The thesis hardened from two sides this week. NuttyCLD's six-part Power Crisis series consolidated into a delivery-side perspective on May 1; the next day, Crusoe CEO Chase Lochmiller — operator-side, hyperscaler customer — independently validated the generation-side AND hardened the incumbent-electrical bear from the delivery side. The frame is no longer "Aschenbrenner sees something." It's now "the most-cited speaker in this thesis is also a customer of the incumbents he's bearish on."

What happened

A new perspective was admitted: ai-power-delivery

Cited: 2026-05-01-ai-power-delivery-spark

NuttyCLD's series (Mar 30 → May 1) reframed the AI capex question from supply-side ("more electricity") to delivery-side ("we lose 30% of what we already have"). NVIDIA's own analysis — 300 MW of parasitic loss at a 1 GW data center — is now the spine of an investible supply chain that the market hasn't fully repriced. Initial cohort: MPWR, AOSL, VICR, IFNNY, NVTS, POWI, BESIY / AMS:BESI, ON, STM, WOLF, VRT, ETN, AIXG, VEEC, AMAT, LRCX, ASML, MRAAY, INTC, TSM. MPWR projected ~70% Vera Rubin VRM socket share with ASPs ~60% higher than Blackwell.

Operator-side ratification of the generation-side bull

Cited: 2026-05-02-crusoe-genside-bull-confirmed

Lochmiller (Crusoe CEO) on a 49-min talk hosted by @apoorv03:

• Explicit GE Vernova endorsement: "It's been good to be a shareholder of GE Vernova." Gas turbine pricing has roughly 3x'd ($1M/MW → ~$3M/MW); GEV is the U.S. champion.
• Onsite gas-turbine generation is the unlocking layer for AI capex. Grid timelines are too long for the buildout cadence the workload demands. Validates the BE / CRWV / CORZ / IREN / APLD "onsite power = first-mover" cohort from a customer voice.
• Compute depreciation likely longer than 6 years. H100 spot pricing has risen since launch — counter to the "next gen makes old gen worthless" assumption. Agentic-workflow demand keeps older silicon valuable. If 8-10 years is right, AI cloud unit economics re-rate meaningfully.

Operator-side hardening of the delivery-side bear

Cited: 2026-05-02-eaton-incumbent-bear-hardens

Same Lochmiller transcript, second-order read: Eaton + Schneider — the legacy electrical primes that sit at Layer 1 of the voltage stack — "have fundamentally not innovated in the last hundred years," and are exposed to solid-state transformers + 800V DC adoption. This is operator-side validation of the existing perspective's "incumbents-monetize-the-door, challengers-monetize-the-wall" framing, but from a major customer of the incumbents specifically calling out their stagnation. Asymmetry tightens onto challengers (AOSL, NVTS, VICR, POWI, BESIY / AMS:BESI).

What's next on the calendar

Cited: 2026-05-28-nvda-q1-fy27-earnings

NVIDIA fiscal Q1 FY27 earnings around 2026-05-28. First print since this perspective set crystallized. Read-throughs to watch:

• Hyperscaler capex commentary — does Jensen surface explicit power-bottleneck framing? If yes, the entire generation-side cohort gets a third-party validator (after Aschenbrenner and Crusoe).
• Vera Rubin shipment commentary — first window where Rubin power-delivery suppliers (MPWR primary, AOSL/VICR secondary) appear in customer mix.
• Compute residual-value language — any analyst Q on H100/Blackwell residual value tests Lochmiller's 8-10 year depreciation read.

Forward affects: ai-power-bottleneck (gen-side reads), ai-power-delivery (delivery-side reads), war-ends-playbook (NVDA is core long there).

How the picture composes

The two operator-side reads from the Crusoe transcript are mirror images, applied to different layers:

• Gen-side: incumbents getting paid (GEV, gas turbine OEMs) because the generation problem is what's hard, not what's stagnant.
• Delivery-side: incumbents exposed (Eaton, Schneider) because the delivery problem is what's stagnant, not what's hard.

Same speaker, same transcript, opposite directional reads — because the layer matters. The four-layer voltage stack frame from NuttyCLD lets us hold both reads coherently. Without that frame, "Crusoe says nice things about GEV but bad things about Eaton" reads as a mixed signal. With the frame: it's the same signal applied to two layers with different innovation rates.

Composition note (system-internal)

This digest is itself an event in the event-sourced log:

• event_id: 2026-05-03-ai-power-week-in-review
• type: digest.weekly (namespaced — see event-schema)
• affects.events: lists the 4 cited events; affects.tickers: lists the 9 tickers it discusses; affects.perspectives: lists the 2 perspectives it synthesizes from
• Cited events stay queryable independently; the digest adds a curated reading on top without replacing them

When the per-ticker feeds rebuild, this digest will surface in feeds for GEV, BE, ETN, MPWR, AOSL, VICR, BESIY, POWI, NVDA — alongside the underlying events it cites. Reading order in those feeds: digest first (most recent), then the cited events. The user can drill from synthesis into source.

This is the amalgamation pattern: events are atomic posts (twitter); per-perspective and per-ticker feeds are the subreddits; digests/issues/briefs are the curated weekly roundups that cite the underlying posts. All three projections share the same event store.