Ask before the deep-dive
Convert claims into artifact requests before the technical deep-dive. The checklist is cheap to send, and the response time is itself data.
INS-01Procurement method · Analysis
How to evaluate AI infrastructure claims
Most AI-infrastructure claims are not false. They are unfalsifiable as written, because no artifact is named that could settle them. The useful question is therefore not whether a claim is true but which document would prove it, and whether that document exists yet. This page turns that into an eight-domain checklist and runs it on PODOS.
PUBLISHED LAST VERIFIED BY JOSEF ELIMELECHREVIEWED PODOS AI ENGINEERING
- 8
- Evidence domains, each borrowing an existing scale
- 7.8 kW
- Screening ratio: 1 MW designed ÷ 128 designed GPUs
- 6th
- Consecutive year average PUE barely moved (Uptime 2025)
The short answer
Evaluate an AI-infrastructure claim by asking which document would settle it, then asking whether that document exists yet. Three things establish it. The qualifier — designed, target, measured, or third-party verified. The boundary — what sat inside the measurement, over what period, and what was excluded. The artifact — the dated test report, interconnection study milestone, in-plant inspection approval, published benchmark entry with system description and code, or patent number a buyer can pull without asking the vendor.
Eight domains cover nearly everything that matters: system maturity, grid interconnection, manufacturing readiness, performance benchmarks, efficiency, intellectual property, references, and capital. Each borrows its scale from a discipline that already solved the disclosure problem, so the buyer is not inventing a standard mid-negotiation. Where the artifact does not exist yet, the accurate label is unverified — not false, and not silence. The checklist below is the working form of that, and the last two sections state plainly what it cannot settle.
First principles
The failure mode is disclosure, not deception
The claims that cost money are rarely lies. They are statements whose truth conditions were never specified: a capacity figure that never says whether it is a service rating or an operating load, a timeline that never says which permits it assumes, an efficiency ratio that never says what sat inside the boundary. Neither party's belief is testable.
The fix is not more scepticism. Every domain below belongs to a discipline that already solved this. Spaceflight has readiness levels.[1] Machine learning has divisions, availability categories, and published code.[5][7] Efficiency has an ISO/IEC standard with measurement categories.[8] Off-site construction has an inspection standard naming who verifies the plant.[4] Grid connection has a regulated queue with study milestones.[2] Patents have numbers.[11] The checklist makes the borrowing explicit, so buyers ask for the artifact, not the adjective.
It matters more now because the aggregate has stopped moving. Uptime Institute's 2025 global survey reports average PUE showing little change for the sixth consecutive year, with improvement constrained by legacy infrastructure and region-specific barriers to efficient cooling.[9] Against that, claims of large individual gains need individual evidence.
Wordings that need a follow-up
- A number without its qualifier: designed, target, measured, or third-party verified.
- A render standing in where a photograph would settle the question.
- “Up to” without the condition that produces the maximum.
- A benchmark quoted without its baseline, hardware, or date.
- A timeline that silently excludes the interconnection and the permit.
- A patent claim count offered in place of a patent number.
Table 1 · The evidence checklist
Eight domains, and the artifact that settles each
Eight domains: the question, the shape of a checkable answer, and the artifact that settles it.
| # | Domain | Ask | What a good answer looks like | The artifact that settles it |
|---|---|---|---|---|
| EC-01 | System maturity | At what readiness level, in which environment? | A level on a named scale. NASA's ladder separates a laboratory proof-of-concept (TRL 3) from a functional prototype (TRL 6) and a system proven in operation (TRL 9).[1] | Test report with date, environment, and witnesses; photographs of the article tested, not a render of the one designed. |
| EC-02 | Grid interconnection | For this site, who owns the energization date? | A named utility or ISO, a queue position, the study milestone reached, and a date in writing. Under FERC's first-ready, first-served cluster process, a queue position is not a study result, and a study result is not an executed agreement.[2][3] | Interconnection request receipt, cluster study assignment, executed interconnection or service agreement. |
| EC-03 | Manufacturing readiness | Who inspects the factory, and who accepts the unit on site? | A named third-party in-plant inspection agency and an acceptance path with the destination authority having jurisdiction. ICC/MBI 1200 covers the plant's quality process; 1205 covers who verifies it.[4] | Inspection agency approval, plant audit reports, AHJ correspondence for the destination. |
| EC-04 | Performance benchmarks | Could someone outside your company reproduce this? | A submission shaped like MLPerf's Closed division: the reference model unmodified, a published system description, a link to the code, and an availability category saying whether the hardware can be bought — Available, Preview, or research/internal.[5][6][7] | A published result entry with system description and code, or full disclosure of model, dataset, precision, hardware, baseline, and date. |
| EC-05 | Efficiency claims | Designed or measured — at which category, boundary, and period? | ISO/IEC 30134-2 defines PUE with measurement categories precisely because a bare ratio is not comparable. Name the category, the boundary, the averaging period, and whether the figure is design intent or measurement.[8][10] | A metering plan mapped to the standard's categories, plus the reporting period. Google publishes a fleet-wide trailing-twelve-month PUE — the shape of a checkable number. |
| EC-06 | Intellectual property | Which numbers? | Publication numbers for pending applications and patent numbers for grants, with jurisdiction and status. A claim count is a filing statistic, not a right; pending is not granted.[11] | Records the buyer pulls from USPTO Patent Public Search, plus assignment records showing current ownership. |
| EC-07 | References | Who has operated this, and may we call them? | A named reference willing to take a call, or an anonymised deployment given by size, date, and workload with an operator reachable under NDA. Confidentiality explains a missing name, not a missing deployment. | A reference call, a site visit, or an agreement given by scope and execution date — distinguished from a non-binding expression of interest or an exploratory discussion. |
| EC-08 | Capital and risk | What does the next tranche buy, and what stops without it? | A use-of-proceeds statement tied to milestones, and the company's own list of its likeliest failure modes. A vendor who cannot name its top three risks has not looked or will not say.[12] | Offering documents. For a US private offering, the Form D notice is public on EDGAR, checkable without asking the company. |
Figure 1 · Original analysis
A worked cross-check: provisioned kilowatts per accelerator
The screening ratio, and where the PODOS Pod's own published design numbers land on the judgement bands.
One piece of arithmetic catches internally inconsistent specifications: divide a unit's nameplate electrical capacity by the accelerators it is designed to hold. Applied to the PODOS Pod, which is designed as a standardized 1 MW building block and designed for 128 GPUs:
1,000 kW ÷ 128 accelerators = 7.8 kW provisioned per accelerator
7.8 kW · designed
Below ~2 kW~3–10 kWAbove ~15 kW
Reading the result
What the ratio can and cannot tell you
Assumptions, stated so they can be rejected: the numerator is designed service capacity, not measured load; the denominator is a design population, not a shipped configuration; and the ratio charges all provisioned power to accelerators alone.
That last assumption is deliberately unfair, which is the point. Accelerators are the largest single load but never the only one — NVIDIA ships the GB200 NVL72 as 72 GPUs alongside 36 CPUs in one liquid-cooled rack,[13]and a unit adds network, storage, pumps, fans, conversion losses, and headroom. A healthy ratio therefore sits well above one device's rated draw.
The ratio never proves a design is sound. It tells you whether two published numbers can both be true at once — and a vendor who cannot explain their own has told you how the specification was assembled. It applies across high-density GPU infrastructure, where the binding constraint is usually thermal; see direct-to-chip liquid cooling for what removes the heat.
Table 2 · The bands behind Figure 1
Three ratios, and the follow-up each one earns
The bands below are judgement, not specification.
| Ratio | What it suggests | Follow-up question |
|---|---|---|
| Below ~2 kW | Accelerators alone would consume nearly all provisioned capacity, leaving nothing for balance of system. | Is the nameplate an IT-load figure presented as a service rating, or is the accelerator count aspirational? |
| ~3–10 kW | Sized for the whole system plus headroom — the ordinary shape for a dense liquid-cooled unit. | What sustained load, as a fraction of nameplate, is the cooling designed to remove continuously? |
| Above ~15 kW | Either deliberate oversizing for future refreshes, or capacity the thermal design cannot use. | Which limit binds first — is the unit heat-limited or power-limited? |
Ask which document would settle the claim, then ask whether that document exists yet.
EC-02
The row that binds hardest — grid interconnection
Practice
What this means for operators
Power leads the schedule
Sequence power first. The interconnection milestone usually sets the project date, and a utility owns it — not the vendor.
Keep the qualifier attached
Make the qualifier travel with the number, in the document rather than the meeting. A figure that loses it gets quoted back as a commitment.
Re-run it at every milestone
Write the artifact column into the RFP, then re-run it at each milestone. Which artifacts arrive on schedule beats how impressive any single one is.
Honest limits
What this does not prove
- A checklist tests disclosure, not truth. A vendor can produce every artifact and still miss the schedule; another can be excellent and document badly. Passing is a floor, not a verdict.
- A missing artifact is not a missing capability. Early-stage companies legitimately hold test data confidential and are bound by NDAs. The inference is that a claim is unverified, not false.
- The maturity scales are borrowed by analogy. TRL was written for spaceflight and MLCommons' divisions for machine-learning systems; neither certifies modular units.
- The kilowatts-per-accelerator bands are heuristic. No standard defines them, and a unit can sit outside them for legitimate reasons — oversizing for future refreshes being the obvious one.
- None of the cited sources evaluate any vendor. IEA, FERC, ISO/IEC, MLCommons, ICC/MBI, Uptime, USPTO, and the SEC describe processes and lookup tools; they endorse nobody.
- This is a method for evaluating engineering claims. It is not investment, legal, or tax advice, and figures identified as targets or estimates are not guarantees.
Applied to us
Running the checklist on PODOS
A checklist a company will not apply to itself is marketing. The PODOS Pod is designed as a standardized 1 MW building block and designed for 128 GPUs, and PODOS targets a 90-day window from order to commissioning for a standard unit. All three are labelled as targets where they appear. Under EC-01 they are specification, not demonstration.
What is not published matters more. PODOS has issued no readiness-level statement or third-party test report (EC-01); no benchmark for the pod or for Syntropic with disclosed model, hardware, baseline, and method (EC-04); no named in-plant inspection agency (EC-03); no PUE at any measurement category (EC-05); no application or grant numbers (EC-06); and no operating reference (EC-07). The investor page is at interest stage with no offering terms (EC-08). Mark those rows unanswered, because they are.
EC-02 binds hardest. A 90-day manufacturing target is a claim about a factory, not a utility. The IEA reports grid-connection bottlenecks tightening as data-centre electricity use surges,[3]and FERC's cluster process sets the milestones a request must clear before anyone can promise an energization date.[2] Every vendor timeline, this one included, depends on a study result nobody in the sales conversation controls — which is why site power readiness is scoped separately from the unit and the readiness checklist starts with the interconnection.
The AI infrastructure glossary defines the recurring terms; modular versus traditional AI data centers covers the differences this keeps surfacing. To size a configuration first, use the configurator.
Sources
- [1] Technology Readiness Levels (TRL 1–9 definitions) — NASA, accessed 2026-08-31
- [2] Explainer on the Interconnection Final Rule (Order No. 2023) — Federal Energy Regulatory Commission, final rule issued 28 Jul 2023
- [3] News: Data centre electricity use surged in 2025, even with tightening bottlenecks — IEA, accessed 2026-08-31
- [4] New brief explores implementation of ICC/MBI Standards 1200 and 1205 for off-site construction — International Code Council / Modular Building Institute, standards 2021 ed.
- [5] MLPerf Inference: Datacenter — divisions, availability categories, required disclosures — MLCommons, accessed 2026-08-31
- [6] General MLPerf Submission Rules (submission_rules.adoc) — MLCommons, accessed 2026-08-31
- [7] MLPerf Inference Benchmark (Reddi et al.), arXiv:1911.02549 — arXiv, 2019
- [8] ISO/IEC 30134-2:2016 — Data centres, key performance indicators, Part 2: Power usage effectiveness (PUE) — ISO/IEC, 2016
- [9] Global Data Center Survey 2025 — average PUE little changed for the sixth consecutive year — Uptime Institute Intelligence, 2025
- [10] Data center efficiency (fleet-wide trailing-twelve-month PUE) — Google, accessed 2026-08-31
- [11] Patent search tools — Patent Public Search, Patent Center, Assignment Center — USPTO, accessed 2026-08-31
- [12] EDGAR Full-Text Search (Form D notices of exempt offering) — U.S. Securities and Exchange Commission, accessed 2026-08-31
- [13] GB200 NVL72 product page — NVIDIA, accessed 2026-08-31
Send us the checklist and hold us to it
Bring the eight domains to the first call. Engineering will answer the rows it can and mark the rest unanswered.