ATAILA Newsroom · Budapest ·
2026-W43
One zone lost power, Revolut stopped. DORA keeps the duty with the bank.
Most Hungarian buyers who ask us about resilience have one cloud region in mind and three zones on the diagram. On 20 October that diagram was tested in London, and a large part of Revolut's Hungarian customers lost their app and their cards for nearly half a day. Google's preliminary report is candid about why the healthy zones failed as well, and it deserves a careful reading. The law puts the consequences on the bank, so the questions that matter are the ones a customer asks its provider before the next outage.
The article we are responding to
“Előszedték a Google-t a Revolutot is kifektető keddi üzemzavar miatt”
hwsw.hu, 22 October 2026
What happened on 20 October
Google Cloud published a preliminary incident report on 22 October. It is our primary source for the cause and the timeline, and it may still change: Google says a final report with preventive actions will follow.
At 06:53 BST on 20 October 2026 a power grid disruption hit the datacentre facility that hosts europe-west2-b, one of the three availability zones of Google Cloud's London region. The load moved to UPS batteries while the diesel generators started up. The generators never took the load. Google writes that the PLC automation system, "both primary and backup PLCs", failed to start and transfer them. When the batteries ran out, from 07:00 BST, the zone's servers, network switching and local storage went offline. Monitoring alerted Google's engineers at 07:08 BST.
Utility power came back within an hour, according to the report. The disruption lasted 6 hours, 10 minutes, from 07:00 to 13:10 BST, which is 08:00 to 14:10 in Budapest. Most services recovered by 12:29 BST, new virtual machines could be created again from 12:46, and by 13:10 over 99% of Compute Engine virtual machines were healthy. Google posted the closure at 15:03 BST. The report lists 40 affected products.
The damage did not stay in one zone. Google writes that customers in europe-west2-a and europe-west2-c, and customers with regional architectures, "also experienced unexpected impact". It gives two causes: "instantaneous demand surges on unimpacted compute capacity" and "synchronous control-plane validation dependencies across zonal boundaries". BigQuery, Looker, Compute Engine, Kubernetes Engine, Regional Persistent Disk and Cloud SQL were among the services hit this way.
hwsw.hu, citing Reuters, names Revolut, ING in Germany, HSBC, Bank of Scotland, DPD, the NHS and Royal Mail among the customers affected. The Register adds Lloyds, Rightmove, BT and Screwfix, and quotes a Keele University lecturer who estimates that 38 services were down.
For Hungarian customers the outage was Revolut. hwsw reported on the morning of 20 October that for an undetermined group of customers many services were unavailable and in the worse cases the app would not start, that currency exchange and transfers failed, and that card payments and cash withdrawals could also be affected, on iOS and Android alike. hwsw notes that Revolut has given about two million customers a Hungarian account number since the summer of 2026, and that this was its third customer-visible fault in October. On 22 October hwsw wrote that a significant share of Hungarian customers had no app, card payments or ATM access until the afternoon. Telex reported that Revolut's status page showed errors on all services from early afternoon until the evening, that the fault was fixed by the evening, and that shortly before 8 pm Revolut confirmed the Google Cloud failure as the cause. Revolut said customers' money and data were safe throughout.
Then the supervisors. hwsw writes that the European Banking Authority said on 21 October that it had opened an investigation, because of the number of financial institutions affected. Portfolio, citing Reuters on 22 October, reports something narrower: an EBA spokesperson said the authority had contacted Google Cloud and asked for more detailed information, to assess the possible impact on EU financial institutions. These are two different claims, and we found no statement on the EBA's own press-release page on 22 October. In the UK, according to the same Reuters report, a parliamentary finance committee wrote to the Bank of England and the FCA, and the FCA said it had contacted Google.
The regime behind those letters is new. According to Portfolio and hwsw, the EU designated 19 technology companies in 2025 as critical providers to the financial sector, under the joint oversight of three European supervisors, and the UK named four of its own. Google Cloud falls under both. Neither regime moves the responsibility. DORA, Article 28(1)(a), says that financial entities using ICT services "shall, at all times, remain fully responsible" for their obligations. hwsw draws the same conclusion: it expects no serious sanction for Google.
