What happened
At 11:50 PM UTC on May 7, 2026 (4:20 PM PDT), multiple cooling units failed inside a single data center hall serving the use1-az4 availability zone in AWS's US-EAST-1 region, in Northern Virginia. As temperatures climbed past safe operating limits, servers on the affected racks shut down automatically to protect the hardware. That shutdown cascaded into a broader power loss, which impaired EC2 instances and degraded EBS volumes across the hall, according to AWS's own status updates (AWS Support).
Restoring cooling capacity to pre-event levels took roughly 20 hours, with AWS reporting stabilization by 1:50 PM PDT on May 8. Full recovery of the affected EC2 instances and EBS volumes stretched the total incident window to somewhere between 19 and 28 hours depending on which services are counted (IT Pro; shattered.io). More than 150 downstream cloud services showed impact during the window (StatusGator).
The visible casualties tell the story of who still concentrates critical workloads in a single AWS availability zone. Coinbase was down for about seven hours, leaving traders unable to close positions during a volatile stretch. FanDuel went dark mid-event, during Game 2 of the Lakers-Thunder Western Conference semifinal, cutting off users trying to cash out live bets. CME Group's CME Direct platform, used by institutional futures traders, also reported disruptions (CNBC; gurufocus).
Why this one is different from a code bug
The first thing worth separating out: this was not a bad deploy, a misconfigured load balancer, or a DNS resolution failure, the categories that have caused most of the high-profile hyperscaler outages of the last few years. It was a building that got too hot. That distinction matters for how platform teams should read the incident.
A software failure is something you can guard against with better testing, canary deployments, or faster rollback. A thermal failure in a single facility is a physical capacity problem, and it points at something structural: the gap between how fast rack power density is rising and how fast cooling infrastructure is being upgraded to match it.
The capacity mismatch behind the incident
Industry data on data center design gives that gap a number. Average rack density rose from roughly 16 kW in 2025 to about 27 kW in 2026, driven by AI accelerator deployments, and only about one in five data center operators report being ready to support the 50 to 70 kW racks that are now common in AI-heavy facilities (Data Center Knowledge). Operators across the industry are shifting from air cooling to liquid and hybrid systems specifically because legacy air-cooled halls, the kind built years before the current AI buildout, cannot dissipate heat at these densities without failing under load (Johnson Controls).
AWS has not said publicly that the affected hall was carrying AI-related density, and it would be speculation to claim it was. What is documented is narrower and still useful: a facility built to an older cooling spec failed when demand on it exceeded what that spec could handle, in a region where AWS, like its competitors, is simultaneously racing to add AI capacity. The mismatch between rack density growth and cooling upgrade cycles is now a documented industry-wide condition, not a one-off explanation for a single incident.
What it cost, and what it did not cover
AWS's standard service credit for an availability event covers about 10% of the customer's monthly compute spend on the impacted instances (tech-insider.org). It does not cover lost trading revenue, regulatory exposure from a financial platform going dark mid-session, or the reputational cost of a betting app failing during a live playoff game. For Coinbase, FanDuel and CME, the credit is a rounding error against what a seven-hour outage during active trading or wagering actually cost them.
That gap between the SLA credit and the real business exposure is the part platform engineering leaders should carry into their own capacity planning conversations, not the headline outage duration.
The buyer-relevant takeaway
This was AWS's third major US-EAST-1 failure since the original December 2021 incident, and the pattern each time has been similar: an outage concentrated in one availability zone within one region, with downstream damage determined almost entirely by which customers had built genuine multi-AZ or multi-region failover and which had architected around the assumption that "us-east-1 just works."
For a platform team, the actionable question after this incident is not whether AWS's cooling design was adequate for one specific hall. It is whether the organization's own critical services can survive the loss of a single availability zone for 20-plus hours without customer-facing failure. Multi-AZ deployment inside a single region reduces exposure to a rack-level or hall-level failure like this one; it does not protect against a regional event. Multi-region failover costs more in both engineering time and steady-state infrastructure spend, and for a service processing financial transactions or live bets during peak demand, this incident is a concrete data point for justifying that spend to a CFO who has previously treated it as redundant cost.
The uncomfortable second-order point is that the underlying capacity mismatch, rising rack density against cooling infrastructure that was not built for it, is not unique to AWS. It is a documented condition across the hyperscaler and colocation industry as all three major cloud providers race to add AI compute capacity in existing footprints. Buyers evaluating reliability commitments from any cloud provider should be asking pointed questions about cooling infrastructure age and rack density plans in the specific facilities hosting their workloads, not just reading the published uptime SLA.
Sources: AWS Support incident update · IT Pro: The AWS outage explained · shattered.io: AWS Outage Hits 28 Hours · StatusGator: AWS outage takes down more than 150 cloud services · CNBC: AWS data center outage hits trading on FanDuel, Coinbase · gurufocus: CME Experiences Disruptions Linked to AWS Outage · tech-insider.org: AWS Outage Hits US-East-1, 28 Hours, 10% Credit · Data Center Knowledge: AI Pushes Cooling to the Forefront of Data Center Design · Johnson Controls: AI infrastructure and increased rack densities