Cloud Computing Watch · October 21, 2026
A Power Event Takes Down a Google London Zone, OVHcloud Holds 110 MW of Open Capacity, and Amazon Buys Racks It Cannot Yet Power — Operational Capacity Is Becoming the Cloud's Scarcest Asset
Cloud Competitiveness Now Depends on Capacity That Is Powered, Connected and Reliable, Not on Announced Investment
Three stories this week make the same argument from different angles. A power event in Google Cloud's europe-west2-b zone in London disrupted UK banks and other businesses on October 20, a reminder that reliability is physical before it is architectural. OVHcloud says 110 MW of its 250 MW of connected capacity is open for new customers, an unusual claim in a market where capacity is usually the constraint. And at the other end of the spectrum, Microsoft has reportedly exited an Nscale data center that cannot get grid power until the 2030s, while Amazon is reportedly taking delivery of Nvidia racks before it can energize them.
The thesis: reliable, operational capacity is becoming a decisive advantage in cloud computing. Money, chip allocations, and contracts only turn into revenue when the power, the building, and the network are ready at the same time. The three stories below show what happens when they are, and when they are not.
Dating note: Most items are dated October 19 or 20. Related coverage appears in The Newsroom (October 21) and AI Watch (October 19); this edition adds the cloud-operations view.
Cloud Computing Watch at a Glance
| Development | Date | Status |
|---|---|---|
| Google Cloud London (europe-west2-b) power event | Oct 20 | Confirmed by Google; root cause analysis pending |
| OVHcloud: 110 MW available, first AI model planned | Oct 20 | CEO statements; model date unsettled |
| Microsoft exits Nscale's Loughton site | Oct 18 | Reported; Microsoft declined comment |
| Nscale and Anthropic talks on a second ~$45B deal | Oct 19 | Unconfirmed talks; terms may change |
| Amazon takes Nvidia racks ahead of power | Oct 20 | Reported from internal emails; Amazon calls it routine |
Cloud Reliability and Resilience
Google Cloud's London Outage Exposes Cloud Resilience Risks
What happened (October 20, confirmed by Google): The Register reports that the disruption began at 06:08 UTC when europe-west2-b, one of three zones in Google's London region, failed. Google's status page said the cause was a power event, and the incident produced elevated errors and intermittent network issues across services including Compute Engine, Google Kubernetes Engine, Cloud Storage and App Engine, per CX Today. Google said most services had returned to normal by about 12:10 GMT, roughly six hours after the incident began, per NZ Herald.
Who was affected: The AP reported disruption at UK banks including Lloyds, Halifax, HSBC and Revolut. Revolut said customer money and data were safe. The Register lists Royal Mail, Rightmove, DPD, BT and Screwfix among affected businesses, and ABC News reports Google said the incident was isolated to a London data center, with parts of Europe seeing errors, and that it would carry out an internal investigation. No data loss has been reported.
Why it matters: One zone out of three failing should, in theory, be survivable for a workload spread across zones. The breadth of disruption suggests that many customers either ran critical services in a single zone, shared a dependency that was not zone-independent, or hit failover limits. The public record does not yet say which, and Google's root cause analysis is the document to wait for. Google's status history also shows an August 2024 power-related incident in the same region, in which a partner facility's emergency power did not carry the load, so power resilience at the facility level is not a new question.
For enterprise architects: Multi-zone deployment is a floor, not a guarantee. Test failover for the whole dependency chain, including identity, networking, and third-party services. Regulated firms should map which critical services depend on one provider's one region, and rehearse how they would operate during a six-hour outage. Connect to: Company Intelligence · Hardware Watch · Cybersecurity Watch
European Cloud Competition
OVHcloud Has 110 MW Available as It Prepares AI Services
What happened (October 20, CEO statements): After publishing annual results, OVHcloud CEO Octave Klaba told Reuters that the company has 250 MW of connected capacity, 140 MW in operation and 110 MW ready to take new business, and that its first AI model should launch in early 2027. The model is reported to use a mixture-of-experts design. The company reported its first positive levered free cash flow since its 2021 IPO and targets revenue growth above 10% in 2027.
Caveats: EconoTimes notes that OVHcloud's results label 110 MW as available while the presentation describes the 140 MW in operation as contracted, and that La Tribune reported an initial open-source model in late November 2026 while OVHcloud's results give no model date. OVHcloud plans an AI Summit on November 19 to detail its AI strategy.
Why it matters: Powered, connected capacity that can onboard customers without new construction is a real advantage when rivals wait on grid connections. It gives OVHcloud a faster path to revenue than a competitor still building. The open question is demand: free capacity is only an edge if customers choose OVHcloud's services, and whether its AI model and sovereignty positioning are enough to win them. Klaba has also been sceptical of national AI-gigafactory plans that lack a matching customer base, which fits his preference for demand-led capacity. Connect to: Company Intelligence · Enterprise AI Intelligence
AI Cloud Delivery Risks
Microsoft's Nscale Setback Highlights AI Cloud Delivery Risk
Development A (October 18, reported): The Telegraph reported, per Reuters, that Microsoft terminated its agreement for Nscale's Loughton, Essex data center after grid officials signaled power would not arrive until the 2030s. Microsoft declined to comment. Nscale said its Microsoft relationship remains strong and that it is exploring alternative power sources.
