The series
Long-form work read sequentially. Each series builds an argument across multiple essays.
These are the eight sequential series on this site. Read them in order for the full argument, or use Topics to jump into specific areas curated across series. New series added roughly monthly.
Currently publishing
The ten regulatory workstreams determining whether AI infrastructure capex gets delivered on time. FERC + ISO reform, state PUCs, EPA, local zoning, CFIUS + international FDI, export controls, EU compliance stack, NEC/UL/NFPA, labour + immigration. Each can be a project killer; best operators treat all ten as Stage-1 architecture inputs.
Retrospective analysis of consequential AI infrastructure transactions and decisions. Blackstone/QTS, Microsoft/Constellation TMI, Amazon/Talen FERC rejection, CoreWeave post-IPO, Iron Mountain pivot, Digital Realty/Teraco, FTC/Nvidia-ARM, CMA UK cloud investigation. Each with retrospective scorecard and transferable lessons.
Specific companies and technologies reshaping AI infrastructure. Vicor factorized power, Menlo Micro SSCBs, Skeleton HSCs, Navitas/EPC GaN, Wolfspeed SiC retrospective, CoolIT + Submer cooling, Kairos + X-energy SMRs, Fervo geothermal, Bloom SOFC, Cerebras + Groq compute, specialised GPU cloud, Ayar Labs/Lightmatter CPO, Compass + Vantage modular.
Foundational series
The transition from 48V AC-DC to 800V DC distribution for AI infrastructure. Historical 48V, why 800V now, arc behaviour + protection, grounding + fault detection, battery + UPS restructure, retrofit vs greenfield, cooling co-emergence, power quality + grid interaction, standards convergence (OCP + IEC + NEC + IEEE), vendor economics, commissioning + skills, ten-year installed-base view.
Competitive framework across the AI-infrastructure power stack: capacitors (hybrid supercapacitors), wide-bandgap semiconductors (SiC + GaN), thermal (liquid cooling), interconnect, on-package delivery, modular datacentre. Each layer with vendor screen, TAM analysis, and design-win dynamics.
Capital flows and structures underwriting AI infrastructure. Hyperscaler capex, PE infrastructure vehicles, ratebase economics, contract-collateralised debt, sovereign wealth, cross-border structuring, exit optionality. Approximately $500B-1T of committed capital analyzed across the eight essays.
Full DD methodology for AI infrastructure transactions. Scoping, team + timeline, market DD, customer/pricing DD, product architecture, manufacturing + supply chain, QoE + working capital, capex + growth economics, org + personnel, legal, regulatory + ESG + cyber, deal structure, financing, post-close validation. Plus three supplements including a full Eaton/Boyd worked example.
Recurring formats
Deep architectural analysis of specific AI infrastructure transactions as they are announced. Format: deal structure, valuation logic, comparable analysis, red flags, transferable lessons.
Analysis of the transaction pipeline as it develops. Preliminary reads on announced or emerging deals before full teardowns.
One-page recap of the month in AI infrastructure. What moved, what the market is getting wrong. Feeds the Living Indexes (once launched) and provides regular structured coverage across NVIDIA, OCP, hyperscalers, and power equipment.
Foundational series · additions
Cross-layer synthesis following the electron from grid interconnect to accelerator die. Vendor concentration, standards state, capex intensity and inflection years per stage. Composes with the AI Power Chain, Investment Layer and Due Diligence frameworks. Anchor for the ongoing series of stage-by-stage deep-dives.
Deep-dive on the OCP Mt. Diablo 400 rack specification (v0.5.0 May 2025 through v0.7.0 March 2026, co-authored by Microsoft + Meta + Google). Four-plane vendor coverage, the August 2026 OCP 800V standardisation with NVIDIA inside the collaboration, and the strategic contest as rack-integration margin migrates above the interface layer. Companion to P4 Hall power at the sidecar/row/facility architecture layer.
