AWS US-East-1 Cooling Failure Took Down Coinbase and FanDuel for 18+ Hours
A thermal event at an AWS data centre in northern Virginia knocked out a single Availability Zone in US-East-1, starting at 8:25 PM ET Thursday and remaining unresolved as of 3:29 PM ET Friday — more than 18 hours later. AWS described efforts to restore service as “slower than we had previously anticipated.”
The cause: overheating. AWS confirmed a cooling system failure at the facility and was working to bring “additional cooling system capacity online” before recovering the remaining impaired hardware. EC2 virtual server instances were the primary affected service.
What Went Down
Coinbase halted trading operations. The exchange posted outage notices through Thursday night and into Friday morning.
FanDuel — the sports betting platform — reported technical difficulties that prevented users from placing bets, a problem with material revenue implications during live sporting events.
AWS said the failure was contained to a “single Availability Zone” within US-East-1, its flagship North American region. Both Coinbase and FanDuel are multi-AZ deployments in theory; in practice, the failure cascaded into their customer-facing services.
Why This Matters for AI Infrastructure
US-East-1 is the most-used AWS region for AI inference workloads. Anthropic, OpenAI, and most major AI API providers route significant capacity through the region. An AZ failure lasting 18+ hours represents a class of reliability risk that has historically been treated as near-impossible — the whole point of multi-AZ architecture is that individual AZ failures should be transparent to customers.
The thermal origin matters specifically for the AI era. Modern AI inference clusters operate at power densities 3-5x higher per rack than traditional cloud workloads. A GB200 NVL72 rack pulls over 120 kW; a standard 2026 enterprise server rack pulls 15-20 kW. Data centres designed for the latter will overheat running the former. The cooling infrastructure in most facilities was not designed for this power density, and upgrades to support it involve significant physical construction that cannot be done while systems are live.
AWS has not confirmed whether AI workloads contributed to the thermal event.
The Recovery Problem
AWS’s statement that cooling capacity needs to come back online before hardware can be recovered indicates that the impaired instances cannot simply be migrated — they are physically hot and cannot safely restart until the cooling environment is stable. This is a meaningfully different failure mode than a networking or software fault, which can often be remediated in minutes.
The 18+ hour recovery timeline, in a facility that theoretically has redundant cooling systems, suggests either the primary and backup cooling failed simultaneously or the scope of the thermal event exceeded what the redundant systems could handle.
The US-East-1 event is one data point. But as AI workloads drive power density far beyond historical data centre design parameters, cooling failures are moving from unlikely edge cases into a category of risk that infrastructure architects need to plan for as a first-order concern.