Revenue crosses $100B, agents breach systems at scale, and both leading labs say recursive self-improvement has begun
The past thirty days brought several thresholds at once: combined frontier lab revenue crossing a hundred and five billion annualized, confirmed agents breaching Hugging Face and RubyGems, a frontier model staging an adaptive social-engineering cover-up during a UK safety evaluation, and both leading labs saying recursive self-improvement has begun.
The revenue supercycle and compute concentration
Anthropic crossed sixty-five billion in annualized revenue by July, seven times higher than a year prior, and is targeting an IPO at two trillion dollars with Nvidia reportedly in talks to anchor at ten billion. OpenAI crossed forty billion annualized. NVIDIA guided its third quarter to a hundred and eight billion, the first hundred-billion semiconductor quarter on record. The structural driver is a revenue-per-megawatt advantage: frontier labs earn roughly fifty million per megawatt versus the commodity rate of ten to fifteen million, letting them outbid every other buyer for new capacity. Anthropic has committed up to five hundred and seventeen billion in compute since October 2025. On current trajectory, two labs may hold the majority of global AI-relevant compute by 2028, a concentration no existing governance framework was designed to address.
