Data Centers — Deep-Dive Brief

Date: 2026-05-24 Type: /research output — comprehensive topic overview Scope: Equipment, energy, unit economics, major players, customer landscape, political concerns, financialization opportunities Author: research-agent (synthesized), reviewed by Bliss

Comprehensive overview for Project TBD financialization wedge orientation. Internal evidence is thin on DC operators specifically — this is largely an external research brief with three rich internal anchors (Lonny/NVIDIA reverse logistics, Vivian/AI DC supply chain, Preston/parametric structures). External data is dated 2025–2026; verify any number before quoting externally.

1. Internal Knowledge (Memory Vault)

What we have heard directly

NVIDIA’s reverse logistics is breaking under hyperscaler GPU volumes. Lonny Orona, 8 weeks into a role at NVIDIA after 8 years at Meta, described a $5T company running its repair/returns operation on email and spreadsheets. Stack is Salesforce (ticketing) + SAP (material planning) + Baxter (demand planning) + Expeditors (3PL), unintegrated. New separate repair lines (Dallas, July 2026) run by Wistron and Foxconn, back-office in Hong Kong, warehouse in Taiwan. Active procurement posture: “We have no time for in-house tooling. If these tools don’t generate revenue, why would we spend time building them?” [Interview: Lonny Orona, 2026-05-12]

The scaling gap is concrete. Meta currently has ~100K GPUs and wants 1M over five years. NVIDIA is “struggling to get hundreds of units back” today — thousands will break the operation. [Interview: Lonny Orona, 2026-05-12]

SLA inventory imbalance from business-unit decision-making. Hyperscaler DC teams cannot unilaterally authorize rack downtime; business units (Instagram, Facebook) often prefer “let it fail” rather than authorize proactive replacement. Advance-replacement units sit idle for weeks/months. [Interview: Lonny Orona, 2026-05-12]

ODM bypass on returns. NVIDIA piloting direct pickup from hyperscale customers, cutting Quanta and other ODMs out of reverse flow because integrators are production-focused and add no value on returns. [Interview: Lonny Orona, 2026-05-12]

NVIDIA warranty reserves are a financialization signal. NVIDIA held $8.22B in warranty reserves at end of FY2025, with additions “primarily related to Compute & Networking segment” — i.e., data center GPUs. The entire US semiconductor industry set aside $1.752B in calendar 2024; NVIDIA alone accrued $2.59B in FY2025. [Public: WarrantyWeek citing NVIDIA SEC filings, 2026-04; synthesized in reverse-supply-chain-research-2026-05-13.md] Max Mirgoli (May 22) independently suggested studying “NVIDIA’s warranty claim size vs. revenue and the potential to reinsure that warranty risk.” [Interview: Max Mirgoli, 2026-05-22]

AI data center components are the “hottest” topic in semis. Beyond GPUs: substrate, cooling systems, power connections, passive components, testing. Each is a separate company; redundancy/2nd-source pressure is driven by (a) margin, (b) supply security, (c) competitive capacity blocking. Google and Amazon “are super active right now — they are trying to approach different kinds of components manufacturers” to lock up capacity. [Interview: Vivian, 2026-04-29]

Hyperscalers (CSPs — Google/Amazon/Meta) are the demand pull. Vivian framed the value chain as: CSPs buy from NVIDIA → NVIDIA sources GPUs from TSMC + components (cooling, power, substrate, PCB) → ODMs (Foxconn/Hon Hai is largest) integrate into rack-scale AI systems → ship to CSPs. [Interview: Vivian, 2026-04-29]

Edge AI as the next cycle vs. saturated data center buildout. Josh flagged edge AI as a possible next major market, contrasting it with the “saturated data center buildout” — timing uncertain. [Interview: Josh, 2026-04-30] Note: this is a contrarian datapoint vs. external “AI capex sprint” framing.