6 hours, 10 minutes
duration of the disruption in Google Cloud's europe-west2 (London) region, 07:00 to 13:10 BST on 20 October (Google Cloud preliminary incident report, 22 October)
within 1 hour
time to restore the utility power feed, per the same report; the rest of the six hours went into recovery, failover and re-admission
over 99%
of Compute Engine virtual machines healthy by 13:10 BST; the remaining customers were directed to file a support case (Google Cloud, 22 October)
40
Google Cloud products listed as impacted in the preliminary report; a Keele University lecturer estimated 38 services down (The Register, 20 October)
about two million
Revolut customers given a Hungarian account number since the summer of 2026 (hwsw.hu, 20 October: "mintegy kétmillió")
19
technology companies the EU designated in 2025 as critical providers to the financial sector; Google Cloud is covered (Portfolio via Reuters, 22 October; hwsw.hu)
Our reading: three zones, one region
The zone failed the way zones are designed to fail: as a unit, with a clear physical cause. The redundancy that should have prevented it was there on paper. There were two generator controllers, and both failed to start the generators. Redundancy that is not exercised under real conditions is an assumption.
The more important lines are about the other two zones. A region with three zones is sold as protection against the loss of one. Google's own report says the healthy zones were degraded because their control planes waited on validations from the dead zone, and because everyone's failover arrived at the same spare capacity at the same moment. That is a structural property of a shared region. A customer who designed for zone loss on the diagram met a dependency that the diagram does not show.
Recovery took longer than the power cut. Utility power was back within an hour; the region took 6 hours, 10 minutes. The time went into restarting infrastructure, checking storage integrity and letting traffic back in gradually. Any operator, ours included, faces the same order of work after a hard power loss.
Revolut's own timeline is longer than Google's. Google reports that most workloads had recovered by 14:10 Budapest time. hwsw says Hungarian customers were cut off until the afternoon; Telex says Revolut was restored by the evening. Part of what Hungarian customers lived through therefore sat in Revolut's own architecture and recovery, which we cannot see. Portfolio, citing Reuters, says Revolut is reviewing the incident to reduce the risk of third-party disruption.
Two portal readings need correcting. Portfolio writes that the failure affected the network serving Google's whole European customer base; Google's report describes one region. And hwsw's "investigation" is a stronger word than the EBA's request for details as Reuters reported it. Under DORA the distinction matters less than it seems. Whatever the EBA asks Google, the bank answers for its customers, and DORA, Article 28(8), requires exit strategies for ICT services that support critical or important functions. We made the same point about NIS2 in the obligation does not move with the regulator.
What to ask any provider, ours included
AI workloads now sit in the same chains. A document pipeline, an internal assistant or an agent that calls a model endpoint in one cloud region fails when that region fails. In a bank or an insurer, if the tool supports a critical or important function, it is an ICT service under DORA like any other. We wrote about the same dependency when three AI services stopped at once in September.
The London report turns into five questions a buyer can put to any provider, including us. Which failures share a control plane, and what does a healthy zone or site wait on when another one dies? Where does failover capacity come from, and is it reserved or shared with every other customer who fails over at the same time? When was failover last tested under load, and how long did re-admission take? Who starts the generator, and when was that tested? Can the workload run somewhere else, with its model and its data, without the provider's permission?
ATAILA offers two forms. ATAILA Cloud is managed and multi-tenant, on our hardware in the EU. Enterprise is single-tenant, on the customer's own hardware. Both run open-weight models, both come with sandbox, dev, uat and prod environments, and both carry 99,9% availability. The relevant difference for this story is the exit. With Enterprise the model files, the data and the hardware stay with the customer, so a second site and a fallback can be the customer's own decision. The uat environment is where a failover test belongs before anyone needs it in prod. None of this removes the generator from the building. The silicon is NVIDIA's, not ours.
Where our argument is weak
This was one zone for about six hours, in a region that recovered the same day. We cannot claim that a typical private server room recovers faster from a hard power loss, and Google published a preliminary root cause within two days.
A single-building private deployment carries the same generator risk. Many server rooms have UPS batteries and no generator at all. An Enterprise installation in one building is one zone with no neighbours, and nothing in our design changes that unless the customer pays for a second site.