Development B (October 19, unconfirmed talks): Bloomberg first reported, and Reuters confirmed via a source, that Nscale is in talks to sell Anthropic AI computing capacity in a new deal worth about $45 billion. A second building on Nscale's Monarch campus in West Virginia would deliver 460 MW using Nvidia Vera Rubin chips, and Anthropic reportedly has a right of first refusal on a third building. Anthropic agreed in August to spend $45 billion at the same campus. The source said discussions are ongoing and details could change. Neither company has confirmed the new deal.
These are separate stories. The Microsoft exit concerns a UK site and a power-delivery date. The Anthropic talks concern a US campus and a different customer. Reading one as evidence of the other would be a mistake. What they share is Nscale's customer concentration: its prospectus, per Reuters, names Microsoft and Anthropic as expected major customers.
Why it matters: Cloud contracts, financing, GPU commitments and grid connections run on different timelines. A customer can sign for capacity that its supplier cannot power, and a supplier can raise capital against backlog that a single delay reshapes. Buyers of AI capacity should treat delivery dates as contingent on utilities and ask for contractual protection when power slips. Nscale's planned IPO will put those assumptions in front of public investors. Connect to: AI Infrastructure Watch · The Term Sheet · Company Intelligence
Infrastructure Economics
AI Infrastructure Spending Runs Ahead of Usable Capacity
What happened (October 20, media report): Business Insider reported, per Investing.com, that internal August emails show Amazon created an "Early Rack Receipt" program to take delivery of roughly $5 billion of Nvidia GB300 server racks at US facilities ahead of final deployment, and expects to accept about $3 billion more each month. The report says some equipment is staged in intermediary warehouses because several AWS sites lack utility capacity to energize it, leaving some hardware idle. Yahoo Finance's summary says Amazon called the staging a routine best practice for fast installation once power arrives and disputed that it lacked a path to bring the facilities live. The emails are internal and the figures are unconfirmed by Amazon.
The power backdrop: Axios reports on a Goldman Sachs analysis expecting global data center power demand to rise 170% from 2025 to 2030, with about $7.6 trillion in capital spending from 2026 to 2031 across computing, data centers and power infrastructure, and roughly 2,600 GW of proposed generation and storage awaiting grid connection. A separate summary says about half of planned capacity may come online on schedule over the next couple of years.
Why it matters: The difference between buying equipment and bringing revenue-generating capacity online is time. Racks taken before power are capital tied up without rental income, as the report itself notes. Whether that is a rational hedge against scarce chips or a growing balance-sheet strain depends on how quickly power arrives. For cloud providers, utilization and time to revenue are now as important as the size of the order book. For customers, announced capacity is a weaker signal than energized capacity. Connect to: AI Infrastructure Watch · Hardware Watch · Enterprise Technology Intelligence · AI Infrastructure Special Report
Strategic Analysis: What Counts as Cloud Capacity?
| Stage | What it means | This week's example |
|---|---|---|
| Announced | Contract or investment pledge | Anthropic and Nscale talks (unconfirmed) |
| Purchased | Hardware delivered and owned | Amazon's reported early rack receipts |
| Connected | Power and network in place | OVHcloud's 250 MW connected |
| Reliable and earning | Serving customers through failures | Google's London zone: a test of this stage |
Investors and buyers often count the first two stages. Customers only pay for the last two. The gap between them is where delays, financing strain and outages show up, and it is the gap that these stories describe.
What to Watch Next
- Google's post-incident analysis. Google said it would investigate. Look for the root cause, why a single-zone failure had such wide impact, and any guidance on regional and zonal architecture.
- OVHcloud customer uptake. Whether the AI Summit on November 19 and the first-quarter fiscal 2027 revenue update on January 7 show new customers filling the 110 MW.
- Nscale. Whether the Anthropic talks produce an announced agreement, how Microsoft responds on Loughton, and how the IPO prospectus describes power risk.
- Amazon's response. Any further detail from AWS on staged hardware and the timing of energization.
- NVIDIA GTC Berlin, today. European AI-factory and sovereign-cloud announcements, which will test claims about where capacity is actually available.
- Grid and regulatory moves. Reported FERC directives on large-load connections, and any UK response to the London outage from financial regulators.
In cloud computing, energized and reliable capacity is the scarce asset. Announcements, orders and hardware purchases are only the early stages.
For cloud providers: Time to power and time to revenue are competitive metrics. Capacity already connected, as at OVHcloud, is a faster route to customers than capacity that is still waiting for the grid.
For enterprise buyers: Test resilience beyond multi-zone deployment, and ask suppliers for energization dates, not just allocation.
For investors: Separate backlog, hardware held, and capacity earning revenue. Treat unconfirmed deals and single-source reports as unresolved.
The CODEW Stat
1 of 3 zones, about 6 hours — Google's London outage began in one of three zones in the region, and most services were restored about six hours later. That is the measure of the lesson: a single power event, in a cloud designed with three zones, still disrupted banks, delivery firms and property sites. It is a prompt to ask which of a company's dependencies are truly independent, not a verdict on any one provider before Google publishes its root cause.
Sources: Reuters, Bloomberg (as relayed), AP, The Register, ABC News, NZ Herald, CX Today, Business Insider (as relayed by Investing.com and Yahoo Finance), Axios and EconoTimes, covering October 18 to 21, 2026. Where a story is relayed through a secondary outlet, the item says so.
Confirmed incidents (Google), CEO statements (OVHcloud), reported terminations (Microsoft and Nscale), unconfirmed negotiations (Anthropic and Nscale), and internal-email reporting (Amazon) are labeled separately. Verify figures and dates against company statements before publication. This series does not provide investment advice.
Reviewed by Erwin Castro
on
Wednesday, October 21, 2026
Rating:
No comments: