The Evidence Machine
We built the test from the failures.
Every data-center marketing document is built the same way. Design a facility, then hunt for evidence that supports the design. We inverted that order.
Fourteen honest answers.
We built the test from the failures
Every data-center marketing document is built the same way. A company designs a facility, then hunts for evidence that supports the design. Zero water. Self-generation. Smaller footprint. Community benefits. A list of attributes in search of a problem.
We inverted that order.
We built the research infrastructure first. The Delay Watchlist and the Moratorium Tracker continuously record what has actually stopped, delayed, downsized, been litigated, or been cancelled across the United States. Generator interconnection queues, utility filings, county and municipal ordinances, public disclosures and cluster studies feed one typed evidence base. It refreshes every Monday at 06:00 Central.
Then we asked a different question. What happens if we run our own architecture through the same failures that are stopping everyone else?
Not a benchmark we invented. Not a marketing score. A matrix where each row is a documented failure mode, each result is scored against the record, and every mitigation names the residual risk it leaves behind.
This page is that matrix, the evidence under it, and the things it says we have not solved.
The failure surface
Before scoring anything, here is the surface we are testing against. These are not projections. They are conditions in the record right now.
Interconnection
- The typical wait from interconnection request to commercial operation has stretched from under two years to roughly five, and only about one in five projects that requested interconnection between 2000 and 2019 had reached commercial operation by the end of 2024 (Berkeley Lab · Queued Up).
- AEP publicly raised the prospect of leaving PJM and SPP in May 2026, citing an inability to serve contracted data-center load at interconnection speed (Energy Central).
Long-lead equipment
- Large power transformers now run past two years, with generator step-up units longer still, against demand growth the manufacturing base has not matched (POWER Magazine).
- Gas-turbine backlogs have pushed delivery years out. TVA has said it must now order turbines roughly five years before delivery and seven before operation (American Public Power Association; Chattanooga Times Free Press).
Water and cooling
- A majority of US data centers still reject heat evaporatively, which is precisely the design choice drawing regulatory and community pressure (USA Today).
- The industry frames this as a tradeoff between water and energy. It is treated that way because most designs have not removed the freshwater draw (E&E News; ITIF).
Community and policy
- Gallup found 71 percent of Americans oppose a data center being built in their area, against 53 percent for a nuclear plant. Opposition to nuclear has never been that high in the years Gallup has asked (Gallup).
- Hundreds of localities have restricted new construction, and multiple states have taken up moratorium legislation (Rockefeller Institute of Government; NCSL).
Litigation
- A wave of nuisance, environmental and land-use litigation is now aimed at data centers and their on-site power, and it is being written up as an emerging practice area rather than as isolated disputes (WilmerHale; Business Law Today, ABA).
Labor
- The Bureau of Labor Statistics projects roughly 81,000 electrician openings a year through 2034, most of them from workers retiring or moving on rather than from growth alone (US Bureau of Labor Statistics).
- Electrical scope is a large share of total build cost, which makes craft-labor availability a schedule risk and not only a cost line (Fortune; WIRED).
That is the test. It is not theoretical, and we did not choose it.
Fourteen rows, scored against the record
Each row below is a failure mode from the record above. Each result is scored against our architecture, and every green mark carries the residual it does not remove. Filter by result to see what we claim and what we do not.
The architecture materially lowers exposure through a design decision we can point to. Residual risk remains and is described.
The architecture mitigates one dimension and leaves another intact. Treated as a live engineering requirement, not a win.
We carry the same exposure the industry carries. Named rather than dressed up.
Economic exposure that has to be modeled per site rather than passed or failed globally.
01 Utility interconnection waitCritical path, and usually a prerequisite to commercial operation REDUCED
On-site primary generation with fuel cell, reformer and storage on a common collection bus. The grid becomes optional import rather than a commissioning dependency.
Not queue-gated for first energization. A grid tie remains available as economic optionality, and any grid-parallel operation still needs a protection scheme.
02 Transmission upgrade after the interconnection agreementOften a prerequisite, and utility-scheduled REDUCED
On-site generation removes the requirement to import baseload. Local distribution only.
Not eliminated. Any grid-parallel operation still requires a protection scheme and a utility relationship.
03 Large power transformer and generator step-up lead timeLong-lead, single-point critical path PARTIAL
Distributed generation reduces dependence on generator step-up class equipment. Conversion happens at the block-level DC bus rather than at a substation.
We still consume distribution-class transformers, medium-voltage step-down where used, and DC conversion units. Step-up risk is substantially removed. Medium-voltage switchgear risk remains and is engineered as a critical-path item from day one.
04 Gas turbine lead timeDirect exposure for anyone self-generating with combustion turbines REDUCED
Primary power is fuel cell, reformer and storage rather than combustion turbines. No turbine slot sits between us and first power.
Fuel-cell supply becomes the new critical path. It is engineered explicitly, with integration completed before shipment.
05 Medium-voltage switchgear lead timeLong-lead item, often paired late with generation PARTIAL
Modular factory-preintegrated blocks. The medium-voltage scope inside each block is scoped, released and slot-locked at design freeze.