Power consumption framed as the Nobel-Prize-worthy unsolved problem. Max Mirgoli described 100x reduction in chip power consumption as the defining technical challenge of the era. [Interview: Max Mirgoli, 2026-05-22]

Brett’s NGP contact was flagged as an O&G/data center perspective intro — not yet actioned in the vault. [Interview: Brett, 2026-04-30] Worth surfacing.

What our parametric specialist already told us about DC risk

Man-made equipment failure in fabs and data centers is the structural gap. Preston’s parametric specialist explicitly identified that no parametric product currently exists for fab overheating, GPU failure, manufacturing process breakdown — because all three non-market pillars are missing: no trusted third-party measurement agent, no agreed metric, no actuarial model tied to historical loss data. “If embedded temperature sensors already exist in fabs, data standardization might be achievable — but calibration, cross-vendor normalization, and third-party trust would still need to be constructed.” [Interview: Preston, 2026-05-22]

Mo Islam’s compute-pricing gap. “What is the index for compute? What is that equivalent in the semiconductor space that doesn’t exist?” Current platforms pull listed neo-cloud prices, which Dustin’s prior contact described as “totally unreliable” — Google’s negotiated rate can be half the listed rate. [Interview: Mo Islam, 2026-05-22]

What we have NOT heard from anyone

No interviews with colocation operators (Equinix, Digital Realty, QTS, CyrusOne, Iron Mountain, Aligned, EdgeConneX).
No interviews with neoclouds (CoreWeave, Crusoe, Lambda, Nebius).
No interviews with DC power/cooling equipment OEMs (Vertiv, Schneider, Eaton, Cummins, Caterpillar, Bloom Energy).
No interviews with hyperscaler infrastructure or procurement teams — Lonny is our deepest signal but he is on the supplier side (NVIDIA), not the buyer side.
No interviews with DC developers, REITs, or infrastructure investors (Blackstone/QTS, Brookfield, DigitalBridge, KKR).
No interviews with utility/grid-interconnect actors or state regulators (PJM, ERCOT, FERC).

2. External Sources (2024–2026)

2a. Equipment

GPUs / accelerators (the prime mover). NVIDIA dominates; rack densities are climbing fast.

GB200 NVL72 racks at ~132kW; Blackwell Ultra/Rubin generations targeting 250–900kW with up to 576 GPUs/rack by 2026–2027; some upcoming designs targeting 600kW per rack. [Public: Introl, Network World, 2026]
Microsoft AI clusters use 100kW/rack as a standard design; NVIDIA’s 120kW 72-GPU rack ships exclusively with direct-to-chip liquid cooling. [Public: Network World, Introl 2026]

Cooling. Liquid is mandatory above ~50–100kW/rack.

Liquid cooling adoption forecast to hit ~40% of AI data centers by 2026. [Public: EnkiAI, 2026]
Chemours/2CRSi (Feb 2026) accelerating two-phase liquid cooling; Vertiv launched Mega Mod HDX (Jan 2026) for pre-defined high-density compute. [Public: industry press, 2026]
Liquid-related losses now ~24% of total DC loss costs — a new insurance risk vector. [Public: Risk & Insurance, 2026]

Power distribution & switchgear is the binding constraint.

Large transformers: 80–120 week lead times; transmission-class units 3–6 years. [Public: Sandstone Group, 2026]
More than half of US data centers planned for 2026 expected to be delayed or cancelled due to insufficient electrical equipment. [Public: Energy News Beat, 2026]
Vertiv Q4 2025 backlog: $15.0B, up 109% YoY, book-to-bill ~2.9x. [Public: Vertiv 8-K, 2026-02]
Eaton Q1 2026: datacenter orders up 240% YoY, datacenter revenue up 50%, backlog $22.8B. [Public: industry reporting, 2026]
GOES (grain-oriented electrical steel) shortage is the underlying physical bottleneck for transformers. [Public: industry analysis, 2026]

Networking. External research not deeply pulled in this brief; gap to fill. Anecdotally, InfiniBand vs. Ethernet (Ultra Ethernet Consortium / NVLink scale-up vs. scale-out) is an active battleground — surface in future brief.

