The GenAI Boom Will Lift Supermicro, But It Will Lift Others, Too

Supermicro reported fiscal Q4 2026 revenue of $10.84 billion for systems, up 92.2% year over year, with AI systems sales reaching $6.69 billion, up 64.2% and comprising 60% of total revenue. CEO Charles Liang noted that 60-70% of revenue came from traditional accelerated machines, 10-20% from agentic AI sandboxing or edge AI, and the rest from traditional servers, storage, and IoT. The company's subsystems business grew 2.4X to $278 million in the quarter.

The GenAI Boom Will Lift Supermicro, But It Will Lift Others, Too It seems like every business, with the exceptions of a few monster companies that gain a monopoly either regulated or unregulated or a piece of an oligopoly where a few almost monopolies control that market and steer clear of colluding, or at least we think they do , is a tough one. Few are tougher than being an original equipment manufacturer, or OEM, that sells a slew of different configurations to a large number of enterprises, or an original design manufacturer, or ODM, that sells a very small number of configurations to the tech titans in very high volume. Both work for very little profit compared to, say, the enterprise software racket, which for now at least can command maybe two, three, or four times the profits out of a given dollar of spending from customers. Supermicro has been bridging the gap between OEM and ODM for a long time, and for its AI systems business, which has exploded in recent years, it really does function more like an ODM than an OEM, although it does have a few thousand customers who like to buy its machines in relatively low volumes as well as channel partners who either resell Supermicro systems or buy components like motherboards, peripheral cards, server enclosures, switches, storage, and racks from Supermicro to tailor systems on behalf of their customers. Yes, Supermicro still has a subsystems business, although it doesn’t talk about it much anymore because it is such a relatively small – but I think still growing – part of its business. Here is my best guess on how the systems versus subsystems business at Supermicro breaks down: This data runs out to the fourth quarter of fiscal 2026, which ended in June for the company. In fact, I think the subsystems business grew by 2.4X in the quarter, to $278 million. The overall systems business, which includes server nodes, rackscale machines, and rowscale “datacenter building block solutions,” as Supermicro calls them, brought in a very tidy $10.84 billion then, by the math, which is a 92.2 percent increase year on year. But, more than anything else, Supermicro has become a go-to maker of accelerated systems based on Nvidia CPUs, GPUs, and scale up and scale out switching, and that is a very large portion of that overall system sales at this point. But, in the fourth quarter, as Supermicro’s sales of AI systems, as it calls machines based on GPU and maybe someday XPU accelerators as well as CPU-only systems that are running agentic AI sandboxes, came to $6.69 billion, up 64.2 percent year on year and comprising 60 percent of its overall revenues. Supermicro founder and chief executive officer Charles Liang said on the call with Wall Street analysts that somewhere between 60 percent to 70 percent of the overall revenues was for traditional accelerated machines GPUs at this point and mostly Nvidia GPUs , with another other 10 percent to 20 percent being for this agentic AI sandboxing or edge AI. The remaining 20 percent or so will be for traditional servers, storage, and IoT systems – the latter being miniscule. The fact is, the transition between the Nvidia “Grace-Blackwell” rackscale systems that ramped last year to the “Vera-Rubin” systems that will be ramping in the coming months, has left a bit of a revenue gap. So this time around, Supermicro sold a ton – ok, we don’t really know how many tens of thousands of tons – of servers to support other workloads at its largest customers, which are being upgraded opportunistically to support the data needs of GenAI workloads and to also consolidate those workloads on fewer, more efficient machines to make room for GenAI accelerated systems. And it almost filled in that AI systems slowdown. Mind you, Supermicro still had incredible year on year growth in AI systems in the June quarter. The traditional HPC supercomputer market was always choppy, and the GenAI supercomputer market is no different, following the ups and downs of budgets that are only loosely coupled to product roadmaps and more constrained by the availability of datacenter space, power, and money to prepay for AI systems and have it all come together at precisely the right time. In fact, some AI system orders that were supposed to land in Q4 F2026 are going to hit in Q1 F2027 which is right now and into Q2 F2027. Supermicro was not precise about this, but