Well Smorgasbord1 (and others), thanks for the education.
I must confess that desperate was my poor word choice, and I was just passing on info I thought might be interesting to someone clearly not as educated as yourself.
For the record, I was an early purchaser of SMCI am a big fan and still hold half my shares even after this parabolic and technically dubious ride up. I was not trying to trash the stock, just providing an additional perspective that if another manufacturer wanted to enter the space it would be easy from a design standpoint.
As I posted earlier in this thread, cutting prices is one way to keep competitors from even deciding to enter the space. And that’s all I was trying to do - provide an additional framework for considering the reduction in margins for SMCI’s latest earnings report.
Thus, I stand by this post and the video as interesting, even though I disagree with the person’s conclusion. And now we all know what only you knew before - that this is common practice. I am going to go to the podcast author now and ask him why he would make a big deal of this…being common practice.
@mizzmonika Let us know what Jose Najarro comes back with. I watched this video as well before you posted it and I, too, was confused by his argument, since there are lots of companies making servers, but I’m not an expert in the space.
Something to keep in mind with Super Micro is they are investing massively in research and development and have been for many years. They’ve mentioned many times that over half their employees are engineers and with a total of 6,000 employees or so, they hire nearly 3,000 engineers.
Breaking it down by spend metrics for the last quarter,
R&D 109M
Sales and marketing 47M
G&A 38M
So they company is spending more on research and development, then the rest of the company combined!
Gross profit for the company was 564M so the company is already GAAP profitable by a wide range.
The low margin of this business is not attractive for competitors. Super Micro is undercutting themselves on price somewhat to gain market share. I believe on the call they are even taking on some loans or leverage to try and gain further market share.
Their competitive advantage lies in their building blocks solution which was developed in house by all these engineers. The components can be put together in anyway the customer specifies. They also have about 20 predefined machines for specific use cases. One of these is an AI optimized server. This technology put together is more advanced than just a GPU being stacked together. They lead the industry in cooling and energy efficiency.
They can effectively build specs with any chip maker. If Nvidia is no longer the leader in 5 years, Super Micro could be partnering with that new chip maker.
This technology can not simply be copied as some are suggesting. It’s the work of a massive R&D team building a competitively superior product over the long term. They can provide the best product in the market and the lowest total cost of ownership for the server. It’s no wonder Meta is coming to them for what sounds like most of their next gen compute needs.
I just re-read the transcript of SMCI’s latest quarterly CC and picked up on something I didn’t really notice the first time. The fact is that the vast majority of their rack sales are air cooled rather than liquid cooled.
The reason is two-fold. First the infrastructure needed for liquid cooled systems is simply not yet developed at most data centers. SMCI had the goods, but most customers are not prepared to purchase it yet. And so far, liquid cooling is not yet necessary. At present most of the racks their customers are ordering run at 600 - 700 watts. This power level can be adequately cooled with the air conditioning units that are commonly installed in these facilities. However, as the H100 racks become the dominant product, liquid cooling will be mandatory as these racks require a kilowatt which generated too much heat for air cooling.
This seems to imply that the demand for large scale AI support has not really kicked in yet. There’s a lot of the market yet to hit the order book.
One thing to keep in mind is that SMCI’s portion of R&D spend to Revenue is very small. It was only 3% last quarter and it has never been above 7% going back to 2020. With that being said it always is more than S &M and G&A. Course we are talking about a company that only has GM of 15.4% last quarter and not above 18 percent since 2020. This is a low margin business and they have to be very mindful of expenses so it is their sales volume that holds up their profits.
SMCI apparently being added to the SP500. Not sure what the net effect from tracking funds buying it and mid cap funds selling it will result in, but seems to be good news in terms of medium term buoyancy in the share price
@buynholdisdead, Yeah, you can look at it as the glass is half full, or half empty. You have highlighted the half empty view. OTOH, as @wpr101 pointed out, of the 6,000 employees at SMCI more than 3,000 are engineers. The R&D spend is more than 50% of total operating budget.
So, as you pointed out, as a percentage of revenue, R&D is not a very much of it. But in fact, the entire operating budget is not very much of it. The bulk of SMCI’s budget goes to inventory and capital expense. It’s kind of inherent to the business. Maybe a more informative way of looking at it would be to compare SMCI’s staffing and R&D spend as compared to Dell or HP or any number of their competitors (I’ve not made the comparison as I just now thought of it). Keep in mind, Supermicro is a widget company, not a software business.
@brittlerock, I think it would be more informative to think why their R&D is so low? You can’t compare it to HP or Dell because both of those comparison’s are not just pure plays. Here is a look at Dell’s 10q,
as you can see they do not break out their costs. Now you might be able to Ferret it out if you keep digging but that, for me, wouldn’t tell me anything useful.
