Internet Health Check 2026: Cloud Outages Climb While ISP Failures Cool — What Hosting Operators Should Do Next
The internet had another bumpy stretch. According to Cisco-owned ThousandEyes’ ongoing tracking, the week of August 17–23 produced 534 global network outage events across ISPs, public cloud provider networks, collaboration platforms, and edge services including DNS, CDNs, and security-as-a-service. That total edged down just 2 percent week-over-week, but the composition shifted in ways that matter to anyone running a website, VPS fleet, or production application: ISP outages fell sharply while public cloud network outages climbed. In 2026, availability is increasingly decided not by whether your server stays powered, but by whether the networks between your users and your infrastructure stay healthy. Below is a breakdown of the numbers, the notable incidents, and the operational lessons hosting buyers should be acting on now.
Related ServerSpan guide: Cloudflare Global Outage November 18, 2025: Why Centralized Infrastructure Is a Single Point of Failure (And What VPS Hosting Gets Right).
A Shifting Mix: Fewer ISP Outages, More Cloud Instability
In the most recent reported week (August 17–23), global ISP outages dropped 15 percent, from 298 to 252, and U.S. ISP incidents fell 12 percent, from 169 to 148. Public cloud moved the opposite way: global cloud network outages rose 17 percent, from 162 to 189, and U.S. cloud outages jumped 22 percent, from 143 to 174. Collaboration app networks stayed nearly silent at one global event and zero in the U.S. Notably, the U.S. total actually ticked up 2 percent (359 to 366) even as the global figure declined — regional swings are normal, but the cloud-versus-ISP divergence has been consistent for weeks.
Zooming out over the summer, the totals tell a story of persistent volatility rather than crisis. Weekly global events ranged from 443 (June 29–July 5) through 509, 587, and a 2026 peak of 610 during the week of July 20–26, before settling back to 514, 502, 546, and finally 534. The low-water mark came in late May at 286. Cloud outages have been especially whipsaw-prone: during the week of July 6–12, public cloud incidents surged 72 percent globally (136 to 234) and 84 percent in the U.S. (120 to 221).
One caveat worth keeping in mind: these counts measure event frequency, not severity. A nine-minute transit blip counts the same as an hour-long regional failure. But when one category — public cloud — repeatedly drives week-over-week spikes, that tells hosting buyers where concentration risk lives.
When Tier 1 Carriers Blink: Arelion, Cox, and a Familiar Chicago Pattern
Two incidents dominated the latest reporting window. On August 21, Arelion (formerly Telia Carrier), the Stockholm-headquartered Tier 1 provider, suffered a disruption lasting 1 hour 21 minutes across a 3-hour-10-minute window beginning around 9:25 PM EDT. It started at nodes in Chicago, spread to Atlanta within five minutes, briefly cleared, recurred, pulled in Seattle, and eventually extended to Dallas, San Jose, Sweden, and the U.K. — with downstream impact across dozens of countries from France and India to Brazil, Japan, and Australia. Arelion also had a shorter 20-minute incident on August 14, again initially centered on Chicago along with Boston, Newark, and Dallas.
Cox Communications appeared twice as well: a 14-minute event on August 11 (Chicago, then Cleveland and Tempe) and a 24-minute event on August 17 centered on Chicago nodes, affecting partners in the U.S., Mexico, the U.K., Austria, Canada, Germany, and Denmark. Earlier in the summer, the roster read like a Tier 1 hall of fame — Cogent (32 minutes, August 6), Zayo (17 minutes, August 5), AT&T (20 minutes, July 27), Liberty Global (67 minutes, July 30), plus Lumen, GTT, Hurricane Electric, Verizon Business, TATA Communications, and others across the year.
Chicago keeps showing up as the initial epicenter, which ThousandEyes reports purely as node geography without stating root causes. Still, the pattern reinforces a simple truth: major interconnection hubs concentrate risk. If your dedicated servers or colocation footprint ride a single dominant transit path through one metro, a carrier hiccup there becomes your outage. Shared and VPS customers inherit their host’s transit mix invisibly, which makes it worth asking prospective providers directly which upstream networks they buy from and whether they peer directly at major exchanges.
