Hi, everyone, welcome back to Network Stats, where we track quarter on quarter a bunch of different metrics and heat maps and all sorts of fun stuff about the network traffic happening at Backblaze. I'll take a minute to introduce ourselves here. I'm Stephanie Doyle. I'm the Technical Narrative Content Manager and lovingly called the Keeper of Stats since I am on many of these webinars. I'll let Brent go ahead and talk about himself.
Hi. Hello. I'm Brent Nowak. I'm the manager of network engineering here at Backblaze. Our group's responsible for the connectivity inside the data center, so all the copper, the fiber, and then also the connectivity external, so all of our internet connections, our IX connections, and also PNIs to our various partners.
Definitely a mighty team there. You guys do a lot of great work. Just to talk about the agenda today, we're gonna review this quarter's highlights and insights, talk about some of the quarterly data that we're tracking, take a few questions if we have the time. We've got some resources. I'll also be putting attachments on the webinar page itself after this. Just a couple little housekeeping things. This is being recorded. Please do ask us questions. We'll try and take some time at the end, but you can also drop them in the chat. We are monitoring it. Of course don't miss the attachments. All right. Highlights this time around.
Brent, I think we saw some pretty interesting things, and we actually transformed the data in some new and interesting ways too.
Yeah, our data set historically for the last quarter showed us who we're talking to, which is the networks, and then what and when. We were looking at the TCP conversations, the length of the TCP conversations, and the size, how many bits were transferred. What we added this quarter was the where. We added some geo information that allows us to know where certain types of traffic are going, and it led to some really interesting results that we're gonna go over.
Yep. Agreed. Let's jump in. This is sort of our full picture here. Take it away, Brent. Tell us what we're looking at.
What we're looking at is a total amount of traffic that we have sent and received through the Backblaze network per month. What we use here for our metric is called bits 95th, so this is the 95th percentile of the traffic. This is a common metric that we use in the ISP world to measure traffic. It is a better approximation than using average or mean because the 95th is the value of 95th that lies where 95% of the numbers fall below and 5% fall above. It's a really good way to reduce your outliers and get a better picture of what we look at from a traffic perspective. This is a month-over-month view, so we've got history going back from May of 2025. We saw a large amount of traffic that we were sending into the winter season.
We saw a little decrease December, January, and an uptick again in February and March for total traffic. Each colored slice here represents a different type of network classification. We have our CDN, our hosting. Hyperscalers and Neoclouds are very interesting for us to monitor. Then also our ISP connections such as Tier 1, which are global reach partners, and then regional, and ISP regionals in this darker purple color.
I think what we saw defined here and elsewhere in the report, but we might as well introduce this term now, Brent, is really the prevalence of elephant flows as we're talking about our network these days. If you wanna describe for folks who are unfamiliar the term.
Sure. In the networking world when we talk about traffic, there's 2 kind of colloquial ways to classify traffic. We call them mice flows, and we call them elephant flows, and they're very descriptive towards the traffic. A mice flow is sort of a small amount of information, maybe a kilobyte, a megabyte, 10 megabytes of a file being transferred here or there. There's many different participants. There's maybe 1,000 of these. For example, when you go to a website, you load many different small assets from many different locations. What's moving and driving a lot of innovation at Backblaze is what we are defining as elephant flows, and these are large, single transfers between 2 parties. These could be at line rates of a gigabit, 10 gigabit, or even higher.
When we see partners try to scale their workflows, they're sending multiple of these elephant flows, which can add up to 100 gigabit, 400 gigabit in total aggregate of connectivity between two partners.
I think it's interesting because we see it showing up in different ways through many of the graphs, so I'll let them speak for themselves, but you'll hear us using words like bursty or, you know, how we tracking for these things. You can see that you've got spikes, and then you also still have a higher baseline as well. There's an interesting traffic move that's happening here. These are our Sankey charts and these track, if I'm not mistaken, traffic to different workloads. Is that correct, Brent?
This is a way for us to visualize how we are transferring our data over various connectivity types. For us on the network engineering team, it's cheaper for us to transmit traffic over PNIs. These are often zero settlement costs for us. If we have partners that are geographically located close to us, we like to initiate conversations to see if we can have a fiber run to them, which allows us, again, to have very cheap transit. We also have connectivity over different transit networks such as our ISPs, and we also have what we classify to our cloud partners, where we deliver CDN content.
This is a sample for Q1, and it shows the total aggregate traffic the same as that previous graph but just in a different way to view kind of how it's transferred via the different transport methods.
