Claude and Grok Went Down Together: The Shared Compute Outage

Three of the four frontier providers degraded within the same two-hour window on the morning of September 3, and the fallout exposes a risk heavy AI users have been underestimating: the model you pay for and the model you think of as its rival can run on the same physical compute. The outage that took down Claude and Grok together traced to a single SpaceXAI data center in Memphis, the same center Anthropic’s Claude models lease capacity from under a deal signed earlier this year. One facility acted as a single point of failure across two of the loudest rivals in the coding-agent race.

What actually happened on September 3

The incident began around 6:30 AM Pacific Time. SpaceXAI reported a problem at its Memphis compute center, and Grok went into a models outage that lasted about three and a half hours, according to the xAI status page. The disruption hit Grok on the web, on X, and in the Android and iOS apps. SpaceXAI apologized publicly to Grok users and, notably, to unnamed “impacted compute partners,” a phrasing confirming the failure was not confined to xAI’s own products.

Anthropic’s Claude status page told the rest of the story. In roughly the same 6:30 AM window, Claude logged elevated errors across multiple models, explicitly including Fable 5.1 and Opus 5, the two models most heavy users run for agentic work. The incident was resolved by around 9:16 AM Pacific Time. Anthropic did not confirm a connection to the SpaceXAI facility, but the timing overlap and the apology to “compute partners” left little room for coincidence given that Anthropic leases capacity from SpaceXAI.

OpenAI added a third data point to the same morning. Users reported trouble with ChatGPT and Codex starting around 7:30 AM Pacific Time, and OpenAI acknowledged elevated errors across both before marking the issue resolved by about 9:55 AM. OpenAI did not tie its problem to SpaceXAI, and it may well have been unrelated. The uncomfortable part is how little that distinction matters operationally: regardless of root cause, a heavy user on September 3 could not rely on Claude, Grok, or Codex at the moment they needed to ship.

The real story: competing models, one shared foundation

The headline is not that an outage lasted two or three hours. It is that frontier AI now concentrates correlated risk inside a few enormous data centers, and the commercial rivalry between Anthropic and xAI does nothing to separate their underlying compute. Claude is Anthropic’s product, but part of the capacity serving Claude runs on hardware that also serves Grok. When that hardware fails, both “competitors” fail at the same time.

That changes the meaning of a redundancy strategy. If your plan is to keep a Claude account and a Grok account so one is always up, this week proved that plan has a hole: both can be down for the same reason in the same hour. Provider diversity is only as strong as infrastructure diversity. Two providers sharing one compute center is not redundancy, it is a single point of failure wearing two logos.

OpenAI’s morning reinforces the point from the other side. Even when outages are independent, the industry now has enough correlated stress, shared hardware, and overlapping traffic surges that simultaneous incidents are becoming normal rather than a freak coincidence. Claude users who hit errors scramble toward Codex and Gemini, and that surge stresses the very services they run as backup.

What an outage actually costs a heavy user

For someone paying $300 or more a month and running agents in production, a three-hour morning outage is a measurable cost event. Agent runs mid-task when the errors start die without completing, and Claude Code sessions holding a long, expensive cached context can lose that warm cache depending on how the session terminates. Scheduled batch work silently fails. And because incident windows consume part of your reservation, a Monday-morning outage can trigger a secondary problem days later when you hit a usage cap on a week with less usable uptime than planned.

There is also an attention cost. In the HN threads that filled during the outage, users described the same behavior: the moment Claude dropped, they went looking at Codex plans, Sonnet alternatives, Grok and Gemini as fallbacks, and at switching costs. One commenter summed up the arithmetic heavy users are arriving at: outages are now frequent enough that relationship durability matters more than the quality delta between models.

The reliability playbook for heavy AI users

You cannot make SpaceXAI’s Memphis center more reliable from your laptop, but you can make your own stack resilient to this kind of correlated failure, and most of the moves cost nothing until you need them.

Assume a provider can be down at the same time as its apparent rival. Audit which of your providers run on shared infrastructure before trusting them as mutual fallbacks. If your two “backup” accounts are Anthropic and xAI, you have one backup, not two. A frontier API plus an open-weight deployment on different hardware gives you the independence you think you are paying for.

Make failover boring. The engineers who sailed through September 3 were the ones who already had a second provider wired in with the same skills, hooks, and model-agnostic scaffolding. If your workflow is tightly coupled to one vendor’s tooling, failover is a rebuild project, not a switch flip. Keep agent scaffolding generic enough that pointing it at another provider is a config change, not a rewrite.

Protect your warm context. A long agent session with cached tokens is your most expensive asset to lose to an outage. Checkpoint state and context outside the vendor session so an interruption costs you rebuilding time, not the entire reasoning context, and instrument runs so you can tell which tasks died mid-flight.

Track uptime the way you track cost. If you monitor spend per provider, add an uptime and incident column to the same report. It is the fastest way to see that a few minutes of downtime each week is not free: it costs completed tasks, burnt reservations, and engineer attention. Cheap per token is expensive per completed week when a provider drops in your highest-traffic window.

The net for heavy AI users

The September 3 outage is a durable signal, not a one-off. A single SpaceXAI data center in Memphis took Grok down for three and a half hours, took Claude down at the same time through shared capacity, and coincided with an OpenAI disruption on top of it. The industry’s loudest rivalry runs on some of the same physical silicon, and that correlation is now the single biggest reliability risk in the stack of anyone paying a serious monthly bill.

The lesson is not to abandon frontier providers. The lesson is to stop mistaking provider diversity for infrastructure diversity, and to apply the same rigor to uptime that you apply to your token budget. Keep an independent fallback genuinely warm, protect your context, and treat every frontier model as a service that can vanish for a morning. The next Memphis-scale incident is a when, not an if.