April 7

Anthropic's CFO dropped a number. ARR went from $1 billion to $9 billion in one year. They projected $30 billion ARR within 15 months. OpenAI took years to reach $1 billion. Anthropic did the same distance in 12 months.

The same day, Project Glasswing quietly launched , the Anthropic cybersecurity consortium that would surface 10,000-plus vulnerabilities in the next 30 days.

April 20

Amazon moved. $5 billion cash plus a $20 billion option. AWS commitment: $100 billion over 10 years.

This was not a bet on a model. It was infrastructure capture. Amazon was saying Anthropic's workloads will define cloud compute for the next decade. The 10-year commitment makes that explicit.

April 23

OpenAI shipped GPT-5.5. The internal codename was SPUD. Everyone expected GPT-6. They called it 5.5.

Two possible reads: they are sandbagging the model name to manage expectations, or they are more worried about Anthropic than they have publicly admitted and not ready to call this their flagship.

Greg Brockman announced they are "70 to 80 percent of the way to AGI."

April 24

Google wired $10 billion to Anthropic. Four days after Amazon. Pledged up to $30 billion more in milestone payments. Total potential: $40 billion.

The same day, DeepSeek V4 dropped , on Huawei chips, at 44 cents per million tokens. 1.6 trillion parameters. 10 to 50 times cheaper than GPT-5.5.

The company that invented the Transformer architecture just wrote a $10 billion check to a competitor whose entire pitch is "we build better AI than you do." That is not how Google normally behaves. That is a company acting exactly like a follower , for the first time in 20 years.

What Spooked Them

Claude Code. That is the answer the numbers point to.

Claude Code reached $2.5 billion ARR in its first months. Enterprise coding market share: 54%. OpenAI Codex: 21% and the gap is widening. PwC rolled out Claude Code to their entire global workforce , all 364,000 staff.

Anthropic's own codebase: 90% written by Claude Code. A senior Google engineer in Seattle used Claude Code on a weekend to reproduce an internal architecture project that had taken his team approximately one year. He finished a working version in under an hour.

Claude Code is writing better versions of Claude Code. That is the recursive self-improvement loop the AI safety community warned about for a decade. It arrived quietly in a $20-per-month developer tool.

Eighteen days. Every major player showed their hand simultaneously. Something had shifted , and everyone knew it.