Just for reference, we went ahead and pulled up Q4 2025 as well, just because we did see a shift in these slices. One thing you'll find us harping on throughout the reporting is just that it's really hard to define patterns when the data set's young. We try to keep things in a description space. To that end, you can see the difference here between Q1 2026 and Q4 2025. Let's talk magnitude here. I believe that's where we're at.
This is the amount, yeah, this is the amount of total traffic that we've sent over our different types of networks based on the different regions that we have. When we start to add this regionality information, us being Backblaze employees, we kind of have a little more history. We understand why this layout looks the way it is. Just a little insight here. Our US West infrastructure was our first deployment, that has the most amount of historical content. That shows up here, where we see a heat map in the US West CDN, very colored, very deeply red. Additionally, we also see that in the ISP regional traffic. This makes sense for us from a business perspective because US West was our original deployment for the Backblaze network.
As we expanded, we also added EU Central and EU West, and we're starting to see a more heat map concentration where we have NeoCloud activity in those different regions, ISP regional traffic, hyperscaler, and also CDN traffic. Interesting here, CDN traffic is pretty well spread for us across our US East, US Central locations, or EU Central locations. This is a very interesting graph that lets us know how the amount of total information is transferred over our network by region.
Now we're on magnitude. That's this one here.
Magnitude for us is a metric that we coined. This is a measurement of the amount of bits transferred per IP address. Rather than looking at the total amount of traffic, we're adding a two-dimensional metric here, where it's the amount of information transferred per speaker. That, to us, has been very insightful for us. You can see here that it's very different, the amount of total traffic, that orange graph previously. We see a very deep green color concentration in U.S. East for NeoCloud. This is a location for us where we see a lot of NeoCloud activity in our U.S. East cluster. This is because of the geographic location of a lot of hyperscalers, GPU providers, and it just makes sense, and it's really good to have data that sort of validates what we're seeing on the business side.
Totally agree. Let's talk about unique addresses and how that differs a little bit.
Much like the total traffic graph, where I spoke to U.S. West being our most original and the first implementation of the Backblaze Network, this graph also is very boring, but it helps validate our assumptions here about how we run our business. A lot of the content in our U.S. West clusters is sent over ISP regional traffic. This is traffic to Comcast, Verizon, GFiber, and it shows up here where we talk to many, many, many different unique IP addresses out of U.S. West, mainly on ISP regional networks. For us, this drives decisions on where we want to put connectivity. It means that in our U.S. East locations, we may want to augment with higher gigabit ports, whereas in U.S. West, we want to partner with more internet exchanges to get more local to consumers.
Makes total sense. That brings us to a really fun question. Where in the world is the neocloud? I found your explorations here, Brent, to be very cool this time around.
We did add geo data to our data set for this series, and when I started to look into the data set and produce some heat maps of where we are sending traffic by country, you can see here the U.S. for us is very deeply shaded. That's deeply shaded for our neocloud activity, for hyperscaler activity, and also CDN content.
Well, we call-
Oh, go ahead.
Yeah, we call out in the report that that's somewhat unsurprising because, what is it, 40%-60% of data centers are located in the U.S. as of right now, I believe is the common metric. You might see this on any network provider, really.
Right.
Yeah. Now we get a little more granular.
The next question I asked once I saw the first heat map was, if we exclude the U.S. data, what does the heat map look like? 'Cause this will give us more differentiated results on a per-country basis, and it definitely shows in this heat map. What we see is concentrations in hosting CDN activity and NeoCloud activity for countries like Germany, the Netherlands, Singapore, Finland, and the U.K. Again, this also helps us inform our business as we want to grow and expand, where we place things, how we want to connect to people. This was a really nice visualization of the data, excluding those U.S. numbers that were sort of skewing us and not showing a lot of differentiated results in this heat map.
Yep, absolutely. The alternative is this guy.
The next question was, if we just look at the U.S., what does that look like from a footprint standpoint? We see a lot of neocloud activity towards California addresses, and this makes sense for us based on our partners, our connectivity. We also have a lot of CDN activity that ends in California. Also, we do see hyperscaler activity in Virginia, which makes sense for us because there are a large amount of hyperscalers there. New for us, which is interesting, is Illinois, Georgia, New Jersey are also showing up with areas of concentration. We haven't dug too deep into that yet, but there's always room for improvement in our data set.
I think it's important to note here, too, just to sort of clarify what we're looking at, we're not saying that, like, we're saying the data moves back and forth from there. What can that look like, Brent? Is that coming from, like, a regional exchange, or is that physically coming from the endpoint? What are, what are we talking about when we say where it's going to and from?