Our own availability figure would not have been breached by this outage. By our arithmetic, 99,9% over a year leaves about 8.8 hours of downtime (0.1% of 8,760 hours), more than the 6 hours, 10 minutes in London. A percentage says how much failure is allowed. It does not say how failures happen.
Revolut's recovery ran into the evening, hours after Google's region had recovered for most workloads. Moving to private hardware does not fix a recovery design. And a bank that leaves a large, supervised provider for a small one like us exchanges concentration risk for small-vendor risk. We are not a designated critical provider, and no European supervisor oversees us directly.
What we concede, and what we will be asked
Google's report is good evidence. It was published within two days, it names a component failure in plain words, and it admits that the healthy zones were hit. Many operators would not publish that much. Multi-zone design still did part of its job in London: the report says data services with failover capability, such as Cloud SQL and AlloyDB, moved to the unaffected zones.
A sceptical reader will ask whether we have ever cut the power to one of our own sites and timed the recovery, whether we can show a report in the same form as Google's, and what happens to an Enterprise customer whose building has no generator. These are fair questions, and the honest answer to the last one is that the building is the customer's.
What we commit to. We will follow Google's final incident report and correct this article if the root cause changes. The five questions above apply to us, and a buyer who asks them should get written answers with test dates. The isolation and exit evidence we said we owe in our reading of the AWS sovereign cloud test should include a failover test under load, with its times.
Sources, figures, and one disclosure
The story we are answering: Előszedték a Google-t a Revolutot is kifektető keddi üzemzavar miatt, Koi Tamás, hwsw.hu, 22 October 2026. Read in full from the text our research pass saved on 22 October.
The primary source for the cause and the timeline: Google Cloud's status incident for europe-west2 and its Preliminary Incident Report, posted 22 October 2026, status.cloud.google.com, read in full from the status feed, with every update since 20 October. Times are as Google prints them in BST; the Budapest times are our conversion. The final report has not been published.
The supervisors: Elestek a bankok, azonnal magyarázatot követelnek a Google-től a pénzügyi szektort megbénító leállás miatt, Portfolio, 22 October 2026, which names Reuters as its source and says an AI assistant helped prepare it. Read in full. The Reuters original was not opened. The EBA press-release page was opened on 22 October and carried no statement on Google; it did carry a notice that the EBA website was offline from 21 October 13:00 CEST to 22 October 14:00 CEST for security updates.
Revolut in Hungary: Súlyosabb leállás lehet a Revolutnál, Koi Tamás, hwsw.hu, 20 October 2026, which relies on the Revb blog; Közel fél napig állt a Revolut, csak estére sikerült helyreállítani a megfelelő működést, Aradi Hanga Zsófia, Telex, 20 October 2026, updated with Revolut's evening statement. Both read in full. hwsw (the afternoon) and Telex (the evening) differ on when service returned; we report both.
The customer list and the 38 services: Egy sor szolgáltatót magával rántott a Google Cloud mai üzemzavara, Koi Tamás, hwsw.hu, 20 October 2026, citing Reuters; Google Cloud says a 'power event' caused today's outage, Paul Kunert, The Register, 20 October 2026, updated the same evening. Both read in full. The Register puts the start at 06:08 UTC, which matches Google's alert time rather than the 07:00 BST start in Google's report.
DORA: Regulation (EU) 2022/2554, Article 28, read on an unofficial copy of the text because EUR-Lex answered with a bot check; the quoted words of Article 28(1)(a) are from it. The 99,9% downtime arithmetic is ours.
Not opened: the Index and hvg reports, two earlier Portfolio reports and one earlier Telex report of 20 October, known to us by title only.
The disclosure: ATAILA sells private and on-premises AI infrastructure, which competes with public cloud for the same budgets. A hyperscaler outage suits our argument, which is why this article names where that argument is weak.
The original article
“Előszedték a Google-t a Revolutot is kifektető keddi üzemzavar miatt”
hwsw.hu, 22 October 2026
- Read it live at hwsw.hu
- ATAILA archive, captured 22 October 2026
The live page has changed since we captured it.
← Back to the Newsroom Press enquiries: contact us