The Common-Mode AI Failure

The Evidence Chain, Part 5 of 10. The first four parts looked at one AI system and the person it acts for. This part zooms out to a question you can ask about your own working day: how many of the AI tools you rely on are really one system?

At 11:05 UTC on 18 November 2025, Cloudflare deployed a routine change to the permissions on one of its databases. Fifteen minutes later its network began to fail: a file the network uses to spot automated traffic had doubled in size, crossed a hard-coded limit, and started crashing the software that handles traffic for the websites and apps that sit behind Cloudflare. Cloudflare’s chief executive called it the company’s worst outage since 2019 (Prince, 2025). For a few hours, people trying to reach ChatGPT, Claude, X and Spotify got errors (Franceschi-Bicchierai, 2025). Two rival AI companies, built by different teams on different models with different safety philosophies, went dark together because of a database setting at a third company most of their users had never heard of.

Engineers call this a common-mode failure: many things that look separate fail together because they share something underneath. I think AI is building the largest common-mode surface we have ever had, and that most organizations, and most people, cannot see it, because it hides behind different brand names. You might use a writing assistant, a coding tool, a customer-service bot and a meeting summarizer from four vendors. Underneath, they may call the same two or three foundation models, run in the same cloud region, sit behind the same network provider and authenticate through the same identity service. Four logos. One point of failure.

How concentrated it already is

Start with the models.

Menlo Ventures, a venture firm that is itself an investor in Anthropic (Menlo Ventures, n.d.), surveyed about five hundred US enterprise decision-makers in November 2025 and estimated that three providers, Anthropic, OpenAI and Google, accounted for roughly 88 percent of enterprise use of large language models (Tully et al., 2025). Given that stake, treat the exact shares with care. The shape, a handful of providers carrying most of the load, is not in dispute.

Then look at the layers below. The Financial Stability Board, which coordinates financial regulators across the G20, reported in October 2025 that concentration exists in the hardware, cloud, training-data and model markets at once, that dependencies appear more prevalent in generative AI than in other kinds of AI, and that this can reduce firms’ ability to cope with disruptions at their suppliers’ suppliers (Financial Stability Board [FSB], 2025).

The events bear this out. On 19 and 20 October 2025, a race condition in the automation that manages DNS records for Amazon’s DynamoDB database in its Northern Virginia region left an empty record where an address should have been, and the disruption rippled through dozens of AWS services for most of a day (Amazon Web Services [AWS], 2025). Perplexity, Signal, Zoom, Duolingo, Roblox, Canva and Snapchat were among the services reported affected (Al Jazeera, 2025; Whittaker & Perez, 2025). In Nigeria, users of the investment app Bamboo could not see their US stock positions, and users of the payments app Raenest could not see their card details (Abiodun, 2025). Nobody in Lagos chose Northern Virginia. It was chosen for them, several contracts up the chain.

The supervisors have noticed, at one layer

Regulators have started to act, and where they have acted tells you something. In November 2025, the three European financial supervisory authorities designated nineteen “critical ICT third-party providers” for direct oversight under the Digital Operational Resilience Act, a list that includes the major cloud providers (European Supervisory Authorities, 2025). In July 2026, the UK designated its first four critical third parties for financial-sector oversight: Amazon Web Services, Google Cloud, Microsoft and Oracle (Bank of England et al., 2026). In June 2026, the European Systemic Risk Board warned that frontier AI models could collapse the time defenders have to patch vulnerabilities, and named concentration, meaning dependence on a small number of AI providers who depend in turn on cloud providers and widely used software, as one of four areas of systemic risk (European Systemic Risk Board [ESRB], 2026).

Look at those lists and notice what is missing. The designated providers are clouds, data centers and IT service firms. As of this writing, no foundation-model provider is on either list. The supervisors are watching the floor the building stands on, and not yet the shared brain that more and more of the building thinks with.

Down together, and wrong together

Outages are the visible half of the problem. The other half is quieter. In September 2025, Anthropic published a postmortem explaining that three separate infrastructure bugs had degraded some of Claude’s responses over several weeks, including a routing error that at its peak affected 16 percent of requests to one model, and that the degradation reached users through Anthropic’s own API and through Amazon’s and Google’s cloud platforms alike (McAllister, 2025). Nothing went down. Answers simply got worse, everywhere that model was used, at the same time. That is common-mode failure without an outage, and it is the bridge to the next part of this series, which asks what happens when systems that keep working all make the same mistake.

Try it on your own tools

Here is an exercise that takes about twenty minutes and will change how you think about your own setup. Write down every AI tool you used in the past week, at work and outside it. Next to each, write the company whose model it runs on, if you can find out. Most tools say so somewhere, in a help page, a privacy policy or a list of subprocessors; some do not say at all, which is itself an answer. Then write down where it is hosted, if that is disclosed. Given the market shares above, do not be surprised if a list that looks like nine independent products collapses into two or three model providers and even fewer clouds.

The point is not that you should panic or switch everything. The point is that the picture of independence most of us carry around is an illusion created by branding. Your study app, your company’s support bot and your bank’s fraud assistant may each feel like a separate decision someone made. Underneath, many of them are the same decision, made once, by someone else.

Who decides, who can intervene, who pays

The power question is simple to state.

The people who decide when a model changes, when a configuration ships and which region carries the load are a small number of engineers at a small number of companies. The organizations that build on them can usually intervene only by waiting. And the consequences land on everyone downstream: the clinic that loses its scheduling assistant, the student whose study tool goes blank the night before an exam, the small business in Nairobi or Manila whose customer service runs through an API in Virginia. Resilience planning that counts vendors instead of dependencies is counting the wrong thing, and it hands the real decision to people who will never meet the ones who carry the cost.

