ChatGPT, Claude and Grok went dark together, right as OpenAI teased ‘the stars are almost aligned’

New Delhi: Thursday night, 8.23 pm. You type a question into ChatGPT. The cursor just sits there.

Refresh. An error screen. You switch to Claude. Same wall.

Grok? Stuck too.

For millions of users in India and across the world, the three most powerful AI assistants on earth went quiet at almost the same minute. Nobody had to tell them it was happening. They all felt it together.

That timing is the story. ChatGPT belongs to OpenAI. Claude belongs to Anthropic. Grok belongs to Elon Musk’s xAI.

Three rival companies. Three rival engineering teams. Three rival data centres. Their chatbots do not normally collapse on the same schedule.

When they do, the question stops being “what broke” and becomes “what do they all share?”

The night the prompts stopped

The first widespread complaints rolled in around 7.57 am US Pacific time on Thursday, September 3. That put it close to 8.27 pm in India, per Downdetector data reported by 9to5Mac and India Today.

Reports rose within minutes, not gradually. Users on the web, the mobile apps and the desktop clients all hit the same failures at roughly the same moment.

Login errors. Hanging prompts. Gateway timeouts.

OpenAI’s ChatGPT was hit hardest. Downdetector logged more than 35,000 problem reports from the United States alone for ChatGPT. Claude drew around 1,400 and Grok over 1,200, according to figures cited by India Today.

PCMag put the ChatGPT peak even higher, above 36,000 reports. Google’s Gemini saw a smaller spike but never formally declared an outage.

The coding assistant Cursor, which many developers use daily, also wobbled. It depends on Claude and Grok underneath.

India felt it too. By 8.23 pm IST, more than 2,800 ChatGPT complaints, about 170 for Claude and over 70 for Grok had already been filed on Downdetector from India, India Today reported.

The Hindu’s PTI report noted outage complaints began surging across Indian evening hours. Users described everything from websites refusing to open to internal server errors mid-conversation.

What each company admitted

The companies did not deny it. OpenAI’s status page listed an active incident titled “Elevated errors across ChatGPT and Codex.”

It said engineers had applied a mitigation and were monitoring recovery. Anthropic’s status page showed elevated errors across a long list of Claude models.

That list included Mythos 5.1, Fable 5.1, Opus 5, Opus 4.8 and Opus 4.6. Partial problems hit claude.ai, the Claude API, Claude Code and Claude Cowork, Moneycontrol reported.

Anthropic later said it had identified the cause. In a subsequent update cited by The Hindu, only Opus 4.8 and Opus 5 were still degraded. The rest returned to baseline error rates.

xAI was quieter. Its status page flagged issues with Grok and said work was underway. It had not posted its own public update by the time many outlets published, Moneycontrol noted.

Services recovered in waves. OpenAI said a fix was restoring service from around 8.49 am PT. By Indian night, most platforms were largely back.

The strangest part? Three separate companies, which compete fiercely with one another, all ended up investigating at the same hour.

Three companies, one suspicion

Outage-watchers immediately floated a shared cause. As 9to5Google put it, ChatGPT, Claude and Grok failing at the same moment points to a shared infrastructure problem rather than three independent accidents.

Cloud researchers describe the pattern as a “shared control-plane failure.” That is the term TechTimes used in its coverage. When a routing or load-balancing layer that many services quietly depend on develops a fault, every service on that layer degrades at once. No matter how different the applications look on top.

The prime suspect is Microsoft Azure.

OpenAI runs most of its production workloads on Azure. Anthropic runs significant inference through Azure too. xAI’s cloud infrastructure also routes through Microsoft data centres.

StatusGator logged a user-submitted report of ingress failures in Azure’s East US region at 10.26 am PT on the same morning, TechTimes reported. Azure East US is one of the most heavily used compute regions for enterprise AI in the world.

Here is the cleanest evidence in favour of the Azure theory: Gemini stayed up.

Google’s assistant runs on Google Cloud. That has entirely separate routing and compute layers from Azure. Three Azure-hosted AI platforms failed almost simultaneously. The one major platform on Google’s cloud kept answering.

TechTimes called Gemini’s survival the key data point of the morning. Not a sign of superior product quality. A live demonstration of what infrastructure independence actually means.

The Cloudflare question

Azure is not the only suspect. Some early reports, including one flagged by aiweekly.co, suggested Cloudflare problems were affecting multiple websites that use its network. That could have compounded the failure.

Cloudflare is the internet’s most common shield. It protects everything from chatbots to transit systems. Its own worst outage in recent years, in November 2025, knocked out huge numbers of services at once after a bot-management software bug went wrong.

No company has publicly confirmed a root cause yet. OpenAI, Anthropic and xAI all said their engineers were investigating. As of midday US time Thursday, none had released a post-incident analysis.

That matters. Until one of them publishes a write-up, the Azure theory, the Cloudflare theory and the plain-coincidence theory all remain live.

