Power is the clearest infrastructure story this morning. The -9.8% move to $56/MWh is real margin relief for data center operators running continuous inference workloads, and it arrives while compute stays flat at $4.09/hr and HBM constraint holds at 71%. The cost stack is loosening at the edges, not the core. Memory remains the binding pressure.
The most important deployment signal is Block's Goose agent reducing headcount by 40% following a record quarter. That's not a prototype claim — it's a live production deployment with a named productivity figure, and it sets a benchmark that enterprise software buyers will now carry into every vendor conversation. Perplexity's orchestration layer launch the same day reinforces the direction: multi-agent workflows are moving from evaluation to procurement decisions, and the agent score's -20-point slide over seven days reflects ecosystem fragmentation more than demand softening.
Yesterday's watch on power as a binding constraint for model scaling by Q2 2026 is still open. No new data today directly resolves it, but the -9.8% power move cuts against the tightening thesis in the near term. We'd watch whether that relief holds through Q1 or whether rising inference density from parallel agent workloads — exactly what Claude Code's team feature and Perplexity's orchestration push — absorbs the slack before it matters.
MatX raising to challenge Nvidia is worth flagging before the call. The compute score hasn't moved in seven days, but a credible challenger entering procurement conversations changes the timeline calculus for operators locked into H100 roadmaps. Meta's AMD chips-for-stock structure is now a third consecutive day of compute partnership news — that's a trend, and it points toward hyperscalers actively diversifying foundry exposure rather than waiting on Nvidia's next cycle.