Six additive economic roles the same physical battery performs: backup (R1), ride-through (R2), power smoothing (R3), peak management (R4), grid services (R5) and grid-forming (R6). Each migration requires a specific technical, contractual and regulatory gate. Screen for R-stage exposure before chemistry, ownership model or geography. Connective frame across the lithium UPS, hall BESS, EU EED overlay and load-side flexibility companions.
Two simultaneous migrations at the interconnect layer. Boundary 1 is medium migration (copper to optical) at intra-rack scale. Boundary 2 is integration migration (pluggable transceiver to co-packaged optics) at rack-to-rack and switch-fabric scales. Silicon photonics fab ecosystem (GF Fotonix, TSMC, Tower, STMicro, IMEC), pluggable + DSP cohort (Broadcom, Marvell, Coherent, Lumentum) and CPO specialists (Ranovus, Ayar Labs, Lightmatter, Celestial AI). The economic control point may migrate toward whoever owns the electrical-to-optical boundary itself.
An investment framework for the reclassification premium between firm AI capacity and firm-plus-flexible AI capacity. Five control models (workload orchestration, storage-mediated facility, BTM dispatchable, DER aggregation, hyperscaler-owned), three monetisation channels (market revenue, cost avoidance, time-to-power NPV), and where the reclassification premium routes. FERC RM26-4 validates flexibility as an interconnection attribute, not just a grid-services product. Fourth P4.2 companion to lithium UPS, hall BESS and the EU EED regulatory overlay.
Fifth and final stage of the Grid-to-Chip framework. Every new power architecture destroys one conversion layer and creates another; the investable question is where the value migrates. SiC + GaN wide-bandgap semiconductors, POL/VRM specialists, factorised power (Vicor), on-package IVR (now inside Analog Devices via the July 2026 Empower acquisition). Nvidia 800V direction of travel drives fewer conversion stages and high-ratio late-stage conversion near the accelerator. Two-tier vendor scorecard with architecture-displacement-risk column.
Fourth stage. Low-voltage transformers, UPS, hall BESS, PDUs and the AC-to-DC rectification boundary migration. Four-vintage architecture (V1 legacy AC / V2 hybrid 800V sidecar / V3 row-level 800V / V4 native 800V facility) with per-vintage vendor scorecard. Companion P4.2 investment stack (see hub).
Third stage. HV-to-MV step-down at the site substation, MV distribution across the campus, protection and switchgear. Solid-state transformer vs conventional MV transformer investment contest, driven by Nvidia 800V DC AI factory demand pull and constrained conventional MV transformer supply.
Second stage. The scarce asset is a contract or a queue position. Interconnection agreements, ISO cluster studies, grid-services contracts, load-serving-entity arrangements. Build-versus-extract framework for grid capacity constraints.
First stage. Firm power procurement: nuclear PPAs, SMRs, behind-the-meter gas plus storage, renewables firmed with dispatchable capacity. The economic unit is a contracted energy stream with defined firmness characteristics.
Standalone deep-dives
Reference underwriting for the Interport multi-port SST architecture. Nvidia MGX qualification chain, single-point-of-failure, customer-adoption trajectory, rent pool, decision tree and the falsification triggers that would break the thesis.
The three US SST pure-plays applied to the 800V DC + MV transformer supply gap. Head-to-head architectures, socket map, adoption scenarios, and the specific milestones that will decide the winner. Standalone tech spotlight ahead of the full Tech Spotlights series.
Canonical reference for on-package delivery (Layer 5 of the AI Power Chain). Four walls of physics, five-way IVR architecture taxonomy (Board VRM to on-die IVR), two-stage PDN, Empower / Ferric magnetics split, the ADI $1.5B acquisition anchor, the Marvell test, seven falsifiers, and the $1.5B reverse-underwrite.
Companion to The IVR Stack. Ferric is an option on the monetisation of merchant IVR scarcity. Value-creation engine, customer economic value (Gate 0), commercialisation ladder (evaluation to platform dependence), evidence-conditioned six-state valuation ladder from $200M to $4B+, three vetoes, tiered acquirer scorecard, reverse-underwrite arithmetic, and the Marvell test.