2b. Energy

Power per facility. AI campuses now measured in GW, not MW. By 2030, global DC capacity projected at ~200 GW (~doubled from 2025). [Public: industry consensus across JLL/Synergy, 2026]

PUE benchmarks.

Global average: ~1.54–1.58. [Public: Statista 2025 survey]
Hyperscalers: Google fleet avg 1.09 (2024); Meta ~1.09. [Public: Google DC efficiency disclosures, 2024]
Hyperscale + liquid-cooled AI: routinely 1.04–1.10. [Public: Huawei DC blog, 2026]
Germany’s Energy Efficiency Act requires new DCs ≤1.2 PUE starting 2026. [Public: clearcomfort.com, 2026]

Grid interconnect lead times.

36 months (Pittsburgh) to 84 months (Columbus) for grid power; ERCOT/MISO/PJM queues 4–5x installed capacity. [Public: Latitude Media, Carbon Direct, 2026]
ERCOT large-load interconnection requests reached 230+ GW in 2025, up ~4x from 63 GW at end-2024. >70% from data center developers. [Public: Latitude Media, 2026]
PJM expects >30 GW demand increase 2024–2030, largely from data centers; new supply adds only 2–3 GW/yr vs. 5–7 GW/yr of DC demand. [Public: Utility Dive, 2026]

On-site / behind-the-meter (BTM) generation accelerating.

56% of developers exploring on-site/co-located generation. [Public: datacenterHawk, 2026]
AWS / Talen Energy: 17-year PPA for 1.92 GW from Susquehanna nuclear (PA), through 2042. [Public: 2024]
Equinix non-binding PPA for 250 MW of SMR capacity. [Public: industry reporting, 2025–2026]
CalEthos/TerraVolt (May 2026) signed nat-gas supply for 200–240 MW BTM plant for ID data center campus. [Public: SEC 8-K, 2026]
First SMR factory groundbreak in Oak Ridge, TN, planned 2026; target 50 reactors/yr by 2028. [Public: DCD, 2026]

2c. Unit Economics

Capex per MW (shell-and-core).

Industry average rose from $7.7M/MW (2020) to $10.7M/MW (2025); 2026 forecast ~$11.3M/MW. [Public: industry benchmarks via Archdesk / iRecruit, 2026]
Standard hyperscale 100 MW: ~$1.07–1.13B shell-and-core; ~$1.6–2.2B fully fit out (pre-server). [Public: iRecruit 2026 benchmarks]
AI-optimized 100 MW: $2B+ in construction alone (~$20M+/MW).
Tenant tech fit-out adds up to $25M/MW for AI infra.
Fully built-out AI campus (power, land, connectivity, equipment): $45–55B per GW. [Public: 2026 benchmark synthesis]

Revenue / pricing.

Wholesale colo (250–500 kW deployments, primary N. American markets): ~$195.94/kW/month average asking. [Public: datacenterHawk 2026 guide]
Equinix retail-style: $150–300/kW/month for additional power depending on metro. [Public: vendor pricing analyses, 2026]
US colocation market sized $72.37B by 2030. [Public: globenewswire databook, 2026-04]

GPU cloud / neocloud pricing (hourly, on-demand).

CoreWeave 8x H100 instance: ~~$49.24/hr (~~$6.16/GPU-hr); range $10–68.80/hr across SKUs. [Public: CoreWeave / costbench, 2026]
Lambda H100: $2.99/GPU-hr; B200: ~$5.50/GPU-hr. [Public: Lambda, 2026]
1-year reserved H100 pricing surged ~40% in six months: $1.70/hr (Oct 2025) → $2.35/hr (Mar 2026). [Public: SemiAnalysis newsletter, 2026]
B200 availability constrained in early 2026; many providers are reservation-only. [Public: Spheron / IntuitionLabs, 2026]

Returns / payback.

