Anthropic's Compute Shortfall Has Become a Trust Crisis
Anthropic is navigating a demand crisis rooted in a capex decision Dario Amodei made in late 2024. Rather than commit to trillion-dollar compute builds — which would have bankrupted the company if revenue growth slowed from 10× to 5× year-over-year — Amodei chose conservatively. Demand did not slow; it accelerated, driven by agentic coding workloads, and Anthropic now cannot serve it at the volumes its subscription model implied.
What the Source Actually Says
Matthew Berman's 2026-04-23 deep-dive enumerates the fallout in specific detail. Anthropic ran a 2% A/B test removing Claude Code from the Pro tier. Peak-hour quota limits — described internally as "affecting 7% of power users" — impose a 5-hour constraint window. On April 3, 2026, at 4 PM Pacific on an Easter Friday, Anthropic banned third-party harness usage (specifically OpenClaw, the most-used agentic harness) via policy changes announced through reply tweets — contradicted by its own documentation for weeks afterward.
Opus 4.7 adds two distinct quota-inflation vectors: a new tokenizer that maps identical input to 1.0–1.35× more tokens, and higher thinking-token output at elevated effort levels — both confirmed by Anthropic's own Boris on social media. Uptime data compounds the picture: claude.ai at 98.8% and the Claude API barely above 99%, versus OpenAI API at 99.8–99.98%.
OpenAI has weaponized every misstep. The Codex team reset usage limits for 3 million weekly users and pledged "transparency and trust" in explicit contrast to Anthropic's approach. Peter Steinberger, OpenClaw's creator, has since been acquired by OpenAI, which now allows OpenClaw usage on Codex subscriptions.
A longer-term fix — a 5 GW Trainium 2–4 capacity deal with AWS worth over $100 billion across 10 years — won't come online for at least a quarter, leaving the current demand gap unresolved.
Strategic Take
For practitioners building on Anthropic models, the short-term picture is constraint. Teams running agentic coding pipelines should actively stress-test OpenAI Codex and Google Gemini as failover paths — not because Claude's quality has degraded, but because quota, uptime, and policy predictability are now live variables that deserve the same engineering attention as model capability.