AI and CDN demand lift cloud network traffic baseline
Cloud network traffic reached a record on Backblaze’s infrastructure in July 2026, driven mainly by regional internet providers, AI-focused compute networks and content delivery activity. Traffic eased in August and September, but several categories remained above earlier baselines, suggesting that capacity planning must account for both sustained growth and sudden bursts.
Backblaze has tracked these patterns since December 2023, although Q3 2026 was only the third quarter covered by its operational dataset for direct quarter-over-quarter comparisons. The company added percentage changes to its regional heatmaps so readers can distinguish heavily used connections from connections whose traffic is growing quickly. A dark cell with no quarterly change indicates a stable, high-volume path, while a large percentage increase can point to a new customer workflow or a wider change in demand.
The traffic graph divides networks into several categories. A CDN, or content delivery network, retrieves material from an origin store and distributes it closer to users. Hosting covers providers running physical or virtual servers for websites, databases and applications. A hyperscaler is a large traditional cloud provider, while a neocloud is a compute network focused on AI workloads. Regional internet service providers, or ISPs, operate close to customer equipment, while Tier 1 ISPs transport traffic over longer national or international distances.
CDN traffic hit a significant peak in July before moderating, but its baseline remained substantially above the same period in 2025. Hosting stayed close to its steady pattern from the previous year. Hyperscaler traffic also peaked in July, then declined slightly while retaining a larger footprint than in Q1 and Q2. Neocloud activity expanded sharply to a record level in July and remained one of the largest continuing sources of traffic after easing in August and September. Regional ISP traffic maintained a heavy band throughout the quarter. One-way data migrations into Backblaze, frequently conducted through Universal Data Migration, produced distinct spikes in all three Q3 months, resembling bursts seen in late 2025 and March 2026.
The regional analysis covers US-West, the oldest and largest region; US-East, which has the greatest observed proximity to neocloud infrastructure; CA-East, the newest Canadian region; and EU-Central. US-West retained one of the darkest cells for regional ISP and CDN traffic, reflecting its large server footprint and Internet Exchange presence. An Internet Exchange, or IX, is a facility where networks connect directly to exchange traffic.
US-East recorded the heatmap’s highest overall CDN volume and an 85% quarter-over-quarter increase. Neocloud volume was concentrated in US-East and EU-Central, while accelerating by 58% in US-West and 46% in EU-Central. EU-Central’s Tier 1 ISP traffic rose 319%, and hosting traffic in US-East increased 133%. CA-East showed increases of 44% for hyperscaler traffic and 35% for neocloud traffic, but Backblaze cautioned that its low starting volumes mean small absolute movements can generate large percentages. The 85% CDN increase in US-East and 46% neocloud increase in EU-Central occurred on already substantial volumes, making them more consequential in absolute bandwidth terms.
Volume alone does not reveal whether growth comes from many users or a few unusually large transfers. Backblaze therefore measures magnitude, defined as bits transferred per unique IP address. High volume spread over many addresses is generally easier to distribute across infrastructure and providers. High volume involving only a few addresses creates heavier individual flows and fewer options for redirecting demand across wide-area network paths.
Neocloud magnitude increased 37% in EU-Central while falling by an almost identical proportion in US-West. US-East recorded increases of 602% for CDN magnitude and 279% for hosting magnitude. EU-Central’s 319% rise in Tier 1 ISP volume also appeared in the magnitude data, indicating high-volume workflows involving relatively few endpoints. The measurements do not identify which workloads produced that change.
The number of distinct addresses provides a third view. US-West had the highest overall uniqueness for the third consecutive quarter, consistent with its larger data-centre footprint and mixture of workloads. Neocloud traffic generally involved fewer, persistent endpoints, matching AI pipelines that maintain stable connections between storage and compute. An AI pipeline is an automated flow that moves data through processing stages. CDN networks in CA-East were the exception, with the number of addresses rising 335% quarter over quarter.
Time-series data showed how AI-related activity differs from steadier CDN, hosting and regional ISP patterns. January and February 2026 had concentrated, low-magnitude neocloud and hyperscaler baselines, followed by a sharp March surge in several neocloud networks. Neocloud peaks compressed slightly during Q2, but many more high-magnitude workflows stayed active. Hyperscaler endpoints moved upward through May and June. In Q3, neocloud activity remained broadly elevated, with a July breakout, some moderation in August and a renewed spread of activity in September. Hyperscaler magnitude remained elevated in July before moving closer to baseline during August and September.
Regional monthly heatmaps covering May 2025 through September 2026 showed neocloud activity expanding beyond US-East into US-West and EU-Central. Hyperscaler activity also remained distributed across the three regions, with a modest shift toward US-East in Q3. Separate CDN data from May 2025 and hosting and regional ISP data from March 2025 showed unusually heavy CDN traffic in US-East and increased EU-Central hosting activity from April through September 2026. Backblaze stressed that these are observations from its own network, not a proxy for global cloud traffic. Customer growth, individual workflows and broader geographic dispersion of AI data could all contribute, while a small number of large transfers can materially alter the results.
The changes are also affecting physical connectivity. Backblaze uses Tier 1 transit, Internet Exchanges and private network interconnections (PNIs), which are dedicated links between networks. It reported expanding its PNI footprint during the past year and moving from paired 100 gigabits-per-second connections toward multiple 100 gigabits-per-second links and dense 400 gigabits-per-second deployments. Traffic over PNIs increased 79% during the year, while traffic carried by internet transit providers rose 61%.
The transferable lesson is that average bandwidth is not enough for planning.