The reason SMCI does not have to spend that much on R&D is because all of the chip companies, NVDA, AMD, Intel, all give them their plans on how to implement their chips into their servers. That is why SMCI does not have a moat. The same plans are given to Dell and HP. What SMCI has is a singular focus on the business and they implemented the plans first. Dell last quarter only had 800 million in Revenue compared to SMCI’s 3.6 billion in Revenue. So SMCI’s first mover advantage will dissipate and is something we all need to watch.
I think you’re both right and wrong. I’ve seen this point made several times in past discussions about SMCI, and I think it grossly understates the challenge of successfully developing a new product.
My background spans over four decades of inventing new components and instruments. As a result, I’m very familiar with the engineering details that different manufactures provide so that other companies can successfully integrate their products into something new. Those details are really important. They include information about idiosyncrasies and capabilities that might require a much more complicated interface than would otherwise (naively) be expected. However, while necessary, this information is far from sufficient.
Every other company has access to information like this in every high-tech industry, but very few manage to use it to develop a really strong product. It’s not just the performance of the actual device, but also the stability and robustness of its operation, the ease of integration, the reliability and cost of operation, and the price of manufacture, assembly and test. Together, these combine to make a big difference between winners and losers, and so establish a meaningful moat.
Any company that can develop a product that is in such demand is doing something very right. If they didn’t have a moat, they would already be being undercut, and outsold, by many other companies. It’s my assessment the SMC has done an outstanding job and will prove a tough competitor for every other company trying to operate in the same space. It’s not as formidable a moat as having complete ownership of a technology, but that’s something that rarely happens. Even Nvidia can’t make that claim.
I have concerns about the valuation of SMCI, But none about their ability to compete and win. Disclosure: I still have a small amount of SMCI having sold during this latest run up, and I’m also long in Nvidia.
I agree with you Larry that they have to be able to implement the plans into their product but the only explanation of why their R&D is so low, can only be that the R&D for implementing it does not take as much as designing the whole product from scratch. Look at NVDA’s R&D. That is what a real budget looks like for designing something and building it. I am not trying to take anything away from SMCI, 13 percent of my portfolio has been in it for awhile, I am just trying to be realistic. The heavy lifting is being done by NVDA, AMD, and Intel.
Another point that people bring up is that SMCI has liquid cooling but here is the problem with that statement. The DATA CENTERS do not have liquid cooling. It won’t be SMCI that develops the liquid cooling for the Data Centers. It will be the Data Centers that do that. So I suspect the Titans will develop it first and then the other Data Centers will follow. The edge Data Centers will need it also. In order for AI to really take off, liquid cooling is a must, and since we haven’t even started building that out in the Data Centers, that tells you how early this is.
One more point, Ethernet will not be able to handle AI speeds due to the latency. That is why NVDA developed Infiniband and handed the plans over to the server companies ie SMCI to develop into their servers.
One of InfiniBand’s most significant advantages over Ethernet is lower latency. However, newer Ethernet standards are closing the latency gap with InfiniBand. Ethernet has an advantage in scalability and may be more cost-effective in various applications. Also, I think the big cloud providers are eventually going to want multi vendor for thier GPUs and those other vendors will likely need Ethernet. It’s why I continue to monitor ANET. If Ethernet closes the gap with InfiniBand or if the big Cloud providers start going multi-vendor to reduce their dependency on NVIDA, I’m thinking, ANET will have a huge NVDA and / or SMCI like opportunity.
The moat for SMCI is quality at a fair price. I worked with them as a QA director for 7years while at Juniper Networks and they provided our SSLVPN appliances in different configurations of price/performance. Secure Micro always beat out Dell and others. How? They would offer many reference configurations that they had customized for others and would tweak for us. These were real world, proven configurations, not textbook examples from Intel or AMD. And SMCI would provide the number of units sold and defect rates with root causes. Imagine the safety this provided us as a vendor with thousands of installations requiring upgrade. And their prices were always competitive.
Now why did I not follow others here and buy SMCI? Argg! guess I am just not that smart and still learning. I got hung up on the low margins. You can’t win them all but I am happy with my recent gains in ELF, CELH, SWAV, FOUR, NVDA, and NU. Fool on!
I’m no networking guy, but some reading I’ve done leads me to a number of candidate beliefs about “latency” claims surrounding InfiniBand:
When InfiniBand’s superior latency is discussed, the claim is not that end-to-end AI workloads always complete faster on InfiniBand. Rather, the “latency” advantage of InfiniBand is only claimed in terms of a single-hop, point-to-point latency measurement.
If you look at the specifics of AI workloads, they are not a predicable assembly of constant, single-hop, point-to-point data movement. So the methodology behind InfiniBand’s “latency” claims is not very useful.
In fact, AI workloads involve a lot of nuanced and complicated logic required to make decisions regarding what data goes to what GPU cluster, and in what sequence.
