ChatGPT, Grok and Claude went down together: the outage that exposed AI's worst-kept secret

WHAT HAPPENED. On September 3, 2026, at around 6:30 a.m. Pacific Time, Claude and Grok began failing almost in unison. Roughly an hour later it was ChatGPT's turn: its status page acknowledged “elevated errors across ChatGPT and Codex,” as The Verge reported. The numbers convey the scale: over 35,000 outage reports on Downdetector for ChatGPT at the peak, versus about 1,400 for Claude and 1,200 for Grok, according to analyst Wisely Chen's compilation. Downdetector is a portal that tracks incidents from user complaints; it isn't official data, but it is a useful thermometer. The durations weren't trivial either: Grok was down for about three and a half hours and Claude for around three, per Engadget; OpenAI marked its incident resolved by about 9:55 a.m. PT. There were scattered complaints about Gemini, but Google never confirmed any problem.

THE MEMPHIS CLUE. The most concrete explanation came from xAI, now folded into SpaceX and operating under the SpaceXAI brand. The company apologized for “an outage at our Memphis compute center” and added a revealing line: it also apologized “to our impacted compute partners.” Memphis is home to Colossus, the supercomputing complex xAI built at record speed starting in 2024. And here is the key: since May 2026, Anthropic — the maker of Claude and a direct rival of xAI — has leased Colossus 1 on a near-exclusive basis. According to TechCrunch and Bloomberg, the contract runs to roughly $1.25 billion per month through May 2029 (about $45 billion in total) for more than 220,000 NVIDIA GPUs and 300 megawatts of power. A GPU is the specialized chip that runs AI models; 300 megawatts is the consumption of a mid-sized city. Translation: when Memphis coughs, Grok and Claude catch a cold at the same time, however hard their logos compete on everything else.

AND CHATGPT? CAUTION REQUIRED. OpenAI attributed its outage to a “routing error” — a failure in directing user requests to the right place within its network. No one has demonstrated any connection to Memphis. Two alternative theories circulated: a failure in Microsoft Azure's East US cloud region, which no party officially confirmed, and a Cloudflare fault, which Cloudflare itself denied, stating its services ran normally the entire time. There is a third hypothesis, suggested by Wisely Chen, which we find plausible but unproven: users pushed off Claude and Grok migrated en masse to ChatGPT, and that demand wave helped topple it an hour later. That would be a demand cascade, not an infrastructure one. The honest framing is this: two outages with a very likely shared physical cause, and a third that may have been coincidence or a domino effect.

OUR READ. This incident matters less for the three-hour blackout than for what it reveals: in frontier AI, three brands do not equal three infrastructures. The scarcity of chips and electrical power is so acute that even the fiercest rivals end up renting capacity to one another: SpaceX hands its crown jewel to Anthropic because, per Bloomberg, its own teams struggled with latency issues in using it, and Anthropic pays because it needs capacity now, not in two years. It is the same dynamic we have flagged with NVIDIA's cross-investments in the big labs: compute is being paid for with equity stakes, alliances and leases that weave an extremely dense web of mutual dependencies. Chen sums up the technical consequence well: failure domains — the set of systems that go down together when something breaks — follow physical location, not vendor logos. A company that contracts Claude and Grok “for diversification” may be buying two doors into the same building. Apparent redundancy is not real redundancy.

THE PRECEDENT THAT MATTERS. None of this is new to the internet: the October 2025 outage of Amazon Web Services' us-east-1 region dragged down thousands of services, and Cloudflare's November 2025 failure knocked out X and ChatGPT itself for hours. What's new is that AI adds an extra layer of concentration on top of the previous ones. Running frontier models requires gigantic, wildly expensive, energy-devouring campuses; there are very few of them, and ever more players depend on the same ones. The usage context has changed too: in 2024, a ChatGPT outage was an annoyance; in 2026, with agents like Claude Code or Codex embedded in the daily work of millions, a three-hour blackout means measurable productivity loss across the economy. AI is taking on the role of critical infrastructure without the reliability and transparency obligations we demand from electricity, banking or telecoms.

IMPLICATIONS. For companies, the lesson is immediate: a multi-model strategy only protects you if it is also multi-location, and today it is nearly impossible to know which data center runs the inference you buy — the information exists, but it is not published systematically. Asking should become part of every serious contract. For regulators, the sensible path is not panic but mandatory dependency transparency, in the spirit of what Europe's DORA regulation already requires of banks regarding their technology providers: map who depends on which building. And for the industry, reliability is about to become a sales argument: “where does your model run, and what happens if that site goes down?” will soon be as routine a question as the price per million tokens.

THE LONG VIEW, WITHOUT DRAMA. It helps to remember how infrastructures mature. The electrical grid of the early 20th century failed constantly; decades of engineering, standards and regulation made it boring — the highest compliment an infrastructure can earn. AI is on the same path: dozens of campuses are under construction, geographic redundancy will come because customers will demand it, and every blackout like September 3's accelerates that demand. In the short term, the fragility is real and must be governed with data, not narratives. In the long term, this episode is exactly the kind of failure an industrializing sector learns from. The day a ChatGPT outage is as routine a news item as a neighborhood power cut, AI will have completed its transition into infrastructure. That day is closer than it looks — and arriving there in good shape, with transparency, redundancy and clear rules, is the precondition for everything else we hope this technology delivers.