Hyperscale ground-up tier-1 US: 25–40% IRR over 3–4 year hold; development margins 50–65%. [Public: Accordant Investments, 2026]
Equity investors generally target 15–20% IRR; debt 6–8%.
Standalone colo hosting model: ~50-month payback, modest 2% IRR / 12.4% ROE (per one financial-model benchmark — variance is enormous depending on power cost / lease terms). [Public: financialmodelslab.com]

2d. Major Players

Hyperscalers (combined 2026 capex tracking $650–725B, ~75% AI-tied). [Public: CNBC, Tom's Hardware citing analyst estimates, Feb 2026]

Amazon: ~$200B
Google: ~$175–185B
Meta: ~$115–135B
Microsoft: ~$110–120B
Oracle: separate, but materially in the mix via Stargate-adjacent buildouts

Colocation / wholesale operators. Equinix (retail leader), Digital Realty (wholesale leader), QTS (Blackstone), CyrusOne, Iron Mountain Data Centers, NTT, Aligned, EdgeConneX, Switch (DigitalBridge).

Neoclouds. CoreWeave (public), Crusoe (lead developer of OpenAI’s Stargate Abilene TX campus), Lambda Labs, Nebius, Vultr. Neoclouds are growing aggressively but their listed prices are described as “unreliable” reference points because actual negotiated enterprise rates can be ~50% lower. [Interview: Mo Islam relaying prior contact, 2026-05-22]

Hyperscale facility footprint. 1,360 hyperscale DCs at end Q4 2025; pipeline of 770 additional facilities in planning/construction. Hyperscalers were 48% of global DC capacity; projected to reach 67% by 2031. [Public: Synergy Research, 2026]

Power & equipment suppliers.

Electrical: Vertiv (rack cooling/PDUs/UPS), Schneider Electric, Eaton, ABB, Siemens.
Backup generation: Cummins, Caterpillar, Generac, Rolls-Royce (mtu).
Backup batteries/UPS: Vertiv, Schneider (APC), Eaton.
Nuclear/SMR: NuScale, X-energy, Oklo, Kairos Power (each with hyperscaler-adjacent partnerships).
Liquid cooling specialists: CoolIT, Asetek, Iceotope, JetCool, Submer (immersion), GRC.

2e. Customer Landscape (Demand)

Enterprise IT and SaaS — historic colo demand, now backseat to AI.
AI training labs — Anthropic, OpenAI, xAI, Meta (in-house), Mistral, others. Anthropic approaching trillion-dollar valuation with no retail exposure — capital-markets pressure point. [Interview: Mo Islam, 2026-05-22; consistent with public reporting]
AI inference — distributed across hyperscaler clouds, neoclouds, and increasingly on-prem enterprise.
Crypto — diminished as a share; many former crypto miners (Crusoe being the canonical example) pivoted to AI.
HPC / national labs / research — DOE labs, weather modeling, pharma.
Government / sovereign AI — EU sovereign-cloud push, Stargate UAE, India compute strategy, DOD JWCC.

2f. Political / Regulatory Concerns

Grid strain is now the binding constraint, not chip availability. “Power infrastructure — not chips — will determine where AI data centre capacity gets built and who can access it at what cost.” [Public: Tom's Hardware citing analyst commentary, 2026]

Local zoning and water-use backlash.

A single large DC can use up to 5M gallons/day of water. [Public: Virginia Mercury, 2026]
Northern Virginia: Prince William County holds Occoquan watershed headwaters (~40% of NoVA drinking water).
VA 2026 General Assembly passed 15 DC bills including energy cost shifts and water reporting; statewide moratorium debated but not passed.
Moratorium bills spreading across multiple states in 2026. [Public: Good Jobs First, 2026]

State tax incentives under pressure. VA sales tax exemption for DC equipment under active legislative fight.

