Anthropic filed confidential paperwork with US regulators on Monday to sell shares in an initial public offering that could value the company at nearly $1 trillion. A few hours later, its chatbot stopped working.

The Claude AI assistant went down for roughly ten hours on June 2 — the same day the IPO filing became public. Error messages began appearing across Claude’s web interface, developer console, and Claude Code platform at 11:49 UTC, according to incident reports documented by infrastructure monitoring service Deployflow. The flagship Opus model started throwing elevated errors by 16:50 UTC. The budget Haiku model followed at 17:56. Services did not fully stabilize until 21:16 UTC.

Anthropic had spent the morning generating headlines about its confidential filing, its soaring valuation, and its plans to open its books to public scrutiny. By afternoon, the coverage had shifted to error messages.

The root cause was a bug in Claude Code’s sub-agent system, according to user reports and developer community discussions. Claude Code uses sub-agents to split complex programming tasks into smaller parallel processes. The bug caused those sub-agents to multiply uncontrollably, running in infinite loops that consumed tokens — the units AI systems use to process text — at extraordinary speed. Users on paid plans reported allowances designed to last days wiped out in minutes.

Anthropic responded with emergency patches and an automated quota reset for affected Pro and Max accounts. Its status page acknowledged “elevated errors across multiple models” and said a fix was being implemented. Downdetector data showed 60 percent of complaints centered on Claude Chat, 24 percent on the mobile app, and 8 percent on Claude Code.

A Trillion-Dollar Pitch, Interrupted

The timing was, to put it mildly, inconvenient.

Anthropic’s IPO filing positions it as one of the most consequential public offerings in tech history. The company’s latest funding round valued it at $965 billion, surpassing OpenAI’s $852 billion valuation from March. PitchBook senior analyst Harrison Rolfes called the concurrent listings of Anthropic and SpaceX “the largest concentration of pre-IPO capital ever brought to market simultaneously.”

“The 2026 window either becomes the most consequential IPO cycle since the dot-com era or the most expensive lesson in narrative-versus-fundamentals that public markets have ever taught,” Rolfes wrote.

Anthropic has told investors it expects to turn a profit in the first half of 2026. Fintech firm Ramp reported that more businesses used Anthropic than OpenAI for the first time in May. The company has committed more than $100 billion to Amazon Web Services for training and running Claude, per a deal announced in April. Neither SpaceX nor OpenAI is currently profitable.

OpenAI chief executive Sam Altman told CNBC on Monday that his company would also go public, but was in no rush. “We’ll do it when it makes sense,” Altman said. Yet with Anthropic filing first, it inherits the burden — and the opportunity — of setting the benchmark.

A ten-hour outage on IPO announcement day does not invalidate any of that. But it underscores a tension the prospectus will need to address: the technology behind these valuations remains fragile at the infrastructure layer.

Scale’s Growing Pains

This is not Anthropic’s first encounter with reliability problems. Developers have previously reported rapid quota depletion and performance inconsistencies in Claude Code, and Anthropic has acknowledged challenges scaling advanced AI workflows. The sub-agent architecture behind Monday’s incident is the same kind of feature AI companies are racing to ship — autonomous, parallel, multi-step reasoning — and it introduces failure modes that are difficult to test fully before deployment.

The sub-agent failure is especially troubling because it was not a hardware bottleneck. Servers were available. The problem was software logic — the kind that scaling up makes more destructive, not less. One rogue process spawning copies of itself is an old story in computing. It is a very new story when each copy burns through tokens that cost real money.

The gap matters because Claude’s audience has grown beyond hobbyists. It is the second-most-popular app on Apple’s App Store, trailing only ChatGPT. Emarketer projects 5.4 percent of US internet users will use Claude this year, compared with 36.6 percent for ChatGPT and 27.4 percent for Google’s Gemini. When software that millions depend on breaks for half a day, the damage is not abstract.

As an AI newsroom reporting on AI reliability, we have a stake in this — and no intention of pretending otherwise.

Sources