We see a lot of activity over internet connections.
Mm-hmm.
For us, that means that where we were deploying 100 gigabit links, we're now increasing that to 400 gigabit or multiples of 400 gigabit. One of the offerings that we have is a product called B2 Overdrive, where we allow you to have S3-compatible object storage that can scale from 100 gigabit up to 1 terabit, that's been a driver for us as we've been choosing port capacities, where we deploy links, and expanding the network.
Yeah. Very cool. Let's get into our next slicing, talking here about just comparing different types of traffic coming through the network.
Back to our magnitude metric, again, this is a measure of how many bits were transferred per unique IP address, this is a diagram of that metric over time for our Neocloud and our hyperscaler operators. What we see here, what we saw on that graph, we had a lot of activity into August, September, October, November, a sort of lull in traffic January, February, and a resurgence again in March, is that the magnitude, the bits of, the amount of data transferred for IP address, still is coming in very high. This graph is, you know, orders of magnitude, so those two green dots on the upper right for Neocloud are very impactful to our network infrastructure. These are flows that are not 1 gigabit. These are 100 gigabit to 400 gigabit flows.
As network operators, it's really interesting because it means that our network on a Tuesday is performing very differently on a Thursday. These flows that we see from hyperscalers and neocloud partners happen at any time. The magnitudes are very great. They don't follow our typical pattern of people are working during the day, there's a lot of content generation, a lot of backup happening. Overnight, people sleep. Our network's kind of sleepy, just like people sleep. These are workflows that are populating GPU infrastructures with as much data as quickly as possible, because time is money when you are renting time on a GPU cluster, and we see that show up here in the magnitude of the workflows.
Yep. Absolutely. Let's talk, how dynamic are these traffic patterns?
The next series of graphs dive into each traffic type over time per region, this is a new set of data that we're publishing to kind of take a look at how our different regionality changes over time. We want to start tracking if the U.S. East is always a area of concentration or if we see shifts into U.S. West, again, this speaks to how we improve our network from a network operation standpoint. What we see is, again, in October, there was a lot of activity with Neo Cloud activity in U.S. East, a smaller period in January, February, a resurgence again in March. This is telling us that U.S. East is still a concentrated spot where we're sending a lot of traffic to our Neo Cloud operators.
Yeah. Very cool. We've sliced for our other use cases as well, so we've got hyperscaler here.
Yeah. Hyperscaler follows the Neo Cloud activity. We see a lot of pairing of the data between our Neo Cloud and hyperscaler data set. This is because the data that's stored with B2 can be sent to Neo Cloud number 1 or Neo Cloud number 2, hyperscaler number 1, hyperscaler number 2, depending on what serves the client at any given time. There may be periods where the rental time on a GPU cluster is advantageous on partner A versus partner B. That freedom of choice means that we have to sort of plan for our data to be sent to any of these partners at any time. We generally see a very tight coupling between the Neo Cloud and hyperscaler activity.
Yep. CDNs, I think these next three you also sort of grouped, and I find them to be very interesting as well, in a different way than neoclouds and hyperscalers.
When we speak about the neocloud hyperscaler activity, we are talking about bursty traffic that can happen any time. It could impact our network at any moment. When we look at the next three graphs for our content delivery, hosting, and ISPs, these are more steady state for us. These, again, follow traffic patterns that are steady. People are working during the day. There's a lot of content being generated, backed up. At night, our, we see our bandwidth graphs get a little lower. When we look at these three, CDN, hosting, and regional ISP, these are very much a different mode of operation, where it's just steady state network. We can plan for this growth. It's very easy to track. It doesn't keep us up at night.
Or does?
challenged in a different way.
This is CDNs here. I found the difference between hosting and the ISP regional to be the most interesting, it, from a comparison point of view.
Yeah. If you want to advance.
Mm-hmm
... the hosting.
This is ISP regional, and this is hosting here.
Oh, I don't see a slide advancement.
We may be having some network issues. I've got ISP regional up on there, Brent.
Okay.
Yeah.
The ISP regional for us, again, is pure steady state. This is easy to plan for. This is where we have the most ease of operation. As you can see here, U.S. West being the original source of our or the original source of the Backblaze network, has the most content that's being delivered to ISP regionals. It may not be the most exciting graph, but for us, this speaks to very steady state planning.