Federal AI infrastructure policy. Stargate (OpenAI / Oracle / SoftBank / Crusoe); CHIPS Act for fab capacity (separate but related).

China export controls on chips into US data centers. Restrictions on advanced GPUs going to China are well-documented; reverse direction (Chinese components entering US DC stack) is less covered publicly but came up via Nicole/NVIDIA in our internal interviews around components compliance. The compliance wedge has been killed for commercial direction — but the existence of these flows is relevant context for any DC operator who cares about supply-chain provenance.

Sovereign AI / EU data localization.

EU “Tech Sovereignty Package” presented May 2026: Cloud and AI Development Act (CADA) + Chips Act 2.0. [Public: Atlantic Council, 2026]
EU AI Act high-risk system enforcement live August 2, 2026.
Shift from “data residency” (where data sits) to “technical sovereignty” (who controls the stack).
US CLOUD Act extraterritorial reach unresolved — creates ongoing friction with EU customers regardless of physical server location. [Public: Orrick, 2026]

Environmental/climate disclosure. Increasing demand for Scope 1/2/3 emissions, water-use intensity (WUE), and PUE reporting. Likely accelerator for sensor-based measurement infrastructure (relevant to parametric measurement-agent role).

2g. Insurance / Financialization (Most relevant to Project TBD wedge)

Insurance market for DCs is strained and fragmenting.

Power supply is responsible for 45% of DC outages. [Public: Risk & Insurance, 2026]
Major tenants (Google, Meta) require 99.99–99.999% uptime; 15-second outages trigger significant SLA penalties. [Public: Hotaling Insurance, 2026]
FM Global 2026 loss-prevention guidance now requires 2-hour fire-resistance wall ratings due to Li-ion battery integration.
Baldwin report (Mar 2026): AI density + battery fire risk + fragmented coverage towers straining DC insurability. [Public: The Insurer, 2026-03]

Parametric insurance for DCs is an active and growing category.

Parametric structures already used for outage/voltage-drop triggers, paying automatically rather than requiring damage proof.
Used in construction / builder’s risk: Zurich launched a first-of-its-kind builders’ risk DC solution in 2025.
Industry literature explicitly advocates “lifecycle-based risk transfer strategies that incorporate parametric triggers, contingent business interruption and technology performance insurance.” [Public: Risk & Insurance, 2026]

Compute futures launched.

CME Group + Silicon Data announced first-in-class compute futures May 12, 2026; pending regulatory review for launch later in 2026. Based on Silicon Data’s daily GPU benchmarks for on-demand rental rates. [Public: CME press release, 2026-05-12; CNBC, 2026]
Explicit purpose: hedge against rising GPU rental rates and operational costs in the AI buildout. Closes part of the gap Mo Islam articulated independently to us 10 days later.

3. Convergences (Internal ↔ External Agreement)

NVIDIA reverse-logistics scaling crisis is real and underreported. Lonny’s first-person account is consistent with public NVIDIA SEC filings showing $8.22B in warranty reserves and the Compute & Networking segment dominating warranty accruals. The operational pain he described would explain those numbers.
Hyperscaler capex / GPU procurement is the supply-side anchor. Vivian’s qualitative framing (CSPs blocking capacity, competitive 2nd-sourcing) matches the public $650–725B 2026 capex tracking.
Power, not chips, is the limiting factor. Internal interviews (Max on power as “Nobel-Prize-worthy,” Lonny on the Dallas repair-line setup waiting on power) and external sources (PJM/ERCOT queues, transformer lead times, Vertiv/Eaton backlogs) converge cleanly.
Financial-product gap for compute/data-center risk is real and being filled. Preston’s parametric specialist + Mo Islam’s compute-index thesis (both internal, May 2026) align with the CME compute futures launch (external, May 12 2026) and the active parametric DC insurance literature. The financialization wedge is forming around us in real time.

