Alphabet's London Cloud Outage Was a Generator Failure — and It Landed 99 Days After the UK Called It Critical Infrastructure

Google's own record of the europe-west2-b incident is blunt about cause and silent about the eight hours that followed: a power event caused an outage within the zone, the power event was mitigated, and services then recovered in stages. Reporting on the recovery adds the part that matters — mains power returned while backup generators failed to engage, so engineers held traffic in a staged drain while technicians verified the backend. the last line of defence failed before the software was ever tested.

The blast radius was wider than the zone that failed, because what Google listed as unavailable was the shared substrate rather than isolated products: Virtual Private Cloud, Cloud NAT, Cloud Load Balancing, Pub/Sub, Cloud SQL, BigQuery, Cloud Storage, Kubernetes Engine and Compute Engine. Google also recorded elevated errors and intermittent network issues across the wider region. A bank running primary and standby inside the same region had no route out of it, and Google has not yet published a post-incident review explaining why the generators did not engage.

On 10 July 2026 HM Treasury designated four global cloud providers as Critical Third Parties to the UK financial sector — Amazon's AWS EMEA entity, Alphabet's Google Cloud EMEA, Microsoft's Ireland entity and Oracle's UK entity — and the FCA, PRA and Bank of England began oversight on 13 July. Designated providers must identify and manage risks to their critical services, keep open and timely communication with regulators and with the firms that rely on them during major incidents, and support joint work on system-level risk. the 20 October failure is the first live test of those expectations.

The gap is wider than the financial sector. NHS England and the courts were knocked over on the same morning, and neither sits inside the financial-services regime, so the part of the outage that reached the public had no regulator with a mandate to intervene in the provider. That is why the episode produced a supervision question rather than a liability question — and why the 99-day gap between designation and failure matters more than the outage itself.

Alphabet's June quarter is the context that makes the outage legible. Revenue reached $119.8bn, up 24%, Google Cloud delivered $24.8bn, up 82%, and Cloud operating income was $8,814m on a 35.6% margin. Capital expenditure in the quarter was $44.9bn — 37.5% of revenue. That is what an AI buildout at this pace looks like on a cash statement: servers, land, substations, switchgear and generation. A cloud whose constraint is megawatts has a reliability problem that is a power problem, and London just demonstrated it. the marginal dollar of cloud capacity is now an electrical asset, not a software feature.

Three suppliers sit directly in that path and all three are already running at cycle-high order books. Vertiv posted Q2 2026 net sales of $3,274m, up 24% year on year, with a 37.7% gross margin and a 19.5% operating margin. Eaton reported second-quarter sales up 21%, 12-month rolling Electrical Americas orders up 41% and Electrical segment backlog up 43% year on year, with segment margins of 23.1%. HD Hyundai Electric signed a $721.3m data centre switchgear and power contract in July, deliveries running through 2028. redundancy-grade power content per megawatt is the new unit of demand.

Google Cloud publishes a separate service level agreement for each product, pairing an uptime commitment with a schedule of service credits. That structure is why the financial answer to \"who pays\" is unsatisfying: with 38 services affected, a customer must reconstruct its loss across 38 different credit tables, and credits are a share of fees, not compensation for the customer who could not withdraw money or attend a hearing. credits cap the provider's exposure at a fraction of the customer's loss.

The Bank of England's own research is why this is a hard policy problem rather than a simple tightening. Its Staff Working Paper 1199 finds that cloud outsourcing is associated with lower operating costs and higher deposits, with the demand-side benefits largest for small and medium banks and building societies; restricting it would raise market concentration, shrink smaller institutions' market shares and reduce depositor welfare, and roughly 32% of the direct welfare gain from lower capital requirements is already offset by weaker incentives to adopt cloud. Concentration creates a correlated tail, but the efficiency that created the concentration is real. regulators have to cap the tail without killing the gain that smaller banks rely on.

Equifax shows the arithmetic on the customer side. Its UK credit reporting platform was affected, yet revenue of $1,692.5m in the quarter ended Sep 30, 2026 was 9.6% higher than a year earlier, with operating income of $355.3m. Nothing in the demand data broke; the cost lands in remediation, second-region architecture and credits, all trivial against $6.59bn of trailing revenue. The uncomfortable part is not the cost to the credit bureau, it is that a credit-decision outage is a financial-stability event that no single vendor's service level agreement was written to absorb.

Seven weeks before London, the failure mode looked entirely different. On 3 September 2026 three of the leading conversational AI services failed within a single roughly 90-minute window, according to a Cloud Security Alliance research note published on 19 September 2026. That was a shared cloud dependency failing. October's was a shared electrical dependency failing: one zone's power feed and the generators that should have covered it. two failures, two different layers, one conclusion — provider names are not independent bets.

This is where the commercial pressure lands. Alphabet disclosed that Google Cloud backlog had nearly doubled quarter on quarter to more than $460bn in the first quarter of 2026, and the segment grew 82% to $24.8bn in the June quarter. Those are multi-year commitments, and every one of them inherits the electrical single point of failure of the region it lands in. Buyers of AI capacity will increasingly price resilience as a term of the contract rather than a feature of the service. resilience stops being an engineering detail and starts being a line item.

Read this event as a capital-allocation signal rather than a disruption. One zone of one region will not show up in Alphabet's profit and loss, and the honest conclusion for a shareholder is that the episode was financially irrelevant to the provider. It is anything but irrelevant in three other places: the capital expenditure line of the hyperscalers, which now buys redundancy as well as compute; the order books of the switchgear, transformer and uninterruptible power suppliers who build that redundancy; and the infrastructure budgets of UK financial institutions that have just been shown what single-provider concentration looks like from the customer side. The first two are investable on published numbers today. The third is a supervision story, and its first data point arrives in March 2027. the outage is worth more to the power chain than to the cloud.