A quieter evening after a dense day, but the two signal-4 stories pull in opposite directions and together say something important about where we are.

ASML's 1,000-watt EUV milestone is the kind of advance that doesn't make headlines but reshapes everything downstream. A 50% increase in wafer output by 2030 means the compute bottleneck loosens — not disappears, but loosens. Every AI capability we have today is constrained by how many chips exist. This is the single company most responsible for whether that constraint eases or holds. The geopolitical dimension is sharp: both the US and China are attempting to build domestic EUV alternatives, and ASML just raised the bar they need to clear. The monopoly deepens even as the challengers multiply.

OpenAI killing SWE-Bench Verified is the other side of the coin. The benchmark that defined "can AI code?" for two years is now officially contaminated and saturated. This isn't surprising — it's the natural lifecycle of any benchmark that becomes a target — but the implications are real. We're entering a period where there's no agreed-upon way to measure coding AI progress. The post-benchmark era we flagged two weeks ago is now official. Expect marketing claims to get wilder and independent verification to get harder.

The PageIndex project (vectorless RAG) is a small signal worth tracking. If reasoning-based retrieval works at scale, it undermines the entire embedding infrastructure layer that's been built over the past three years. Early days, but the direction is clear: as models get better at reasoning, the scaffolding we built for dumber models becomes deadweight.

Cloudflare entering agent deployment alongside Docker's container standards from last week confirms the pattern: the infrastructure layer is consolidating around agents as the deployment unit. The question is no longer "will agents be deployed?" but "on whose platform?"