taking the former guidance of $11 billion to $12.5 billion for Q4 F2026 and it only did a little more than $11.1 billion, assuming most of the regular, non-accelerated CPU was baked in into that original guidance but perhaps not all of it , that implies that somewhere between $100 million and $1.4 billion in AI systems got pushed. Q1 F2027 is guiding between $14.5 billion and $15.5 billion, with some of that being some Vera-Rubin ramp but also a whole lot of Grace-Blackwell systems, it is reasonable to guess that maybe somewhere between $600 million and $700 million of AI systems got pushed out and some extra effort on CPU systems helped fill in behind that drop in expected AI systems revenue. Add it all up, and Supermicro raked in $11.12 billion in sales in the June quarter, up 93.2 percent. Operating income rose by a factor of 6.6X to $1.49 billion and net income rocketed up by 6X to $1.18 billion. This is not the highest revenue Supermicro has ever had – that happened in the second quarter of fiscal 2026, when it posted sales of $12.68 billion – but this is by far the most profitable the company has ever been in one quarter in an absolute dollars sense and at 10.6 percent of revenues, this is the highest ratio of net income to revenue that the company has ever seen. To be fair, Supermicro hit 10.5 percent of revenues in Q3 F2024, but it was a third the size it is today and that was easier. Supermicro ended the quarter with $7.52 billion in cash and equivalents, with $1.4 billion coming from sales of new common stock to Wall Street and $4.2 billion coming from convertible preferred shares. Given how choppy the AI systems business is, I think it is wise to look at Supermicro on an annual basis as much as on a quarterly one. Here is the annual revenues and net income for the company from fiscal 2010, when the cloud buildout was booming up through the mid-point of the estimated $65 billion to $72 billion that Supermicro is guiding to for fiscal 2027, when the GenAI buildout has gone nuclear: For all of fiscal 2026, Supermicro brought in $39.06 billion in sales, up 77.8 percent. Operating income for the year was $2.77 billion, up 2.2X year on year and net income was $2.23 billion, up 2.1X and 5.5 percent of revenues. Looking ahead to fiscal 2027, at the mid-point of guidance Supermicro expects $68.5 billion in sales, which would be about a 78 percent increase compared to the fiscal year that just ended. I project operating income will grow faster because of higher prices on Vera-Rubin systems and because Supermicro is no doubt getting better at building rowscale systems complete with power and liquid cooling and driving its costs down. This is what Supermicro does, and always has done, from motherboards on up. So I am pegging operating income at that revenue level for fiscal 2027 at $5.21 billion, and net income will be around $4.18 billion, or 6.1 percent of revenues. With more revenue, it will generate more absolute income and slightly more income as a share of revenue. Supermicro has just opened up a new rowscale computing facility in Silicon Valley to work on optical networking and datacenter-scale systems, and with that it has nearly 4 million square feet of design and factory space in the United States alone; it has substantial factories in Taiwan, Malaysia, and the Netherlands as well. Right now, it can produce more than 3,000 direct liquid cooling racks a month up to 240 kilowatts per rack and 3,000 air-cooled racks per month; 500 kilowatt racks are being designed and prototyped now. As you might expect, Supermicro is focused very much on the market for AI systems in the United States, and that stands to reason because tech titans here are spending the most money. Interestingly Supermicro had one customer who alone accounted for 28 percent of revenues in fiscal 2026. This is almost certainly SpacexAI, with the Colossus 2 supercomputer, which is mostly being rented by Google and Anthropic, for an estimated $920 million and $1.25 billion per month, respectively. Supermicro also said that it had nine customers who spent over $1 billion. So that is CoreWeave and seven other companies. Those nine customers are around half of Supermicro’s revenue in the year that was just closed. The thing is, as fast as Supermicro expects to grow, Dell, its OEM and quasi-ODM rival here in the United States, will have $60 billion in AI system sales in its fiscal 2027 ended in January of next year, and maybe on the order of $120 billion in overall datacenter sales, where it has a huge enterprise customer base numbering in the millions of companies worldwide. If you normalize Dell to Supermicro’s fiscal years, the gap will be even larger between Dell and Supermicro. But Dell and Supermicro will have about the same sized AI systems businesses, at around $60 billion each. It all comes down to CPU and GPU allocations, and we all know who controls them.