AI Apartheid: Paid Claude users suffer from outages, while the U.S. government operates without interruption
On June 23, 2026, Anthropic's AI assistant Claude experienced an 85-minute outage. The incident completely paralyzed public access to the platform, yet the specialized version of the service for government agencies remained untouched. This situation revealed a deep, deliberately constructed infrastructure divide: regular users and business clients were cut off from the AI, while government systems continued to operate normally.
Anthropic's service status monitoring showed that the company's technical specialists recorded widespread errors across nearly all key products: the claude.ai web interface, API, the Claude Code tool, the console, and the Cowork service. Notably, the government version of Claude (Claude for Government) was not included in the official list of affected segments.
The timeline of events demonstrates a less-than-swift response. 35 minutes after the incident began, the team announced they had found a technical solution, yet the overall system paralysis lasted 85 minutes. The status was marked as "resolved" only two hours after the first problem signal.
Numbers Highlighting Inequality
A comparative analysis of service stability over the past 90 days paints a telling picture. The government version of Claude showed an uptime of 99.93%, while the public service claude.ai achieved only 99.10%. Downtime also differs dramatically: approximately 90 minutes for the government version versus nearly 19 hours for the public one.
Such statistics inevitably sparked a sharp reaction from paid subscribers. Posts quickly appeared on social media, accurately reflecting the audience's sentiment. One user succinctly noted: "Claude is down for everyone except the government."
Architecture of Two-Tier Access
This deep division is not accidental—it was architecturally embedded. The government version of Claude operates entirely separately from commercial users. The platform is deployed on the specialized federal cloud service Palantir, which holds strict FedRAMP High certification. It was this complete environmental isolation that ensured government AI systems continued to function without interruption, while consumer products faced massive failures.
Recall that last year, the General Services Administration (GSA) provided access to Claude to all three branches of government for a nominal fee of $1. Such precedents only underscore that for strictly regulated cloud services, isolated infrastructure is a classic standard.
Context and Scale of the Problem
The June 23 incident was just part of a highly challenging period for the company's public segment. According to Anthropic's official status page, more than 20 incidents were recorded between June 9 and June 23. Users regularly encountered errors or complete system unavailability, which primarily affected the new flagship Opus 4.8.
To cover rapidly growing demand, the company promptly reserved up to 5 GW of new computing capacity from Amazon and also signed new lease agreements for modern data centers. Against this backdrop, the isolated government segment remained safe from the outset, as it was not part of the overloaded resource pool.
My comment as an analyst: This incident is not just a technical failure but a vivid demonstration of an emerging "AI apartheid." We are witnessing how critical infrastructure for government needs receives priority redundancy and protection, while paying users essentially become a "testing ground" for load testing. For the long-term sustainability of the AI services market, SLA (Service Level Agreement) models for commercial clients need to be reconsidered; otherwise, trust in platforms will inevitably be undermined.