The Memory Crunch Pinches Cisco’s Profits

It has taken many years for the AI boom to reach the general ledgers and balance sheets of the world’s largest original equipment manufacturers, and one might say that it has taken particularly long for Cisco Systems, the dominant supplier of switching and routing in the enterprise and traditional telco/service provider spaces as well as a respectable systems supplier with over 90,000 customers using its UCS converged server-switch platforms.

But Cisco has been playing a very long game under chief executive officer Chuck Robbins, starting with rearchitecting and unifying its switch and router ASICs with the Silicon One chips, which it uses in its own gear and which it also makes available as merchant silicon that hyperscalers and cloud builders can acquire and use in their own whitebox gear. The acquisition of Acacia Communications back in 2021 for $4.5 billion also gives Cisco the optical transceivers that appeal to the hyperscalers and cloud builders that are driving the lion’s share of AI spending. And now, Cisco is getting traction as a provider of servers and switching to neoclouds and sovereigns need as they build out their infrastructure for the GenAI boom.

This AI business is still relatively small, but it is growing:

As fiscal 2025 was rolling along and AI orders to hyperscalers and cloud builders were starting to ramp, the company originally expected only about $1 billion in AI orders for fiscal 2025. But when the fiscal year ended last July, it was more than double that – our model says around $2.1 billion. Well, that is as much as Cisco booked in AI orders to these elite customers, and Robbins confirmed that number in a conference call with Wall Street analysts, which let us know that our model was spot on. So hooray for that!

So far this fiscal year, Cisco has $1.3 billion in AI orders with hyperscalers and cloud builders in Q1 F2026 and $2.1 billion in Q2 F2026; the company also confirmed that it had $350 million in additional AI orders to neoclouds, sovereigns, and enterprises.

This is the first time that Robbins & Co broke out the latter category, and we did a little backwards extrapolating that pretty much shows why only now is sales to this latter group becoming material. We think most of that is for a combination of servers and switching sold to these customers, with a fair amount of optical transceivers in the mix. But at $3.35 million for each rack of GB200 NVL72 iron (72 GPU sockets per rack) and maybe a little bit less for two racks of less dense DGX NVL8 machinery with a total of 64 GPUs, we are only talking about maybe 100 racks or 224 racks of iron, with either 7,200 or 7,168 GPUs, respectively.

A couple of hundred racks is one HPC supercomputer or one risk quantification cluster at a big bank or, heck, one large Hadoop cluster from a decade ago. This is material for Cisco, which should be commended for being specific. But this had better be the beginning of a very big ramp outside of the hyperscalers and the cloud builders.

Now, of that $2.1 billion booked so far with the hyperscalers and cloud builders, none of that is for systems, which stands to reason since these companies design their own iron and have it custom built by original design manufacturers (ODMs), thereby cutting out the OEMs and their profit margins from the picture. But these hyperscalers and customers do want to buy Cisco Silicon One chips and sometimes switches or routers – the Cisco 8000 whitebox series of switches and routers based on Silicon One are explicitly designed to run the open source SONiC network operating system or custom ones made by the hyperscalers and cloud builders.

In Q2 F2026, Robbins said that of that $2.1 billion in orders from the tech titans, about 60 percent was for switch and router silicon or systems and 40 percent was for optics. Some quarters it is half and half, sometimes it is two thirds to one third, and if you do the math for the trailing twelve months, then there was $2.89 billion in Silicon One stuff sold and $1.93 billion of optics stuff (a lot of it from the Acacia line) sold.

Cisco is telling Wall Street that it will do more than $5 billion in sales of AI stuff to the hyperscalers and cloud builders, to which we say: “Yup, Cisco is low-balling it to be careful again.” Just as it did at the beginning of fiscal 2025. Cisco does not include sales of the just-announced G300 and P200 Silicon One chips and switches and routers based on them in this forecast.

Our most pessimistic forecast for Cisco, shows AI orders coming in at around $6.2 billion for Silicon One chips and switches and optics sold to the tech titans in fiscal 2026, which is a factor of 3X growth year on year. Cisco could do another $1 billion in AI orders from neoclouds, sovereigns, and enterprises (anchored by Saudi Arabia’s Humain effort).

With that, let’s drill down further into Cisco’s financials for the second quarter of fiscal 2026.

In the quarter, Cisco brought in $15.35 billion in sales, up 9.7 percent year on year and up 3.1 sequentially. Operating income was $3.78 billion, up 21.5 percent thanks to ongoing cost cutting efforts, and net income was $3.18 billion thanks to an incremental $25 million in other net income compared to a $60 million in extra charges this time last year.

Cisco ended the quarter with $15.78 billion in cash and equivalents in the bank, and had a revenue backlog of $43.4 billion, up 5.1 percent. Cisco’s revenue backlog grows slowly and every once in a while shrinks a bit.

Sales of services, as you can see, just kind of hums along at Cisco, and products, being a capital expense instead of a recurring operating expense, are choppy as they have been for the past six decades of enterprise computing. Product sales were $11.64 billion, up 13.8 percent, while services revenues were $3.71 billion, down 1.3 percent.

The group that Cisco calls Networking – which includes datacenter switching and routing, campus switching, wireless stuff, servers, and IoT gear – was the star of the quarter, with revenues up 21.1 percent to $8.29 billion. All of the categories mentioned above in the networking group all saw double digit growth year on year.

We wish that Cisco separated out edge and IoT stuff from datacenter stuff, but it no longer does. But we have some past trend data that dates from the Great Recession when Cisco entered the datacenter systems market, and since that time we have been trying to figure out what Cisco’s “real” datacenter business looks like. Here is the trendline for that:

As best we can figure, Cisco’s core datacenter business was around $7.47 billion in Q2 F2026, with an operating income of $1.82 billion, representing about 24.4 percent of revenues.

Those operating margins took a bit of a hit thanks to DRAM and flash memory shortages, which are driving up prices. Robbins said that Cisco is changing contracts to have provisions for flex pricing on memory so it can pass through higher costs – memory prices have doubled in a year and probably will triple before the year is out is what we are hearing – to customers. Cisco has been buying up DRAM and flash memory as much as it can, which is why its advanced purchase commitments had another $1.8 billion added in the quarter. That is up 73 percent year on year, and “a big chunk of that,” according to Mark Patterson, Cisco’s chief financial officer.

“As we talk to customers, I think they understand it,” Patterson said on the call about the memory price increases. “They understand this is industry-wide. And I haven’t talked to any customers that are really willing to delay or defer any sort of strategic investments that they are making in technology. And I think there’s no concern that we’ve seen there yet whatsoever.”

Just another way you know the GenAI boom is exceptional. Companies will just defer acquisitions of general purpose machines to buy AI iron, we think. It has happened before.