AI workloads are characterized by lots of “microbursts” of data, subject to 3) above, and CANNOT be measured/modeled based off some assumption that point-to-point latency is the only valid measurement.
The MOST valid measurement is how long a AI workload takes from beginning to completion.
Ethernet is a superior framework for solving the problems of 3) and 4), above
Even at the speeds Ethernet CURRENTLY supports, it’s a potential contender as a valid solution in the back-end GPU network
Because of $ANET’s superior software architecture, it’s able to program the relevant parts of its framework to better handle the logic demands of AI workloads, making $ANET the superior supplier of Ethernet solutions to 3) and 4) above,.
That is correct intjudo because once you put the equipment on it that adds latency that you can not know. So, while someone might measure latency across the network it can change depending on the hop you take.
Infiniband was developed to be used inside the data center and latency is very germane to the discussion.
I do not think this is correct. There are to many nuances and differences to know. All AI workloads are not the same. It makes more sense to measure your hops and compare them to the end point. Then you could realize which equipment was causing you the most latency.
Ethernet has been around for a long time and is due for an upgrade. It isn’t that it is a bad technology, it isn’t. It won out in a long list of topologies. But as the world changes you wouldn’t expect it to be around forever.
Anet, Cisco, Ciena,etc do not care what they use. They only care that they supply to their customers what they need. So if Infiniband or some other technology comes along they will make their input/output ie NIC cards compatible with that technology. After all, Infiniband isn’t going to change how they implement the technology, but, when the customers examine the hops for latency they will give their money to the business that supplies the fastest and least latency.
@buynholdisdead SMCI addressed the migration to liquid cooling during their most recent quarterly CC. It’s not complicated.
Most of the racks currently sold demand 600 - 700 watts. These racks are air cooled because cooling can be adequately addressed by the HVAC systems commonly installed at data centers around the globe.
However, as the data centers migrate to Nvidia H100 as the dominant processor used in rack installations liquid cooling will become mandatory as these racks run at 1.0 - 1.1 KW. The added power throws off enough additional heat such that the HVAC systems are simply inadequate for the cooling requirements of these higher power racks.
Yes, you are correct, the majority of data centers do not have the appropriate infrastructure to handle the power demands of these racks. Not just the cooling requirements but the electrical requirements as well. A lot of upgrades will be required before installations of liquid cooled devices become the standard. But, once the data centers upgrade their infrastructure to handle the increased power requirements, SMCI is fully prepared to work with them during the design phase. They will base their recommendation on real hardware characteristics that can be delivered in short order, as opposed to the theoretical devices likely to be proposed by their competitors.
I will have to disagree with this statement. If it was not complicated they would have already implemented it. Liquid cooling has been around for about 100 years.
But none of that is the important part of what I was trying to state.
If liquid cooling is just starting out, and it is, and if they need liquid cooling to keep the servers cooled with AI, and they do, Then this is just beginning and everyone who thought they missed out haven’t.
@buynholdisdead I’m not sure I understand what we are in disagreement about. Maybe I just didn’t say what I was trying to convey very clearly. What I meant is what I paraphrased from the quarterly CC. Understanding the requirement is not complicated. The migration to data centers that support AI at scale demands racks that consume 1.0 - 1.1 KW. These racks generate more heat than the current 600 - 700 Watt racks. The additional heat is too much to be addressed by air cooling. Therefore, liquid cooling will be mandatory. That’s not difficult to understand and it’s not a complicated requirement. But, by no means does it mean that the physical implementation will be simple.
@intjudo asked what will need to be addressed in order to migrate to liquid cooling, electrical, HVAC and anything else. I am not an expert on this, I have zero experience in this area, but I can guess that there’s a lot more than electrical and HVAC upgrades. The physical spacing between the racks may need to be altered, though that doesn’t sound like a big deal, actually that would be a very big deal.
Data center fire protection uses freon gas rather than a water sprinkler systems. I don’t know exactly what liquid is used with respect to liquid cooling, but I’m reasonably certain that it’s not water. My guess is that it is a proprietary fluid of some sort that won’t damage the equipment if a server hemorrhages. It wouldn’t surprise me if plumbing (by that I mean physical pipes), pumps, valves and so forth might be required. I just don’t know, but keep in mind that we’re talking about hundreds, maybe thousands of devices. This is something more than a refridgerator.
Again, I’m not knowledgable about this stuff, I’m just speculating. So like I said at the outset, understanding the need for liquid cooling is not complicated. That doesn’t mean that the implementation is simple.
We are not in disagreement. I agree with you completely.
Great question injudo and one that I have been thinking alot about. There are many ways to do liquid cooling and we have to figure out which one is most likely to win. I am going to look into how smci wants to implement it, NVDIA wants to implement it, AMD wants to implement it. Then look at META, AMZN and GOOG to see if they have started to implement it. That should give us a lead.