Cisco Can Finally Sell Lots Of Supercomputers And Their Networks

Cisco Systems reported Q4 fiscal 2026 revenues of $17.25 billion, up 17.6% year-over-year, the highest quarterly revenue in company history, driven by AI infrastructure demand. The company's Silicon One ASICs and Acadia optical transceivers are key to selling UCS systems enhanced with GPUs and XPUs for AI clusters, positioning Cisco to sell large AI systems to neoclouds, sovereigns, and enterprises.

Cisco Can Finally Sell Lots Of Supercomputers And Their Networks There are those that believe most or all applications – including AI inference models, which are a kind of massively structured relational database management system if you expand your view a bit – will inevitably end up on the big clouds, and those that believe companies will embrace this new technology and deploy it in their own datacenters or those controlled by third party co-location services who act on their behalf. In my mind, the glass is neither half full nor half empty when it comes to cloud versus on premises, but rather we need three different glasses to pour three different stiff drinks. Because we are going to need them as we sort out the technology and the economics of these different approaches to corporate computing. The one thing that we do know is that eventually the GenAI technologies that are being created by the big AI model builders will trickle down and this will create a massive need for supercomputing-style systems that can deliver sustained performance for chewing on and spitting out tokens that is akin to the performance that is required for – and delivered by – mainframe-class online transaction and batch processing application in the back office of the enterprise. And the shift from chattybot farting around to true agentic AI is going to be profound, and the hardware and systems software to deliver it are going to be expensive. This is what usage looks like: This chart above, released by Cisco Systems during its financial presentation for the fourth quarter of its fiscal 2026 year ended in June, could easily describe the difference between online and batch workloads on a System/360 mainframe over the past six decades, or between web infrastructure workloads and data analytics in the past three decades. This is ever the way, since applications always push hardware envelopes, but it is different in that half of corporate computing is now based on what is undeniably a supercomputing architecture supremo – with a price tag to meet its exotic architecture and high performance. For Cisco, which got into the converged switch-system business in 2009 when the Great Recession was roaring and built up a very good and sizable systems business with a $4 billion run rate – as big as the individual systems businesses of Sun Microsystems, Hewlett Packard, Dell, and IBM in a very short time. We do not expect for Cisco to become an ODM-class AI cluster maker as Supermicro and to a certain extent Dell, Lenovo, and Inspur have done. But we do expect it to be able to sell a few large AI systems to neoclouds, sovereigns, and large enterprises who already buy its UCS systems and its datacenter and campus switching infrastructure. The company’s move into the merchant silicon market with its Silicon One ASICs, which converge switching and routing down to a single architecture but which allow for tuning for these workloads as well as variants for scale up networking to share memory across GPUs and XPUs in AI clusters, has been transformative for Cisco’s datacenter business and is, I think, a key to selling UCS systems enhanced with GPUs and XPUs much as trust in Cisco for datacenter switching and routing was foundational to the original UCS business two and a half decades ago. A modern AI rackscale system bears much resemblance to a converged UCS chassis back then. Everything is just beefier and hotter and more expensive. Silicon One has been enhanced with the Acadia optical transceiver business, and this is driving a lot of the company’s AI sales because good transceivers running at 400 Gb/sec or 800 Gb/sec are hard to find. More on that and what is happening with Cisco’s AI supercomputing business in a moment. For now, let’s start at the top of the company’s numbers for Q4 F2026 and drill down. In the June quarter, Cisco’s overall revenues came in at $17.25 billion, up 17.6 percent and the highest quarterly revenues in the company’s very long history. Operating income grew faster than revenues, up 23.9 percent to $4.26 billion, and net income grew even faster at 36.7 percent to $3.86 billion. This is the distribution of growth that every IT supplier craves. The company ended the quarter with $15.92 billion in the cashbox and $46.7 billion in revenue backlog across its products. That backlog has been kinda flattish, as has the cash pile in the past year two years, but both may be poised to grow during the late 2026 and 2027 GenAI spending that is expected. These are still very, very good revenue and profit numbers for Cisco despite that the hyperscaler and cloud business it is doing in systems is a drag on the company’s overall business. There is a dearth of 400 Gb/sec and 800 Gb/sec switching and optical transceivers out there thanks to the GenAI boom, and Cisco has figured out how to benefit from it selling datacenter networks or components for them as well as upgraded campus networks to feed the AI beast. The mix of AI systems will change as the quarters come and go, as has always been the case with the supercomputing racket. Hell, Cisco may even be able to sell machines to some of the big HPC and now HPC/AI national labs – stranger things have happened. And someone has to keep Hewlett Packard Enterprise honest as well as Dell and Lenovo, who all have sizeable proper HPC businesses. You can see the big jump in Cisco product revenues after the launch of the “California” Unified Computing System platform in fiscal 2009, which helped Cisco grow its product business by 50 percent to a run rate of around $9 billion per quarter. Some other things helped there, too. Like Splunk and merchant chip silicon and other software. Cisco now drives more than half of its revenues from software and services, which has been a company goal since 2016, and as you can see, the services business has grown very steadily and doubled over the term of our coverage. These transitions have not always been easy for Cisco, but difficulty is just something that happens to all incumbents when