13 November 2026
Microsoft and Claude down: my local AI didn't even notice
Microsoft down globally, Claude kicking users out and NVIDIA without a CEO within 48 hours. Meanwhile, an on-premise local AI kept serving clients without noticing.
This week was a fast-track reminder of how fragile our “trusted” digital infrastructure really is.
48 hours of shaky cloud
On Thursday, November 12, Microsoft suffered a global outage: Outlook, Teams, SharePoint and Copilot down for users in every region. No minor scare: we are talking about the email, calendar and collaboration stack of millions of companies. Hours later, live coverage still described the fixes with a disheartening phrase: not going as expected, with no estimated time to resolution (Microsoft’s own incident EX1550543).
On Friday the 13th, the domino chain continued: Claude, Anthropic’s AI platform, piled up incidents — first trouble with its Microsoft 365 connectors (mail, calendar, search), then a bigger outage caused by authentication errors that threw users out of their claude.ai and Claude Code sessions. People who paid for an “always available” service, locked out.
And underneath it all, the whole industry was asking bigger questions: Jensen Huang, the face of NVIDIA and of the AI boom, resigned as CEO on Thursday, effective immediately. The media cover his personal life; the deeper reading is unavoidable: his bet — affordable AI hardware and open weights — never sat well with the closed, pay-per-token AI camp, in the middle of the US-China tech war. Whatever the final reason, the message is the same: the board is being reshuffled, and the pieces your business is built on move without notice.
Meanwhile, in an office in Barcelona
Oblivious to all of it, on Thursday and Friday I kept working normally for my clients. My email lives on Google Workspace, like almost everyone’s — but documents, private RAG over each client’s data, generation and analysis run on my own hardware, open weights. No queues, no “we are investigating the incident”, no waiting for a team in Redmond or San Francisco to restore my business.
It is not luck: it is architecture. When the critical parts of your work do not depend on an external provider:
- No foreign ETA: the incident is yours, you diagnose it and you fix it.
- No session evictions: your access to your data does not expire because someone else’s login broke.
- No price hikes or terms changes while you sleep.
- Your data stays put: privacy by design, not by privacy policy.
To be fair: local breaks too. A disk can die, the power can go — and electrical and backup redundancy are part of the design. The difference is that a local failure is small, visible and yours; a cloud failure is global, opaque and nobody’s.
The cloud is not the enemy (but it must not be the only foundation)
I am not against the cloud: I use it, and it makes sense for many things. The question is different: which part of your business can afford to stop when — not if — the provider goes down? This week we saw it with Microsoft and Claude. And earlier we documented a very clear case: Revolut, the bank serving millions of customers, dropping Claude and GPT and running its inference on open-weight models (we covered it here).
If a bank can do it, an SME with bounded critical needs can do it much sooner, and without an infrastructure team: a modest on-premise setup covers private RAG, assistants and automation with an investment that pays for itself.
If you want to know where to start in your case, tell me about your situation and I will tell you exactly how I would build it.
Sources: coverage of the Microsoft outage 2026-11-12 (BBC, TechRadar — incident EX1550543), Claude status page (status.claude.com), official NVIDIA announcement (nvidianews.nvidia.com).
Free assessment
Local, private AI for businesses, freelancers and organisations. Your data never leaves your office.