Google says control wiring and configuration left UK generators without a start signal

Google says a power failure at a UK data centre facility hosting its europe-west2-b zone began at 22:53 US/Pacific on 19 October 2026 (06:53 BST on 20 October), and that the site's backup generators never received a signal to start. Its incident report, last updated on 4 November, traces the cause to a fault in an underground cable on the utility's network, outside the facility, which took out primary utility feed A.

The report says the site had two primary feeds, A and B. Feed B's breaker had been opened for repairs after problems turned up in a generator load test it runs every quarter, on 14 October, so feed A was carrying the whole critical load. When feed A failed, the controllers that command the generators did not start them. Google says the controllers' voltage-sensing inputs were wired upstream of the open feed breakers; they still saw voltage on feed B's line side, read it as a healthy utility source and issued no start command. Google blames a flaw in the voltage-sensing design and a misconfigured controller program, which did not check both undervoltage and breaker position on each feed.

Google says uninterruptible power supply batteries, which give roughly five minutes of runtime, carried the load, as designed, until the batteries were drained, and the report puts the loss of power in the data halls at 23:00 US/Pacific. Site staff tried a remote load transfer through the controllers, but Google says the system blocked it because feed B still showed live voltage. The team then moved the facility to a standby utility connection, and power reached the data halls again at 00:20 US/Pacific on 20 October, the report says. On-site teams and the equipment makers then reconfigured and validated the controllers, according to the report.

Google says services in europe-west2 were disrupted for six hours and eight minutes, from 23:00 to 05:08 US/Pacific, when 99.5 per cent of its Compute Engine VMs were back in normal operation. The report says a small number of isolated workloads recovered after that, and the incident was closed by 07:03. It adds that some customers running in the region's other two zones were also affected, which it attributes to sudden demand surges on those neighbouring zones and to control-plane dependencies that cross zone boundaries.

According to the report, the utility located the cable fault on 26 October. Google's estimate for the utility to repair the cable and restore feed A moved from 30 October, in the 27 October version of the report, to 13 November in the 4 November version, and its estimate for switching the facility from the standby connection back to feeds A and B moved from 31 October to 17 November. Google says the controllers were reconfigured and validated on 20 and 26 October, complete transfers of the load to the generators were tested on both dates, the on-site team was practising updated emergency operating procedures twice per shift by 24 October, and planned maintenance at sites with similar designs was frozen, a step completed on 30 October. Still planned, the report says, are a controller code review with the facility vendor (estimated for 6 November), a check of controller settings at sites with similar layouts and controller makers (15 November), generator transfer tests once both primary feeds are back (17 November), and controller code reviews during the commissioning of future facilities (18 December).