Maintenance Is Quietly Becoming the Leading Outage Vector
Some of the year’s most instructive failures weren’t cable cuts or hardware deaths — they were self-inflicted during routine work. On July 23, Microsoft confirmed that a bug encountered during regular maintenance removed IP routes between its West U.S. datacenter and its WAN from more devices than intended. ThousandEyes observed roughly an hour of disruption; Microsoft initiated a rollback around 1:45 PM EDT, restored the network by 2:26 PM, and saw full service recovery by 3:41 PM. Customers across the U.S., Canada, Asia, and Europe felt it.
For a more detailed walkthrough of this part of the topic, read Server Monitoring on Your VPS: Self-Hosted Uptime and Health Monitoring for Developers.
Cloudflare’s February 20 BYOIP incident followed the same script: a bug introduced during an automated internal maintenance task caused the company to unintentionally withdraw customer IP address advertisements from the internet, producing connection timeouts for about 1 hour 40 minutes until routing stabilized. Add Microsoft 365’s April 2 Central U.S. degradation (infrastructure entering an unexpected degraded state, 1 hour 14 minutes) and Google Gemini’s June 10 backend failure — roughly seven hours of degraded chatbot responses with no network problems visible on the frontend path — and a theme emerges clearly.
ThousandEyes itself flagged this in its 2025 retrospective: single symptoms mislead, and if the network looks healthy while users struggle, the fault is often in the backend. The Uptime Institute’s 2026 Annual Outage Analysis backs the broader trend — network and connectivity issues are now the most frequently reported cause of IT service outages, human error contributed to some degree in 92 percent of significant incidents over the past three years, and while impactful-outage rates improved (half of operators versus 74 percent in 2020), gains are slowing under pressure from AI workloads and growing third-party dependencies.
For hosting buyers, the takeaway is uncomfortable but useful: change management quality is now a vendor-selection criterion. Ask how providers stage rollouts, whether maintenance windows are published, and how quickly they can roll back. And keep offsite backups with a different provider so recovery never depends on whoever is currently broken.
Hardening a Stack You Don’t Fully Control
None of this means the internet is failing — weekly counts oscillated between roughly 290 and 610 events all year, and most were short and contained. But it does mean resilience has to come from architecture, not vendor promises:
- Monitor externally. Synthetic checks from independent vantage points catch upstream reachability failures your host’s dashboard will never show.
- Split your DNS. Authoritative DNS across two providers and networks keeps resolution alive when an edge or cloud event hits one vendor.
- Diversify deliberately. Critical workloads deserve distinct providers and regions — a backup instance in the same facility behind the same carrier isn’t a backup.
- Interrogate transit. Ask hosts for their upstream mix, peering presence, and failover behavior before signing.
- Treat SLAs realistically. Credits won’t repay lost revenue; design for graceful degradation instead.
- Rehearse recovery. Backups only count if the last restore actually succeeded — test on a schedule.
Key Takeaways
- Cloud network outages rose 17 percent globally (22 percent U.S.) in the latest week while ISP events fell 15 and 12 percent respectively.
- Tier 1 carriers — Arelion, Cogent, Zayo, AT&T, Lumen — all logged multi-region incidents this summer; multihoming still pays.
- Chicago repeatedly appeared as an initial outage epicenter, highlighting interconnection concentration risk.
- Headline 2026 failures traced back to routine maintenance bugs at Microsoft and Cloudflare, not exotic attacks.
- Uptime Institute ranks network/connectivity as the top IT-service outage cause, with human error implicated in 92 percent of significant incidents.
- External monitoring, redundant DNS, and cross-provider backups remain the cheapest insurance available.
The internet’s health in 2026 is best described as resilient-but-fragile-at-the-edges: totals fluctuate, categories trade places, and the biggest risks hide inside ordinary change processes. Watch next week’s ThousandEyes update, but more importantly, audit whether your own stack would survive a repeat of this month’s incidents.