“Put all your eggs in the one basket and watch that basket.”
Mark Twain, Pudd’nhead Wilson
You don’t have three AI vendors. You have three front doors, and nobody on your team has seen the building behind them.
THE SIGNAL
On the morning of September 3, Grok, Claude and ChatGPT all failed within about 90 minutes of each other.
Grok went dark at 6:30 AM Pacific and stayed down past 10.
Claude ran a partial outage for just over three hours, across the app, the API and Claude Code.
OpenAI reported elevated errors on ChatGPT and Codex starting around 7:30.
SpaceXAI blamed its Memphis compute center and apologized to “impacted compute partners.” It never named them. Anthropic agreed in May to take the full output of SpaceX’s Colossus 1 facility in Memphis. It called its own problem an infrastructure issue and has said nothing more. OpenAI pointed to routing. Nobody has confirmed a shared cause.
Two weeks before the outage, according to The Commercial Appeal, SpaceXAI told a court that losing power at its Memphis plant would do catastrophic damage. It said Grok would largely stop working, along with other AI services.
The headline was “AI goes down.” The signal was different. A warning about a single point of failure sat in a public court record, and almost nobody buying these tools read it.
THE FAILURE POINT
The break didn’t happen at 6:30 AM. It happened the day a company added a second AI provider to its risk register and wrote “redundant” next to it.
That one word turned a physical question into a vendor count. The physical question was where this compute runs and whose power keeps it on. On September 3, the count said three. For most of that morning, the working number was closer to zero.
SIGNAL WITHIN THE SIGNAL
Follow the truth as it moved.
In court, it was specific: one plant, one power supply, catastrophic if lost. In the apology, it shrank to “compute partners,” with no names. On the procurement sheet, it shrank again to a logo and a contract term. By the time it reached the executive approving the AI rollout, a dependency had become a diversification story.
Nobody lied at any layer. Each layer just removed the part that would have forced a decision. That’s Signal Compression with a power cord attached.
Worth checking where your own team sits on that same gap before assuming it's someone else's problem.
Score your gap →
BEHAVIOR UNDER PRESSURE
When the tools went down, teams did what the plan told them. They switched providers. Then the second one threw errors too.
Feeds filled with screenshots of error messages and guesses about a shared cloud. Speculation filled the space a dependency map should have filled.
Most leaders probably treated it like weather. Wait three hours and it comes back. It did come back, and that’s the danger. An outage that heals itself teaches an organization nothing unless someone decides it should.
SYSTEM DRIVER - MOS
The system is producing false redundancy because the vendor register tracks the wrong unit. Fix the unit.
Every critical AI dependency gets one added field: the facility and power source behind your workloads, and who else runs there. Then add one rule. Two providers that share a facility or a power feed count as one provider. Unknown counts as shared.
That rule changes what “redundant” means in every review that follows. A backup has to survive the same bad morning as the primary, or it isn’t a backup.
LEADER DRIVER - INTERNAL OPERATING SYSTEM (IOS) - REGULATE
The leader needed a lower threshold for coincidence. Three separate tools failing inside 90 minutes is information about structure. The nervous system files it as a nuisance and waits it out.
The override is one question, asked before anyone says “it’s back”: what do these systems share? The read here suggests most leaders never asked. The outage ended before the discomfort had time to do its job.
IF YOU DO ONE THING TODAY
Pull up the AI tools your operation can’t run a shift without.
That might be scheduling, code, quoting or customer response.
Next to each one, write the facility and power provider behind it.
Where you can’t, email the account rep before lunch with one question: which data centers serve our workloads, and which other AI companies run in them? Every blank you can’t fill is a dependency you’ve been calling a backup.
FINAL SIGNAL
Redundancy you can’t locate is a rumor.
CTA
Send MOSei.org to the one leader who needs it before Friday.
SOURCES
SpaceXAI statement on X (Sept. 3, 2026) · Anthropic status page (Sept. 3, 2026) · Engadget · Cloud Security Alliance research note (Sept. 9, 2026) · The Commercial Appeal (Aug. 19, 2026) · China Daily / Xinhua (Sept. 4, 2026)