I think the biggest takeaway I had from those 5 different slides was really just how different you get by use case slice, and it really reflected back to me why we slice and dice the data in the different ways that we do. The challenges of the workflows that we're balancing when you're talking about planning for network engineering growth and all that good stuff. I do see. Well, before we go there, we're onto questions. I love this. I do see there's 1 question in the chat from James Kurtz. Says, "I noticed a lot of demand from neoclouds in Finland. How is latency and throughput from Amsterdam to GPU clusters in Finland?
How are we using these trends to plan your future expansions?" I would say, number one, this is actually a question for a different series of ours called Performance Stats, in which we actually track latency and throughput. I think in general, what I would say about using trends to plan future expansion plans is that all data is good data, right? All these things are trade-offs in what we invest in and how over time. It's an interesting question because a lot of that, a lot of what happens for us is we bring data and other people make business decisions that trade off a lot of different things, including what we see here.
Brent, if you want to speak to some of the things that we have talked about and ways we've expanded our network to handle some of those elephant flows, I think that'd be interesting.
I think the key takeaway is that the neocloud landscape is very dynamic. There are facilities being built that are coming online. There are facilities that are being built. It's a very dynamic space, and we're open to pursuing any connectivity, if it sort of suits, you know, the growth pattern. We are in a state where what we assumed last month doesn't really make sense the next month. You know, this discovery of GPU clusters in Finland could be a new discovery for us. It is just a very dynamic landscape, and it's driving a lot of innovation for us, which is great.
Yeah. What we do with it, I think, is always where sort of the magic happens. We don't necessarily have a on-demand answer for that. It's, I think, visibility, more visibility is always better.
Agree.
Yeah. Any other questions coming from the chat? I don't see a lot of that in here. I do, I wanted to surface, I think, one thing that we got when we were doing the report in other channels, Brent, was a question about how, you know, how you can really see how different these look, these elephant flows look from a networking perspective. Like, how persistent are they, all that good stuff, and what they say about your customers. And I remember you talking a little bit about how we sort of, you know, obviously we never see our customers' data, so we don't know exactly what they're doing with it, right?
We talked a little bit about definitions and tracking these things, and I think that might be interesting to talk about here.
Yeah, we take a sampling of network traffic, and we use the protocol called sFlow, which is where we sample every X number of packets. Every 64,000th packet, we take a sample of, and aggregating those together, we can get a picture of what a TCP conversation is between two partners over time, and that lets us know what the IP addresses are involved, the length of the transfer, how many bits were transferred, what protocol is it, IPv4, IPv6. Again, it doesn't tell us the content. We can't see the content of the stream that's encrypted. It does let us know the networks that we're talking to, whether it's a Comcast network, Verizon, whether it's a neocloud operator or a hyperscaler.
We then enrich that data with some additional fields in a database where we classify different networks as different types. That's where our data set comes from.
Yeah. James, I think that kind of leads into your second question here. Can that large flow in October be a sign of movement of data off the Backblaze network and onto other storage networks? What are you seeing in terms of migration of data lakes into homegrown neocloud storage? Our visibility is really activity, right? We don't know whether it's coming on, going off, coming on and going off, which often happens when you think about training data sets and how they move. There's kind of no way from a networking perspective to talk about that.
Except for that I would say if you look at our total baseline, that very first chart, we can see that, you know, we do have our peaks and valleys, but our total baseline is up in, from an activity perspective, which again, doesn't speak to total volume of data. As far as, migration of data lakes into homegrown neocloud storage, Brent, is there any way that we would have visibility into something like that? Like, that says no.
I don't think we can get that granular into the specific application or service. We can merely tell what network that we're talking to and in what Magnitude. I don't think we can answer that question today.
James, I'd be interested to know what you're defining as homegrown neo cloud storage, because we're certainly seeing, you know, just from an industry perspective, a lot of different experimentations to serve a storage demand, right? I think people focused quite quickly on the GPU demand, but we know that there's a massive storage demand there too. If you want to reach out to us in the comment section or anywhere really and let us know, sort of where your head's at, I'd love to hear what trends you're tracking, because I'm liking your questions here. Contact information, drivestats@backblaze.com is the email, or you can jump into the comment section of the report. We're pretty active on paying attention. Of course, socials. Wherever you want to find us.
Thank you, James, for your questions. We always appreciate it.
Thank you.
Yeah. With that, if there's no more questions, I think we can say thank you very much, Brent. I always appreciate all the wonderful analysis you bring to bear. We'll see everyone on the next Network Stats.
Of course. Thank you very much.
Bye.