The lead time exists. We mitigate through procurement sequencing and by minimizing switchgear count per megawatt. It is not eliminated.
06 Water and evaporative cooling oppositionGrowing regulatory and social risk across many states REDUCED
Direct-to-chip liquid cooling with dry-cooler heat rejection. Closed loop, with a product-water reuse cascade from the reformer.
Zero freshwater draw for cooling in the primary case. A local regulator must still issue a written product-water reuse determination. That is not automatic.
07 Community oppositionGrowing, and demonstrably fatal to projects REDUCED
A host-partner campus model. Training, community programming and workforce development sit at the front of the site, and the host institution is a partner rather than a landlord.
Only works where a legitimate host institution wants the compute and workforce package. It is not a substitute for a bad site.
08 State and local moratoriumsSite kill risk after capital is committed REDUCED
The Delay Watchlist and Moratorium Tracker refresh the map continuously. Site selection is evidence-driven rather than sales-driven.
Selection risk is measurable before commitment. We cannot immunize against a jurisdiction that adopts a moratorium after commitment.
09 Zoning and permitting, per siteSite-dependent and frequently fatal EXPOSED
The same statutory requirements apply to us. A smaller modular footprint is easier to place and improves community fit.
Zoning still applies. Any jurisdiction can block a rezoning. We are not exempt and we do not claim to be.
10 Nuisance, open-meetings and air-permit litigationGrowing, especially against gas-turbine power islands PARTIAL
Fuel-cell primary generation removes the specific fact patterns behind turbine-noise and unpermitted-turbine emissions cases. A community-first campus reduces the standing base.
Air-emissions and noise attack surfaces are materially reduced. Open-meetings and public-notice challenges are procedural and cannot be architected away.
11 Skilled labor availabilityIndustry-wide, with electrical scope a large share of build cost PARTIAL
A training institute at each campus, designed as workforce production rather than as marketing.
The institute produces future capacity, not current-quarter crews. Near-term projects still compete in the same craft-labor market. This is mitigation, not a solution.
12 Upstream supply chain concentrationStructural, with no near-term relief PARTIAL
Vertical integration across the power block and compute block, with factory pre-integration before shipment, reduces on-site trade coordination.
Coordination risk drops. Raw-material concentration in electrical steel and copper hits us the same way it hits everyone.
13 Incentive and tax-credit volatilityJurisdiction-dependent, and model-breaking if assumed QUANTIFY
Economics are tested without aggressive credit stacking. Base cases are modeled without adders, and credits are treated as upside rather than as structure.
Not scored pass or fail. Exposure is quantified per jurisdiction and per statute, in the model rather than on a slide.
14 Post-approval delayProjects now spend more time between agreement and operation than in the queue REDUCED
On-site generation collapses the approval-to-operation leg for first-block energization.
True for the primary energization path. Any subsequent grid-parallel operation or export adds the exposure back.
Seven of fourteen reduced. Five partial. One exposed. One to quantify. Those counts move when the evidence base moves, not when a deck needs refreshing.
What we do not claim
This is the section most infrastructure documents do not have, and it is the reason to trust the rest of the page.
Seven of the fourteen rows are not wins. One is a straightforward exposure. Five are partial mitigations with the untouched dimension named. One is an economic question that has to be answered per site rather than claimed globally.
Zoning can still stop us
Row 9 is marked exposed. A smaller footprint places more easily and reads better in a hearing, but the statute applies to us exactly as it applies to anyone else, and a jurisdiction can decline. No architecture removes that.
Switchgear and transformers are still long-lead
Rows 3 and 5. We remove generator step-up class exposure and reduce switchgear count per megawatt. We still buy distribution-class equipment in a constrained market, and we treat it as critical path from day one rather than discovering it late.
We compete for the same electricians
Row 11. A training institute is real workforce production, and it produces capacity on a multi-year horizon. It does not staff next quarter's job. Near-term builds compete in the same craft market as everyone else.
Procedural litigation cannot be designed away
Row 10. Removing combustion turbines removes specific noise and emissions fact patterns. Open-meetings and public-notice challenges are procedural, and any project can be delayed on procedure.
Raw material concentration hits us too
Row 12. Vertical integration reduces coordination risk between trades. It does not create grain-oriented electrical steel or copper that the market does not have.
Incentives are modeled, not assumed
Row 13. We do not score this pass or fail. Base cases are built without aggressive credit stacking so the project stands on its own economics, and credits are quantified per jurisdiction as upside.
The loop underneath it
The matrix is not the point. What produces the matrix is the point, and what happens when the world produces a failure the matrix has not yet ingested.
Nothing in the loop below is aspirational. Every stage runs today.
The same discipline runs on the software side. Our own model work is gated the same way: a capability is not a capability until it clears a held-out evaluation, and we have retracted our own results when the measuring instrument turned out to be wrong. That story is written up in The Model Was Fine. The Instrument Was Not.
One loop tests the physical factory against reality. The other tests the intelligence layer against the business. Same rule underneath both: do not convert a hypothesis into a capability claim.