conditions in the market change. Like IBM, HPE, Dell, and Lenovo, Cisco has learned to pivot to meet the market where it is. At the front of that chart, you can see the product revenues reach up towards $14 billion a quarter, and that incremental $5 billion or so a quarter is coming from the GenAI boom either directly or indirectly. Indirectly means customers are upgrading front end Ethernet networks in the datacenter or upgrading campus networks that hook users into the AI datacenter. Cisco has not broken out the UCS business separately from switching and routing systems for many years, so it is hard to guess how much revenue and profit this business is individually driving. I strongly suspect that Cisco has in excess of 95,000 UCS customers worldwide, and while they tend to upgrade incrementally as do customers of IBM, HPE, Dell, and Lenovo , the money is “starting to add up,” as Charlie Sheen once correctly observed https://www.latimes.com/archives/la-xpm-1995-07-21-me-26278-story.html , as they upgrade their systems to make room for GenAI as well as add GenAI systems to their datacenters. Of course, Cisco is getting net new business, as is the case with the several sovereign wealth funds in the Middle East that are building out their own AI capabilities and that have chosen Cisco as their platform provider. You can see the Cisco upgrade cycle at work in fiscal 2023, after the capacity of the coronavirus pandemic buildout was consumed and customers wanted to upgrade their systems. This is also when the Silicon One merchant ASIC and switch and router systems businesses started to take off. After a slump in fiscal 2024, the networking business, which has switching, routing, and serving all glommed together, is up and to the right, and it kissed $10 billion in Q4 F2026. The company’s other businesses are steady freddy, but haven’t grown much in the past year and a half. I like to try to figure out the real systems businesses of the publicly traded OEMs, and as best as I can reckon, the datacenter server, switch, and router businesses including their software components comprised $8.81 billion in sales in Q4 F2026, up 28.3 percent year on year. I estimate that operating income for this aggregate business was on the order of $2.16 billion, up 25 percent. Now, let’s drill down into the AI business that is driving Cisco’s growth in recent quarters. Let’s start with the AI system and component orders that Cisco says have come in from the hyperscalers and the big clouds which it just calls hyperscalers : Cisco said it had $9.3 billion in orders from these tech titans in all of fiscal 2026, and if you do the math based on what it has said in the prior three quarters of the fiscal year, that means it took in $4 billion in orders from the hyperscalers and clouds in Q4, up by a factor of 4.9X year on year. The company said it had over $1 billion in orders from these customers for optics, which means just under $3 billion went to AI systems orders, which means switch ASICs, switches, servers, and maybe some routers plus software for the gear where these customers opt for it. Many have their own Linux operating system for servers as well as their own network operating systems for switches and routers, or they use SONiC or another NOS. I reckon that Cisco did maybe $1.04 billion in optical transceivers in the quarter, up 4X, and $2.96 billion in AI systems components, up 5.4X year on year. Cisco took in another $475 million in orders from neoclouds, sovereigns for AI components and systems, up 4.9X compared to Q4 F2025. Total AI orders were up 4X to $4.48 billion when you do the math on that. The part of the Cisco AI business that is not the hyperscalers and cloud builders is growing faster than that business with these tech titans, but it will probably be many years before the numbers cross – if ever, considering how much merchant networking business Cisco will do with these tech titans. In its presentation to Wall Street going over the Q4 F2026 numbers, Cisco’s top brass said that it had sold a cumulative 850,000 transceivers running at 400 Gb/sec, and another 75,000 running at 800 Gb/sec. The 800 Gb/sec ramp is just getting started, and 1.6 Tb/sec is starting to ramp in the next generation of systems coming out later this year and into early next year. Last quarter, Cisco said it had 750,000 cumulative 400 Gb/sec optics shipments and 40,000 cumulative 800 Gb/sec optics shipments, so it is simple math to figure out what shipped in Q4 F2026 proper. That’s 100,000 transceivers at 400 Gb/sec and 35,000 at 800 Gb/sec. Assuming the 800 Gb/sec transceiver costs at least 2X that of the 400 Gb/sec device, and knowing when these things started to ramp after the Acacia deal for $4.5 billion closed in March 2021, I built a rough model of transceiver sales. Best I can figure, Cisco has peddled a little more than $6 billion in 400 Gb/sec transceivers and $915 million in 800 Gb/sec transceivers. And if MRC networking takes off on AI systems, as I think it will given the resiliency it yields and as I detailed back in May https://www.nextplatform.com/connect/2026/05/12/openai-microsoft-and-friends-build-a-better-more-scalable-ethernet/5239078 , companies will be spending a hell of a lot of money on transceivers and cables to make their AI clusters more fault tolerant – and coming out better in the bargain because there is nothing more expensive than a cluster of GPUs or XPUs just sitting there burning idle power. Of course, all those transceivers and cables are not going to be cheap, either. For the full year, Cisco had $63.33 billion in sales, up 11.8 percent, with a net income of $13.27 billion, up 26.9 percent. AI hyperscale and cloud revenue was $4 billion of that, up by a factor of 4X compared to fiscal 2025. It is hard to say how accretive this AI business is to earnings. Cisco cannot make much margin reselling Nvidia components and designs, but it can do fine selling its own chippery, systems, and software. Looking ahead to fiscal 2027, Cisco expects for sales to be between $72.2 billion and $73.4 billion, and that is up 15 percent at the midpoint of $72.8 billion. It is hard to say how much profits Cisco can extract from those sales, but if it is just a lot of shuffling Nvidia gear and not selling a lot of its own scale out and scale up networking, profits could come under pressure and not grow as much as revenues. This is clearly a situation that Chuck Robbins & Co want to avoid. The fiscal 2027 year is still young. . . .