Preliminary Incident Report — UKB Power Outage, Tuesday 15th September 2026

Introduction

We architect our platform to be as resilient as possible, but on Tuesday 15th September 2026, for a total of 36 users across 5 of our customers, that promise was not met. Computle provides the platform for staff to work, and we recognise the practical and financial impact that outages have on users. In this preliminary incident report we break down what happened, and what we are doing to mitigate this issue occurring again. The root cause has not yet been confirmed, and we will issue a full Root Cause Analysis once the data centre provider has accounted for the works carried out on their side.

Impact summary

Background

Computle operates from third party data centres, and in each data centre, two power feeds are provided. These power feeds are powered by two independent power substations, independent battery backup systems, and independent generators. In addition, Computle operates its own network, AS213513, which is provided by at least two routers per site, connected to multiple upstream internet providers. For further resiliency, each user is assigned a single workstation, and customers are split across multiple racks, so that any failure at a rack level mitigates the impact on a customer. We select data centre partners that contractually guarantee 100% power availability, and until this incident we had maintained that standard across our estate.

Timeline

January 2026

Our UKB data centre provider informed us that the site had reached its power capacity and would be undergoing major building works. In response, Computle began work on a new site, UKC, which carries a significantly higher MW allocation than UKB and has no major building works planned or underway.

Monday 10th August 2026

Computle was informed by our UKB data centre provider that they would be “undertaking planned works to install new devices across the UPS A and UPS B output switchboards”. A scheduled date of Tuesday 1st September 2026 was provided for power feeds A, and Wednesday 2nd September 2026 for power feeds B. We raised concerns that this work posed a risk for our customers and were informed that the work was mandatory. As part of this, we requested manual bus-bar power readings to confirm the live power load, so that we could plan the purchase of ancillary power equipment and ensure that each A and B power feed could handle the load of each rack.

Tuesday 11th August 2026

Computle began work to install Automatic Transfer Switches (ATS) in all the UKB racks. This work was completed by Saturday 22nd August 2026. In the event of a power feed being removed, these devices would ensure that each workstation maintained its power availability.

Tuesday 1st September 2026

Computle was informed that the A feeds were isolated and that power had been transferred to the B feeds. No customer impact was noticed and the ATS units handled the power interruption as expected. At the time of the switchover, 15.1kW of power draw was recorded across racks 3, 6, and 7. We were then informed that the works on the B feeds were “cancelled and will no longer proceed as previously communicated” and that “the revised maintenance dates have not yet been confirmed.”

Wednesday 2nd September 2026

We were informed that the B feed work would now take place on Monday 14th September 2026.

Thursday 3rd September 2026

We were informed that the work on B feed would now be undertaken on Tuesday 15th September 2026. This notice was sent only to Jake Elsley at Computle and, owing to a birth in the family, was missed.

Tuesday 15th September 2026

Detection and impact

At 8am, we received an email from the data centre provider informing us that the works were going ahead that day, with the UPSB01 feed to be isolated at 10am and the UPSB02 feed at 12pm. Had the earlier change of date not been missed, we would have informed customers ahead of the works. Once the data centre provider isolated UPSB02, at 12:05, we started to receive telemetry alerts for a subset of machines with a large increase in activity over the next few minutes. 36 machines in total were flagged as unreachable. At 12:10pm, our automated recovery systems began a series of outbound emails to impacted users. At the same time, manual communications were sent to the impacted customers and a support ticket was raised with the data centre.

Power load at the time of the outage

Prior to the time of the outage, power draw across racks 3, 6 and 7 was 16.3kW; within the safety rating of 21.68kW for the ATS units and within the 21.68kW ceiling for the APC power strips that supply the ATS units. These readings rule out overload as a cause: the racks were drawing well within the rated capacity of both the ATS units and the power strips feeding them. We were then informed that power would be restored in the next four hours.

Working theory

The ATS units in racks 3, 6 and 7 are designed to hold a workstation up on whichever feed remains live, and they did exactly that on Tuesday 1st September. For all three racks to lose their workstations simultaneously implies that neither feed was live — that is, a total loss of both the A and B supplies at those racks, rather than a fault in the ATS units themselves. As it is highly unlikely that three racks’ power supplies fail independently at the same moment, our current working theory — not yet confirmed — is that the method used to isolate the B feeds on Tuesday 15th September differed from that used on the A feeds on Tuesday 1st September, and that this left the A feed inoperable. We requested confirmation that the A feed was in operation and were informed that it was, however, the 36 machines were still showing a no power state.

Recovery

Each machine is configured to automatically power on when power is restored. When power was restored to the B feed, we observed 7 machines return to a healthy state, with 29 remaining unreachable. Manual intervention was required on the remaining 29 hosts, and service has since been restored for all 36 users. Unfortunately, we have no detail from the data centre provider as to what was done differently on Tuesday 15th September, if anything, that would result in this loss of power. We have formally requested a full account from them and will update this RCA once the cause is confirmed.

Remediation and next steps

Site attendance: our immediate priority is confirming the physical state of the affected hardware. We are attending the UKB site in person to verify each of the 36 machines, and will update this RCA once that visit is complete.

Relocation: we are relocating all machines in UKB to UKC, the site we began building out in January 2026. UKC carries a significantly higher MW allocation than UKB and has no major building works planned or underway. The move is scheduled to take place over the Christmas period to minimise disruption, and we will confirm exact dates and any per-customer impact closer to the time.

Provider communications: we are asking the UKB data centre provider to send all communications to everyone at Computle, not only to Jake Elsley, so that other members of the team can pick them up and inform customers.

In-person meetings: we will be holding in-person meetings with all impacted customers.

Service credit: a service credit of 50% of the monthly fee for each affected machine will be applied automatically. No action is required from you.

Corrective actions

Redundant power supplies: we are starting a programme to replace the power supplies in every workstation with per-workstation redundant units, at no additional cost to customers. Each workstation will then draw from both the A and B feeds directly, removing any dependence on the ATS units and ruling them out as a point of failure.

Live host respawn: we are bringing forward work to introduce live host respawn, which restores a host from snapshot onto alternative hardware. We should be clear that this would not have helped in this incident — where both feeds are lost at a site, there is no healthy local host to respawn onto — but it materially shortens recovery from single-host and single-rack failures.

Provider accountability: we have asked the UKB data centre provider for a full written account of the works carried out on Tuesday 15th September and of any difference in method from the A feed works on Tuesday 1st September. We will publish a full Root Cause Analysis once we have it.