Beta. Kit runs on your own Mac, in open beta. Try the beta.
← All notes

The day the inbox went dark

Listen · narrated by Kit 8 min

What happened

On Thursday afternoon, millions of people opened Outlook and found nothing there.

Microsoft 365 went down across at least 74 countries. Exchange Online, Teams, SharePoint, Copilot, OneDrive and Defender all struggled at once, and for many organisations it took the best part of a day to come back.1 Microsoft says it was not an attack. By the company's own account, the trouble began with a network capacity upgrade and left “a large number of servers” unavailable.24 Something routine, in other words, done by people who are good at their jobs, went wrong in one place and was felt everywhere.

What it touched is the interesting part.

  • Workhorse Group, a listed electric truck maker, moved its quarterly results from 12 to 16 November, citing business disruption.2
  • Richemont's earnings were delayed.2
  • Air India had to reroute its email so customers could still reach it.2
  • In Colorado, a sheriff's office told people that anyone needing to post bond for immediate release would have to do it in person.3
  • In Los Angeles, the county medical examiner and sheriff's department lost communications.3 In the UK, NHS services were affected and one hospital group warned of long waits on its phone lines.2
  • In Australia, the federal police and the intelligence services were hit. In the Philippines, the Supreme Court's communications went down.1 US federal agencies and Capitol Hill lost email.3

Nobody has a final bill yet, because at the time of writing Microsoft is still restoring service. The first serious estimate, from the cyber-risk firm Kovrr, puts lost productivity at about $16 billion, with insured losses of around $700 million.5 Kovrr's co-founder called it a back-of-the-envelope figure, and it assumed the outage would be over by Friday evening, so it may well rise. Read those two numbers together: if they are roughly right, more than 95% of the damage falls on the businesses themselves. Service credits under Microsoft's service level agreements, as one security adviser put it, “typically represent a fraction of actual business losses”.6 At the small end, one Queensland shower-screen business reckoned the lost day would cost it at least $30,000.7

None of these organisations did anything wrong on Thursday. They made a reasonable decision years ago, the same one almost everyone made, and on Thursday they found out what it cost.

James Owen at Control Risks called it a “resilience and dependency issue that has been decades in the making”.2 That is exactly right, and it is about to get much bigger.

Email was the easy part

Losing email for a day is painful, but email is mostly a pipe. When it comes back, the messages are still there.

The next wave of dependency is not a pipe. It is memory.

Companies are now handing their assistants and agents the context that makes them useful: the decisions, the customer history, the reasons behind the pricing, the way this team actually ships, the thing the founder said in March that everyone still works from. Increasingly that context lives inside someone else's product, indexed by someone else's system, shaped to fit someone else's model.

Imagine Thursday again, but this time it is not the inbox that goes dark. It is the layer that remembers how your company thinks. Your agents stop. Your new hire's assistant forgets everything it knew about the account they took over. And when it comes back, you are reminded that the thing you most depend on is something you never actually held.

Sovereignty means different things to different people

For an individual, sovereignty is mostly about privacy. Your memory, the record of what you have thought about, who you have spoken to, what you are worried about, should belong to you. You should be able to read it, carry it to another tool, and delete it, without asking anyone.

For a company, it is about four harder things.

Continuity. Can you keep working when a provider you do not control has a bad day? Thursday answered that for millions of organisations. A memory that runs on your own machines keeps answering when the cloud does not.

Risk. Every critical system held by a single vendor is a concentration risk, whether or not it appears on a risk register. Outages are one form of it. Price changes, product shutdowns, terms that shift under you and acquisitions are others.

Compliance. Regulators increasingly want to know where data lives, who can see it and how it is deleted. “It is somewhere in our vendor's index” is a weaker answer every year, especially in Europe, where digital sovereignty has moved from policy papers to procurement.

IP, or alpha. This is the one people underestimate. What makes a firm good is rarely in its documents. It is in the accumulated judgement: why the last deal was structured the way it was, which supplier quietly can't be trusted, the playbook that only exists in three people's heads. Once that is captured as memory for AI, it is the most valuable dataset the company owns. It should not be training material, a negotiating lever or a single point of failure in someone else's data centre.

What running locally actually looks like

This is why Kit is built the way it is.

Kit's memory runs on your own machine or your own servers. The store is a local database. The embeddings, which are what make semantic search work, are generated locally, so finding a memory needs no network call at all. Retrieval involves no outside model and no outside service. And everything can be exported as plain, readable text, so if Kit itself disappeared tomorrow, your memory would not.

You can go further. Kit can run its own thinking on local models too, so a company that wants to can operate with no cloud provider in the loop at all, model or otherwise.

I want to be honest about the line here, because it matters. If you choose to use a frontier model from the cloud, which most people will for much of their work, then whatever memories that model is shown in a conversation go to that provider for that conversation. That is how these models work, and anyone who tells you otherwise is selling something. Our terms list exactly what leaves your Mac. What Kit gives you is storage sovereignty and inference choice: the memory is yours and stays with you, and you decide, task by task, which model gets to see which part of it. When the cloud is up, you can use the best model in the world. When it is down, or when the material is too sensitive, you still have your memory, and you can run locally.

That is the shape I think every serious organisation will want: rent the compute, own the memory.

The lesson of Thursday

The cloud is not going away, and it should not. It is brilliant at most of what it does, and Thursday was a bad day, not a verdict.

And for all its cost, Thursday was the recoverable version. Email came back. Results were moved by four days. People posted bond in person.

The expensive version of this lesson is the one where the thing that goes dark is your company's memory, and you discover it was never really yours. That one is still avoidable. It just has to be decided now, while the memory is still being built, rather than after it is somewhere you cannot reach.

Kit for companies. Try the beta.