs, that might not be enough leverage for HPE to have a big AI server business without some help from Nvidia, too.
If we had to guess, we think that the AI server business is profitable for Dell to a certain extent, and that other servers and storage are both struggling. If you have been waiting for Dell to deliver your AI machines for six to twelve months, or more, you are not going to be able to go to another vendor and get a better price with GPU supplies still constrained.
But here is the point: When GPU allocations do ease, then pricing gets softer by definition, and if history is any guide, then it will be the OEMs of the world that take most of the profit hit, not the chip makers like AMD and Nvidia. So maybe Dell (and HPE) might want just enough GPUs to maximize profits for a quarter or two. . . .
Looking ahead, Dell says it can boost revenues to between $93.5 billion and $97.5 billion in fiscal 2025, with a midpoint of 8 percent growth to $95.5 billion. Dell further expects for the ISG business to grow in excess of 20 percent for the year. At 20 percent growth, that works out to around $40.7 billion, which would be an all-time record for the Dell datacenter business. Of that total ISG business, and depending wildly on GPU allocations, maybe somewhere between $6 billion and $7.5 billion of that could come from AI servers, we think. But, as many have said, it is hard to predict this. GPU allocations are based on whether the end customer has the datacenter power and space to take the machines right now, and if you don’t, you go back to the end of the line and Nvidia or AMD will give those GPUs to someone who can deploy now.
All we know for sure is that Nvidia should be giving more margin to OEMs so they don’t start looking for alternatives.
So, what does this mean for SMCI?
> It is hard to predict this. GPU allocations are based on whether the end customer has the datacenter power and space to take the machines right now, and if you don’t, you go back to the end of the line and Nvidia or AMD will give those GPUs to someone who can deploy now.
SMCI also has been an NVIDIA favoured partner AFAIK. And it is kind of known for its fast deployments. What do you think their margins will be like?
Supermicro is making machinery for Meta Platforms and very likely is the manufacturer of DGX systems for Nvidia. The same rules no doubt apply. Supermicro is getting the allocations of the big customers it is building machinery for as well as its own. You can see its numbers. While it is profitable, the margins are not high. They are getting a little better.