SC25 Recap – AI and More

Getting your Trinity Audio player ready...

AI has taken over HPC

Despite this being a bit upsetting for some HPC industry veterans, SC25, more than any SC event before it, was all about AI, and mainly AI infrastructure. While AI can be considered a subset of HPC, it has grown so big that there was little room for giving attention to anything other than AI. This is not a big surprise, since AI has taken over, practically, every other aspect of current technology; yet still, it was very evident this year in St. Louis.

Networking has taken over AI

While the previous conclusion is not very surprising, this one is a bit more interesting. This year, more and more industry leaders came to realize that networking is a crucial part in the success of AI. This was discussed intensively during the show, as people understand that while networking is a relatively small part of the investment in AI infrastructure, it is, by far, the largest “headache” of such deployments and needs special attention and consideration.

By mere coincidence, during the event Nvidia posted its financial results for Q3 of its 2026 fiscal year. What amazed me about those results is that Nvidia’s networking revenues topped $8B this quarter. This unprecedented record may seem like an error, as Nvidia is not, by all means, a networking company. But, when you think about it, it makes perfect sense as networking is a growing part of any AI deployment, and Nvidia, today, is THE AI company. This is why Nvidia, for the first time ever, became the largest networking company by revenues, dethroning the previous leader Cisco, which generated “only” $7.7B from networking this quarter.

Download now!

Fabric-scheduled Ethernet as an effective backend interconnect for large AI compute clusters

Ethernet has taken over AI networking

The HPC market in general, and the AI infrastructure market in particular, have been heavily biased towards non-standard, semi-open networking protocols and architectures, with InfiniBand being the dominant one. This is, to some extent, still the default go-to technology for some designers in this industry. The main reason for that, other than tradition, is InfiniBand’s performance superiority, especially when compared to the default technology for data centers: Ethernet.

But, this year, it seems the industry has realized that Ethernet is no longer inferior to technologies like InfiniBand. When you look at performance figures like NCCL bus bandwidth or even job completion time (JCT), there are Ethernet technologies that have become on par and even superior to InfiniBand. People on the show floor understood that, and realized they can move to an open, cost-effective, fast-to-deploy technology. This is something they have wanted to do for ages, but now can do so without compromising on performance.

Scheduling has taken over Ethernet

And the change in Ethernet that made all of this possible? Scheduling..

As written here many times before, there are two methods of adding a scheduling mechanism to Ethernet in order to reduce congestion, packet-loss and jitter, and improve overall performance. One is based on the endpoints (i.e., NICs), and the other is based on the fabric itself. Both are valid, and both were discussed last week in St. Louis – drawing a lot of attention and interest.

What’s next for AI infrastructure?

So, what will we talk about a year from now at SC26 in Chicago?

Chances are we will see all of the above changes and trends materialize in the year leading up to that event. Lots of AI, lots of networking, lots of Ethernet, and multiple scheduling architectures.

Will there be other new topics and trends by then? Most likely – time will tell.

Key Takeaways

AI has taken over the HPC world

SC25 demonstrated that AI now dominates discussions, priorities, and innovation within the HPC community, leaving little room for traditional HPC topics.

Networking has become the most critical challenge and point of focus in AI deployments

SC25 highlighted a growing industry realization: while networking represents a small portion of AI infrastructure spend, it creates the largest “headache,” drawing intense attention as deployments scale.

Schedule in driving Ethernet ahead

Once seen as inferior for high-performance workloads, Ethernet is now outperforming InfiniBand in key metrics due to new scheduling approaches—making open, cost-efficient AI fabrics a reality..