4. Divergences (Internal ↔ External Disagreement)

Saturation framing. Josh (Apr 30) flagged the data center buildout as “saturated” and pointed to edge AI as the next cycle. External consensus (JLL, Synergy, hyperscaler capex announcements) treats the AI DC build as still mid-sprint with ~100 GW more capacity coming by 2030. Either Josh is contrarian-correct (priced-in / cyclical-peak) or external consensus is wrong; this is worth probing on a future call.
In-house vs. external tooling stance. Lonny said NVIDIA cannot build internally. Max (May 22) was skeptical that “the digital twin will ever happen because companies won’t share the data.” These are not strictly contradictory (one is “we need to buy,” the other is “data sharing is the bottleneck”), but they bound the addressable market differently. The IMEC neutral-broker analogy Max raised is the structural bridge.
Compliance pain in DC stack. Vivian said compliance is “not a pain point” for big companies because they have in-house counsel — yet US export controls on components going into AI DCs are an active regulatory area. The compliance wedge has been killed for our purposes (commercially), so this divergence is recorded but not actionable.

5. Open Questions (Frame as Probes, Not Conclusions)

Directly tied to financialization wedge

Do DC operators currently hedge GPU price or availability risk, and how? Forward purchase commitments? Take-or-pay contracts with NVIDIA? Once CME compute futures launch, who is the natural first buyer — a neocloud locking in its sell-side price, or a hyperscaler hedging its buy-side cost? Who could answer: Silicon Data team, CME product manager, CoreWeave/Crusoe CFO, a hyperscaler procurement lead.
Where is the natural parametric trigger in a data center? Power outage minutes? Temperature excursion beyond rack threshold? GPU failure rate above a sensor-defined baseline? Which of these already has third-party measurement infrastructure that could serve as a credible measurement agent? Who could answer: Parametrix, Vertiv’s data-services group, an FM Global DC underwriter, Preston’s parametric specialist (re-engage).
What is the actual structure of a DC business-interruption claim today, and what is the gap parametric fills? Specifically for AI workloads where contractual SLA penalties cascade across tenants. Who could answer: Hotaling Insurance, a Marsh/Aon DC broker, a colo operator’s risk manager.
Is there a tradable inventory-hedge structure for GPUs analogous to the LME warehouse model? Given (a) secondary market for H100/A100 exists with $8–18K range pricing, (b) NVIDIA’s advance-replacement ARMA program, (c) 1-yr H100 rental price up 40% in 6 months, the conditions for a physical-hedge product look more present than ever. Who could answer: ALTA Technologies, a GPU secondary-market broker, a neocloud treasurer.
Can NVIDIA’s warranty reserve become a reinsurance opportunity? Max independently surfaced this. Lonny is on the operations side of the same problem. Is the answer a warranty-risk MGA selling into Munich/Swiss Re, or a reverse-logistics SaaS that prices the data feed into the warranty claim curve? Who could answer: Greg DeLoccio (Lonny’s intro), an NVIDIA finance/treasury contact, a specialty warranty reinsurer.

Adjacent but worth tracking

What is the actual unit economics of a 100 MW AI DC at current GPU densities? Public benchmarks give $11.3M/MW shell, ~$20M+/MW AI-optimized, ~$25M/MW tenant fit-out — but rack-power densities are rising so fast (132kW → 600kW projected) that the per-MW base may be the wrong denominator. Who could answer: JLL data center capital markets team, CBRE DC research, Brett’s NGP intro (still un-actioned per our vault).
How material is the equipment-shortage delay risk to project completion timelines? If >50% of 2026-planned US DCs are at risk of delay/cancellation due to transformer/switchgear shortages, the delay itself becomes an insurable event for project finance lenders. Worth probing as a parametric construction product. Who could answer: Zurich builders’ risk team, a project-finance lender to a DC developer.
What does the sovereign-AI dynamic do to DC location decisions? If EU customers can no longer trust US-headquartered cloud providers post-CLOUD Act / CADA, does that create demand for sovereign-DC operators (Schwarz Group, OVH, Scaleway) that don’t currently buy from the same supply chain? Implications for our financial-product market sizing. Who could answer: A European specialty insurer (Munich Re/Swiss Re sovereign-cloud desk), an EU DC operator.
Is Josh right that DC buildout is saturated? Worth pressure-testing with somebody who would short the trade. Who could answer: A hedge fund analyst covering DC REITs, a debt analyst covering DigitalBridge/QTS, Mo Islam (he is in this circle).