The prediction

My prediction is that AI dependency will become a mainstream operational-resilience issue, first in finance, where supervisors already have the tools, and then in any sector where an outage or a correlated error causes real harm, and that the foundation-model layer will eventually be supervised the way cloud is now. The research question for this part is practical, and you can ask it of your own organization on Monday: what is the actual dependency graph behind the AI systems you think of as independent?

For Mercury Intelligence members: the section below walks the Cloudflare outage through the dependency graph, maps the rules clause by clause, sets out a quarter-long method for finding your shared nodes, and gives the strongest case for concentration a fair hearing.

---

Support independent investigation. Mapping the dependencies under the AI economy means reading post-incident reports that the companies involved would rather you skimmed, and asking supervisors why the model layer is still missing from their lists. That work only stays independent if readers fund it. Give once ($5, $10, $25, $50 or any amount), give monthly ($10 or $25), or become a patron: fairbydesign.org/patrons-2.

References

Abiodun, B. (2025, October 20). AWS outage disrupt Bamboo, Raenest services. Techpoint Africa. https://techpoint.africa/news/aws-outage-crash-snapchat/

Al Jazeera. (2025, October 20). Amazon Web Services largely recovered from chaotic global web outages. https://www.aljazeera.com/news/2025/10/20/amazon-cloud-problems-spur-outage-of-global-websites-and-apps

Amazon Web Services. (2025). Summary of the Amazon DynamoDB service disruption in the Northern Virginia (US-EAST-1) region. https://aws.amazon.com/message/101925/

Bank of England, Prudential Regulation Authority, Financial Conduct Authority, & HM Treasury. (2026, July 10). UK financial regulators to begin overseeing critical third parties announced by Treasury [Statement]. https://www.fca.org.uk/news/statements/uk-financial-regulators-overseeing-critical-third-parties-announced-treasury

Directive (EU) 2022/2555 of the European Parliament and of the Council of 14 December 2022 on measures for a high common level of cybersecurity across the Union (NIS 2 Directive). (2022). Official Journal of the European Union, L 333, 80–152. http://data.europa.eu/eli/dir/2022/2555/oj

European Commission. (2026). AI Act: Regulatory framework for artificial intelligence [Policy page, updated August 3, 2026]. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

European Supervisory Authorities. (2025, November 18). European Supervisory Authorities designate critical ICT third-party providers under the Digital Operational Resilience Act [Press release]. European Securities and Markets Authority. https://www.esma.europa.eu/press-news/esma-news/european-supervisory-authorities-designate-critical-ict-third-party-providers

European Systemic Risk Board. (2026). Warning of the European Systemic Risk Board of 25 June 2026 on systemic cyber risks stemming from frontier artificial intelligence models (ESRB/2026/3). https://www.esrb.europa.eu/pub/pdf/warnings/esrb.warning260625_on_systemic_cyber_risks_stemming_from_frontier_ai_models~ef424708cf.en.pdf

Financial Stability Board. (2025, October 10). Monitoring adoption of artificial intelligence and related vulnerabilities in the financial sector. https://www.fsb.org/uploads/P101025.pdf

Franceschi-Bicchierai, L. (2025, November 18). Cloudflare blames massive internet outage on “latent bug.” TechCrunch. https://techcrunch.com/2025/11/18/cloudflare-blames-massive-internet-outage-on-latent-bug/

Knecht, D. (2025, December 19). Code Orange: Fail small — our resilience plan following recent incidents. The Cloudflare Blog. https://blog.cloudflare.com/fail-small-resilience-plan/

Knight, J. C., & Leveson, N. G. (1986). An experimental evaluation of the assumption of independence in multiversion programming. IEEE Transactions on Software Engineering, SE-12(1), 96–109. https://doi.org/10.1109/TSE.1986.6312924

McAllister, S. (2025, September 17). A postmortem of three recent issues. Anthropic Engineering. https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues

Menlo Ventures. (n.d.). Anthropic [Portfolio company page]. https://menlovc.com/portfolio/

Microsoft Azure. (2026). Mitigated – Intermittent request failures and increased latency across Azure OpenAI, Azure AI Foundry, and Cognitive Services [Post incident review 7QL5-Z50, September 29, 2026]. Azure status history. https://azure.status.microsoft/en-us/status/history/

OpenAI. (2026, April 20). Users unable to load ChatGPT, Codex and API Platform [Incident write-up]. OpenAI Status. https://status.openai.com/incidents/01KPNN2V2SMP3TAN3MCJK87W50/write-up

Parham, A. (2026, September 3). Gemini survived when ChatGPT, Claude, Grok collapsed: Azure is at fault. Tech Times. https://www.techtimes.com/articles/326509/20260903/gemini-survived-when-chatgpt-claude-grok-collapsed-azure-fault.htm

Prince, M. (2025, November 18). Cloudflare outage on November 18, 2025. The Cloudflare Blog. https://blog.cloudflare.com/18-november-2025-outage/

Regulation (EU) 2022/2554 of the European Parliament and of the Council of 14 December 2022 on digital operational resilience for the financial sector. (2022). Official Journal of the European Union, L 333, 1–79. http://data.europa.eu/eli/reg/2022/2554/oj

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). (2024). Official Journal of the European Union, L, 2024/1689. http://data.europa.eu/eli/reg/2024/1689/oj

Tully, T., Redfern, J., Das, D., & Xiao, D. (2025, December 9). 2025: The state of generative AI in the enterprise. Menlo Ventures. https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/

Whittaker, Z., & Perez, S. (2025, October 21). Amazon identifies the issue that broke much of the internet, says AWS is back to normal. TechCrunch. https://techcrunch.com/2025/10/21/amazon-dns-outage-breaks-much-of-the-internet/

Designing for failure: Chapters 4 and 7 of Operational AI Readiness (free).