Datacenter Power — What Matters This Week (May 2026)

1. Top 3 Things Happening (User View)

1. Behind-the-meter SOFC + gas-turbine PPAs replace the 5-7yr utility interconnect queue as the default path for 50-200MW AI campuses (May 2026, products module: "1–500 MW prime-mover assets... sited between the utility meter and the campus busway"; Bloom signal_id 7ee2bd4f FY25 $1.65-1.85B; GE Vernova bb99bbf8 $19B/$129B backlog).

• Why it matters to: POWER-GRID-LIAISON-01 and POWER-FACILITY-ENG-01
• Action: Pull your current site list, mark every project >50MW whose interconnect date slips past 2028, and request a behind-the-meter quote from Bloom or GE Vernova this week.

1. Rack/row power+cooling has collapsed into a single SKU for GB200/GB300 NVL72 racks — Vertiv and Schneider now sell CDU+busway+UPS+PDU as one line item, not four (products module "coupled liquid-cool-plus-power SKU"; Vertiv 06847802 $8.5B AI-attributable; Schneider f76e172f $22B Energy Mgmt portion).

• Why it matters to: POWER-FACILITY-ENG-01 evaluating vendors for 50-200MW campuses
• Action: Re-run your last RFP as a coupled SKU — single-vendor liquid+power likely beats 4-vendor bids on lead time, not price; demand 18-month lead-time SLA in writing.

1. Hiring signal is flat zero across all L1 OEMs for 90 days while order books grow — backlogs are being filled from existing headcount, not by scaling teams (hiring module: 0 roles across Vertiv/Eaton/Schneider/GE Vernova; market module $55-65B TAM with named tailwinds).

• Why it matters to: POWER-VC-ANALYST-01 modeling order-book conversion
• Action: Short the thesis that L1 OEM revenue scales linearly with AI capex — model a margin-expansion (not unit-volume) quarter into Q3 2026 estimates.

2. Strategic Implications (Our View)

1. The bottleneck has migrated from chips to electrons and from electrons to permits — which means the winning vendor map is now permit-arbitrage, not technology.

• Evidence: products module — "Standby & permit-arbitrage gensets... selected on lead time, fuel mix, and air-permit posture rather than steady-state efficiency."
• Implication for next 30 days: Our coverage should reweight from chip-cycle pieces toward state-by-state PUC + air-permit reporting; one explainer per ISO/RTO queue beats another GB300 teardown.

1. Low-signal-density across 6 of 7 company profiles plus 0 hiring roles is itself the signal: this is a backlog-burn quarter, not a discovery quarter — coverage should treat news absence as a real data point.

• Evidence: 6 company excerpts marked `low_signal_density`; hiring module `total_roles_90d: 0`.
• Implication for next 30 days: Replace "what's new" framing with "what's shipping from the 2024 backlog" framing; lead with delivery dates, not announcements.

1. Persona evidence shows POWER-FACILITY-ENG-01 and POWER-RELIABILITY-ENG-01 both name vendor monitoring lock-in as a top blocker — a wedge product opportunity sits there, but no L1 OEM will build it.

• Evidence: POWER-RELIABILITY-ENG-01 — "Vendor-specific monitoring lock-in (Vertiv / Schneider / Eaton each have own)."
• Implication for next 30 days: Scout one early-stage company building a vendor-neutral power-telemetry layer; if none exists in our pipeline, that gap is the next briefing.

3. Newsletter Issue Spotlight

• Lead story: Behind-the-meter SOFC + gas turbines are now the default path around 5-7yr utility queues for AI campuses
• 3-second hook: The chip shortage is over. The new shortage is a substation — and hyperscalers just stopped waiting.
• Image direction: V16 isometric 3D — a transparent AI datacenter campus split down the middle: left half waits in a frozen queue of grey substations stamped "2031"; right half is already live, fed by a glowing row of Bloom SOFC boxes and a GE Vernova gas turbine sitting between the utility meter and the busway. Single warm light source on the right.
• Subject line: Hyperscalers stopped waiting for the grid

4. Reverse-Hype Watch

• Overhyped right now: "AI datacenter hiring boom" — hiring module shows 0 roles across all 7 L1 OEMs for 90 days; the narrative is a year behind the actual headcount data.
• Underrated right now: Air-permit posture as a vendor-selection criterion — products module explicitly names it ahead of efficiency, but no major coverage treats permit-engineering as the moat it now is.

5. Pattern Library Updates

• Success pattern observed: "Sell the coupled SKU, not the components" — Vertiv and Schneider winning GB200/GB300 sockets by bundling CDU+busway+UPS+PDU into one liquid-cool-plus-power line item (products module).
• Failure pattern observed: "Treat low-signal-density as nothing-to-report" — 6 of 7 company profiles ran with `low_signal_density: true` and produced near-identical "riding AI capex" snapshots; the silence itself was the story (backlog-burn, not discovery) and was missed.