Internal pipeline actions

Re-engage Brett’s NGP/data center intro — never actioned in the vault.
Greg DeLoccio intro from Lonny is the highest-leverage next conversation in the reverse-logistics direction.
Connect to a colocation operator (Equinix, Digital Realty, QTS, Aligned). We have zero direct interviews here — the largest gap in our DC coverage.

6. Confidence Summary

Topic	Internal Signal	External Signal	Confidence
NVIDIA reverse logistics scaling crisis	Strong (Lonny, primary)	Strong (SEC filings, warranty data)	High
AI DC supply chain structure (CSP → NVIDIA → ODM → components)	Moderate (Vivian, qualitative)	Strong (Synergy, JLL)	High
Hyperscaler 2026 capex ~$650–725B	None direct	Strong (multiple analysts, earnings)	High but consensus-dependent
Power as the binding constraint	Moderate (Max, Lonny indirect)	Strong (PJM/ERCOT queues, Vertiv/Eaton backlogs)	High
Rack densities 100–600kW trajectory	None direct	Strong (NVIDIA roadmap, OEM disclosures)	Moderate-High (vendor-claim-heavy)
Capex ~$11–25M/MW range	None direct	Moderate (benchmark databases, wide variance)	Moderate
Colocation pricing ~$150–300/kW/mo	None direct	Moderate (vendor self-reporting, broker estimates)	Moderate
Parametric / financial product gap in DC risk	Strong (Preston, Mo, Max)	Strong (CME launch, parametric DC literature)	High
Saturation vs. continued buildout	Mixed (Josh contrarian)	Strongly bullish (consensus)	Mixed — surface as open question
China export controls / sovereign AI impact on DC	Weak (compliance wedge killed)	Moderate (EU CADA, CLOUD Act)	Low — preserve as record only
Internal DC operator (colo, neocloud) perspective	Zero	Available externally	Critical gap

Overall epistemic state: We have one strong internal anchor (Lonny/NVIDIA reverse logistics — supplier side) and two structural anchors that intersect the financialization wedge (Preston on parametric, Mo on compute pricing). We have no direct conversations with the actual DC operator buyer archetype — colocation, neocloud, or hyperscaler infrastructure. External research suggests the financialization opportunity in the DC stack is forming right now (CME compute futures launching this year; parametric DC insurance growing; warranty reinsurance gap visible in NVIDIA filings) — which means the window to insert proprietary data infrastructure or measurement-agent positioning is open but narrowing as incumbents move.

Internal files referenced

memory/interviews/2026-05-12-lonny-orona.md
memory/interviews/2026-04-29-vivian-bliss-dustin.md
memory/interviews/2026-05-22-prestonbliss.md
memory/interviews/2026-05-22-mo-islam.md
memory/interviews/2026-05-22-max-dustin-bliss-1-1.md
memory/interviews/2026-04-30-josh-bliss-dustin.md
memory/interviews/2026-04-30-brett-bliss-dustin.md (NGP/DC intro lead, unactioned)
memory/synthesis/reverse-supply-chain-research-2026-05-13.md
memory/synthesis/glencore-of-semiconductors-2026-05-13.md
memory/synthesis/briefing-max-mirgoli-2026-05-15.md

Project TBD Memory Vault

Explorer

data-centers-research-2026-05-24