When users flee, they flood the survivors

There may be a second mechanism at work underneath the first. When ChatGPT went dark first, millions of displaced users instantly opened Claude and Grok as alternatives.

That created demand spikes those platforms would never have seen on a normal evening. It happened exactly when their own infrastructure may already have been strained.

The same pattern was documented in June 2024. A long ChatGPT outage was followed by degraded performance at Claude and Perplexity. TechCrunch reported at the time that overflow traffic from ChatGPT’s failure was the likely cause.

In Thursday’s case, the overflow cascade and any underlying cloud fault may have reinforced each other. From outside, the two effects are almost impossible to separate. Neither company has shared enough data to untangle them.

What is certain is the outcome. For roughly half an hour, most of the world’s best-known AI could not answer a single reliable query. Unless you happened to be on Gemini.

The Astra theory is tempting. It is also unproven.

The outage landed on a day when OpenAI was teasing something big.

Its official ChatGPT account posted a video of drifting stars with the line “The stars are almost aligned.” Its main account shared a 12-second clip of an old MIT demo called “Put That There.”

OpenAI has confirmed a next-generation model, reportedly codenamed Astra, is in development. Speculation has swirled that it could mark a jump toward a GPT-6 generation.

The internet did the math out loud. Some users joked that ASTRA degrades letter by letter into RESET. Others compared the outage to the Apple Store going offline before an iPhone launch.

It is a wonderful meme. It is also almost certainly wrong.

Here is the counterargument, and it deserves equal weight. An outage like this costs OpenAI money, trust and headlines. It also took down its rivals.

If OpenAI were preparing an Astra launch, only OpenAI would have needed to go down. Not Anthropic and xAI.

A coordinated three-company blackout is not a launch stunt. It is a bad day for everyone involved.

The Economic Times reported the same caution. The outage and the Astra rumours remain two separate developments unless a company provides evidence linking them.

Even the most suspicious observer should note one thing. Anthropic said it identified its own cause. That is hard to square with an OpenAI publicity exercise.

The concentration that nobody voted for

Strip away the theories and one uncomfortable fact remains.

The frontier AI industry is not just concentrated at the model layer. It is embedded inside the hyperscaler cloud market underneath.

The Cloud Security Alliance’s June 2026 analysis of AI provider concentration risk warned that enterprises face correlated risk on two stacked levels. A handful of model companies, running on a handful of clouds, all exposed to the same failure domains.

Thursday was that warning in action. Three rival chatbots went down because, at the layer that actually moves the traffic, they may not be rivals at all. They are tenants of the same landlord.

For enterprise teams whose AI workflows went dark this morning, the lesson is structural. Thoughtworks made it bluntly after an earlier Claude outage. Treating a single provider’s API as always-on is a single point of failure and a genuine threat to business continuity.

The fix is not glamorous. Documented fallback routes, tested regularly, pointing at genuinely separate infrastructure. Running Claude through AWS Bedrock when the direct API fails. Keeping a Google Cloud path warm.

India’s exposure is the same story, louder

For Indian users, this outage was a reminder that the country’s AI boom runs on borrowed plumbing.

Indian startups, schools, courts, hospitals and millions of professionals now push daily work through ChatGPT, Claude and Grok. Every one of those tools routes through data centres on the other side of the world.

When the US clouds hiccup, India feels it in its evening prime time. That is exactly when it happened on Thursday.

India does not yet have a frontier model that could have absorbed that traffic. No Indian platform was a beneficiary of the chaos.

The government’s push for domestic compute and data centres, through the IndiaAI Mission and related programs, is partly a response to this dependence. But policy intent does not fill a prompt box at 8.30 pm.

The outage also lands at a moment when regulators are circling the same companies. The European Union has placed ChatGPT under its strictest platform rules. OpenAI is building what it described to lawmakers as automated shutdown capabilities for AI tools, reported The Hindu.

Global attention on who controls AI infrastructure, and what happens when it fails, is not going away.

The morning after

By Friday morning, ChatGPT, Claude and Grok were largely back online. Most users will have moved on.

The status pages went quiet. The memes about Astra faded. But the questions from Thursday evening do not resolve themselves.

Why did three competitors fail in the same minute? Was it Azure, Cloudflare, or a cascade that started with one platform and crushed the others?

Will any company publish a candid post-incident analysis, or will the silence continue?

And if the world’s AI runs on a handful of shared clouds, what happens the day the fault is bigger than East US?

The cursor is moving again tonight. That is the easy part.

The harder question is what happens the next time it stops. Because Thursday proved that when it stops, it stops everywhere at once.

Sources: OpenAI status page, Anthropic status page, Downdetector data as cited by India Today and PCMag, The Hindu (PTI), Moneycontrol, TechTimes, Economic Times, 9to5Mac/9to5Google, aiweekly.co, Cloud Security Alliance June 2026 analysis, Cloudflare post-incident reports.