Adjectives and ledgers
There are two ways to sell infrastructure.
Most vendors sell with adjectives. Seamless. Resilient. Turnkey. Best-in-class. Every one of those is a promise with no artifact behind it, and the counterparty has no way to check it short of committing capital.
An evidence base sells with ledgers.
The difference is not tone. Adjective-driven vendors are asking to be believed. We are asking to be checked.
Buyers who have watched projects delay, downsize and die respond better to being checked than to being reassured.
Test the claims yourself
Everything on this page is meant to be verifiable. The fastest ways to test it:
Read the trackers
The Delay Watchlist and the Moratorium Tracker are the evidence base behind the failure surface above. They refresh every Monday.
Open the trackers →Follow the sources
Every figure on this page links to the primary document, filing, agency or report that states it. Nothing here rests on our own summary of someone else's work.
See the research hub →Test a specific site
The useful version of this matrix is the one run against your jurisdiction, your interconnection class and your schedule. That is a working session, not a brochure.
Start a working session →We did not design a factory and then look for evidence that supported it. We built a system that continuously tests the architecture against the evidence, and publishes what the test says, including the rows we do not pass.
When the record changes, the score changes. That is the point of building it this way.
Want the next one?
When a new piece publishes on SAVRN Insights, you get one email with what it covers and a link to read it. No digests, no promotions.
One email when it publishes. Unsubscribe in one click. Privacy
Frequently asked questions
What is the architecture stress test?
A standing matrix of fourteen documented data-center failure modes, each scored against SAVRN's architecture. Every row names the failure observed in the public record, the conventional exposure, our design response, and the residual risk our response does not remove.
Where does the evidence come from?
From the SAVRN Delay Watchlist and Moratorium Tracker, which continuously ingest generator interconnection queues, utility filings, county and municipal ordinances, public disclosures and litigation dockets. They refresh every Monday at 06:00 Central. Every external figure on this page also links to its own primary source.
What do REDUCED, PARTIAL, EXPOSED and QUANTIFY mean?
REDUCED means a design decision materially lowers exposure, with the residual named. PARTIAL means one dimension is mitigated and another is left intact, so it stays a live engineering requirement. EXPOSED means we carry the same risk the industry carries. QUANTIFY means the exposure is economic and has to be modeled per site rather than passed or failed globally. Nothing is scored as solved.
Why publish the rows you do not pass?
Because a scorecard with no losses is marketing. Seven of the fourteen rows carry residual risk, and those are published as prominently as the seven that do not. A buyer who has watched projects fail can tell the difference between a document that is trying to be believed and one that is trying to be checked.
Does on-site generation actually remove interconnection risk?
It removes the queue from the critical path for first energization, which is a different claim from removing it entirely. A grid tie remains available as economic optionality, and any grid-parallel operation still requires a protection scheme and a utility relationship. That residual is stated in rows 1, 2 and 14.
Is the zero-water claim absolute?
For cooling in the primary case, yes: direct-to-chip liquid cooling with dry-cooler heat rejection means no freshwater draw. The residual is regulatory rather than thermal. A local regulator still has to issue a written determination on product-water reuse, and that is not automatic.
How often does the score change?
Whenever the evidence base changes. The trackers refresh Mondays at 06:00 Central, and a new class of failure that the matrix has not seen becomes an engineering question rather than a footnote. The counts move because the world moves.
How is this different from a capabilities deck?
A capabilities deck lists attributes. This lists failures that actually happened to other people's projects, then shows what our architecture does about each one and what it does not do. The organizing unit is the failure, not the feature.
Does the same discipline apply to SAVRN's software?
Yes, and it has cost us. Our own model work is gated on held-out evaluation, and we retracted a promoted result after auditing the instrument that produced it. That is documented in The Model Was Fine. The Instrument Was Not.
Can I see this run against my own project?
That is the useful version. The matrix on this page is generic; run against a specific jurisdiction, interconnection class and schedule, it produces a distribution rather than an adjective. That is a working session, and you can start one from the section above.
Every source behind this page
16 outside sources, each linked to the primary document, agency or report that states it.
Interconnection and grid
Equipment and supply
Water and cooling
Community, policy and litigation
- Gallup · Americans oppose AI data centers in their area (71% vs 53% for nuclear)
- Rockefeller Institute of Government · Moratoriums on data centers
- NCSL · Which states are considering data-center restrictions
- WilmerHale · Data centers in court: nuisance, environmental and land-use litigation
- Business Law Today (ABA) · Emerging data-center litigation
Labor
Every external figure on this page links to the primary document, agency or report that states it. Where a widely quoted number was traceable only to secondary coverage, we went to the source and used its wording instead. The Bureau of Labor Statistics projects roughly 81,000 electrician openings a year, most from retirement and transfer rather than growth alone, which is a different statement from an equivalent number of unfilled posts; we use the agency's framing. The Gallup finding is opposition to a data center in the respondent's own area, not opposition in general. The stress-test scoring, the residual descriptions and the architecture claims are SAVRN's own.