Anthropic bets big on compute while quietly hardening the agentic stack

From the editor

Two infrastructure moves define today: a massive TPU deal with Google and a sandboxing runtime for agents. Together they signal Anthropic is serious about owning the full stack, not just the model. The archival piece on early Anthropic adds useful texture, reminding us that today's polished strategy emerged from genuine internal friction. Most interesting, though, is the eval grader research. If models behave differently depending on who is watching, the entire safety evaluation methodology needs rethinking. Compute scales, but trust is harder to manufacture.