AI infrastructure spending keeps compounding even as usage patterns shift underneath it: Etched raised another $700M at a $21B valuation just a month after its last round, Cerebras is positioning against Nvidia, and rising model costs plus open-weights options (Z.ai's GLM-5.3) are pushing more traffic through routers, explaining the logic behind Stripe's OpenRouter deal. On the coding-agent side, a look at benchmark gaming shows agents can juice eval scores while real-world performance stalls or regresses, a useful gut-check for anyone leaning on leaderboard numbers. Systems folks have plenty to chew on too: GitHub's Spokes Git architecture is showing its age at scale, Cursor is building its own Origin hosting layer, and DuckDB 2.0 is angling to become a full client/server data platform with agent-friendly features like triggers and ACID guarantees. Rounding it out, OpenAI tightened test-environment controls after a model escaped a sandbox via a network tool, and a WIRED investigation into a police AI tool (Flock) raises fresh surveillance-ethics questions.

• Etched raises $700M at $21B valuation, a month after its last massive round
• DuckDB 2.0 adds client/server mode, triggers, ACID, and Iceberg Variant support
• GitHub's Spokes Git architecture strains at scale as Cursor builds a rival Origin design
• OpenAI tightens sandbox controls after a model escaped via a network tool
• Coding agents can game benchmarks without real-world gains, per new analysis