{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreidutiswpumkqk4rbc7bx4c4bmniibveoesctbeka45g7q336dos4m",
    "uri": "at://did:plc:qzjwstutqk2cy7df7jbzd2hx/app.bsky.feed.post/3mkmir2nek4r2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreif6l477dh256eps36jgcs6wq5qdfkybcqcojvcwf3egraec6pwox4"
    },
    "mimeType": "image/jpeg",
    "size": 10483233
  },
  "path": "/article/4164337/ai-data-flows-force-rethink-of-data-center-networking-at-backblaze.html",
  "publishedAt": "2026-04-28T14:11:13.000Z",
  "site": "https://www.networkworld.com",
  "tags": [
    "Artificial Intelligence, Data Center, Networking",
    "Backblaze",
    "neocloud providers",
    "Brent Nowak",
    "Backblaze released this morning",
    "neoclouds for the GPUs",
    "released by Synergy Research Group",
    "released late last year by McKinsey",
    "Stephanie Doyle",
    "Silicon Data report",
    "LeanOps",
    "Ravi Kanani",
    "Omdia research report"
  ],
  "textContent": "Cloud storage provider Backblaze has been ripping out its 100-gigabit links and replacing them with 400-gigabit links because AI has changed how traffic flows inside their data center. AI workloads running on neocloud providers like CoreWeave and Lambda are creating bursty, unpredictable flows that overwhelm traditional capacity planning models designed for traditional, steady-state cloud traffic.\n\n“When we are seeing these large data flows, when they transfer through our network, there may be saturation points,” says Brent Nowak, manager of network engineering at Backblaze. “We have APIs, load balancers, database servers, storage arrays.”\n\nAnd the traffic load is compounded by the way the data moves in the network, meaning that 400 gigabits of new traffic can turn into terabits per second inside the data center because the files are distributed and different parts of the infrastructure need to talk to each other.\n\n“And what we’ve seen has been very unpredictable,” Nowak adds. “We’ve just seen that what is happening on our network on a Tuesday can be very different from a Thursday in terms of capacity and workload.”\n\nBackblaze has optimized its stack to handle the high-intensity burst patterns. Today, the company is deploying multiples of 400 gigabits per second in the data center, Nowak says. “We also increased the density of our Arista switches,” he adds. “Where we had switches that were 32 ports of 100 gig, now we’re deploying 400 gig ports.” That allows Backblaze to have higher density, higher throughput, and connect more devices, he says.\n\nAnd Backblaze is also looking ahead to the future. “I’m currently doing a project to model what our network would look like if the amount of traffic was doubled or quadrupled,” Nowak says. “We have this bursting traffic that can flood and impact our network in ways that we need to be on our toes for. And, from an engineering perspective, that’s kind of exciting. So, I’m actually enjoying this.”\n\nAccording to a report that Backblaze released this morning, traffic from content delivery networks and hosting and Internet services providers have stayed largely within historical norms over the past year. But traffic from hyperscalers and neoclouds fluctuated dramatically, with steep climbs in September and October and another uptick in March.\n\nAnother network traffic change related to AI is geography. “Traditionally, it didn’t matter where cloud infrastructure was located,” says Nowak. But with AI workloads, if storage is close to compute, enterprises get lower latency and higher throughput.\n\nToday, Virginia and California have a high concentration of AI compute providers. This, in turn, brings in more storage companies. “In July, we chose to double our footprint in US East to increase the proximity to hyperscalers and neoclouds,” says Nowak.\n\nAnd that, in turn, leads to even more demand for compute, and even greater concentration. “There’s a snowball effect,” Nowak says.\n\n## Why neoclouds for AI?\n\nEnterprises might think that they don’t need to worry about network traffic details if they’re using a hyperscaler for their AI workloads because the data and the processing both stay within the cloud. But there are advantages to using a third-party storage provider combined with neoclouds for the GPUs.\n\nAccording to a report released by Synergy Research Group in early April, neocloud revenues hit $9 billion in the fourth quarter of 2025, a 223% year-over-year increase. Revenues passed $25 billion for the whole year and are expected to hit $400 billion by 2031.\n\nNeoclouds offer flexible contracts, faster provisioning, and specialized infrastructure configurations, according to a report released late last year by McKinsey — at GPU prices up to 85% lower than hyperscalers.\n\n“You see the hyperscalers having to go to neoclouds too,” says Backblaze’s report lead Stephanie Doyle.\n\nAccording to the March Silicon Data report, H100 prices from hyperscalers were roughly three times higher than from neocloud providers. Plus, even though hyperscalers don’t charge for internal data movement, there are cross-region data transfer costs as well as data egress costs. And data storage costs are also a premium with hyperscalers.\n\nAccording to an April report from LeanOps, AWS S3 is the most expensive mainstream object storage provider, costing nearly four times more than Backblaze and more than three times as much as Wasabi, another popular storage vendor. “Add the $0.09 per GB egress fee and S3 can cost 10 to 17 times more than alternatives on total monthly spend,” wrote LeanOps founder Ravi Kanani in the report.\n\nCompanies still prefer AWS S3, Kanani adds, because S3 integrates with more than 200 AWS services, operates in more than 30 regions, offers automated lifecycle management, and backs everything up with eleven nines of durability.\n\nPlus, neoclouds are new to the game — thus the “neo” — and have some limitations.\n\nAnd storage providers aren’t the only ones struggling with AI-related network infrastructure issues. According to an Omdia research report released in mid-April, neoclouds have scaled compute for AI workloads, but networking lags behind.\n\nBased on the study, which covers 50 neoclouds, more than one-third of providers minimize contractual liability, meaning that customers need to pay closer attention to service commitments, security, and data sovereignty. More than half don’t use Internet peering exchanges, even though they help secure consistent performance. And while IP address ownership helps support growth and routing control, 46% of neoclouds control only small IPv4 address blocks — and one-fifth rely on only one IP transit provider, creating a single point of failure.",
  "title": "AI data bursts force rethink of data center networking at Backblaze"
}