T2C / Methodology
The single most common error in AI infrastructure coverage is treating three different quantities as interchangeable. T2C keeps them apart at the data level: they are separate fields, they are never added together, and the build fails if an aggregate mixes them.
Power contracted at the utility connection, before conversion, cooling and distribution losses. The largest number a company can quote, and the least indicative of compute delivered.
Power available to IT equipment inside the hall. The industry's comparable measure of usable data-centre capacity, and typically well below the gross utility figure for the same site.
Power drawn by accelerators specifically, excluding other IT and infrastructure load.
The second error is treating an ambition as a capability. A company can hold a 5 GW year-end target, a 2.2 GW development pipeline and 648 MW of secured power simultaneously — and only one of those is capacity it controls today.
A current, operating figure as at the stated date.
At least this much, from the components the company has itemised. The company has not published an exhaustive total, so the real figure may be higher.
A management goal for a future date. An intention, not a capability that exists today, and never counted in a current-capacity total.
Identified development opportunity — sites under study, options held, applications lodged. Not secured and not built.
A theoretical ceiling, typically phrased "up to". Conditional on approvals, financing and demand that do not yet exist.
Where a company itemises individual contracts but never publishes a company total, the sum of the
items is a disclosed minimum, shown with a ≥. IREN discloses 200 MW to Microsoft
and 60 MW to NVIDIA, plus further contracts worth approximately $2.8bn whose megawatts are not
disclosed — so 260 MW is a floor, not a total. Applied Digital, by contrast, publishes an exhaustive
1,410 MW across five campuses, and that figure is exact.
| Gate | Meaning |
|---|---|
| Site control | Ownership, lease or option over the land secured. |
| Utility agreement | Executed electric service or energy-services agreement with the supplier. |
| Interconnection approved | Grid interconnection study complete and connection approved. |
| Zoning approved | Local land-use consent granted for data-centre use. |
| Regulatory approval | Federal or state authorisation for the transaction or the asset — a FERC order, a state utility commission approval. Distinct from zoning and from environmental permits, and frequently narrower than it appears: approval to acquire a power plant is not approval to build a data centre on it. |
| Environmental approval | Environmental permits and assessments granted. |
| Building permits | Construction permits issued. |
| Financing committed | Capital committed and available to build. |
| Long-lead equipment ordered | Transformers, switchgear, generators and chillers on order. |
| Construction started | Physical construction under way on site. |
| Utility energised | Grid power delivered to the site. |
| Critical IT energised | IT load live and drawing power in the hall. |
| Customer contracted | A paying customer has signed for the capacity. |
| Customer accepted | Capacity formally accepted under the customer agreement. |
| Billing / revenue commenced | The company has disclosed that billing or revenue generation has begun. |
Power capacity controlled through executed utility, interconnection, energy-service, lease or comparable arrangements. It may include capacity already operating as well as capacity awaiting development. Secured power indicates supply control, not customer demand or completed compute capacity.
Capacity formally accepted under the relevant customer agreement. Acceptance may trigger billing or revenue recognition, but the commercial effect is contract-specific and must be separately evidenced.
Capacity for which the company has explicitly disclosed that billing, rent commencement or revenue generation has begun. Billing, GAAP revenue recognition and cash receipt are not assumed to occur simultaneously.
Energised critical IT divided by secured power, computed per company from the same records the tables display, and shown only when both figures are disclosed. Because the two are usually measured on different bases — critical IT against gross utility — the ratio is a rough indicator of progress, not a like-for-like efficiency measure, and the site labels it as crossing bases wherever it appears. Where either figure is missing the conversion reads Not disclosed; it is never computed against an assumed zero.
An undisclosed figure is stored as null with confidence unknown, and the
data layer refuses to coerce it. In aggregates it is excluded and counted separately, so every total
states how many companies contributed and which did not. There is currently no genuine
numeric zero in the capacity dataset: Keel has no announced lease, which is recorded as not
disclosed with that label rather than as 0 MW, because a zero would imply a measured figure.
Stated in a primary document — an SEC filing, an official investor-relations release, a shareholder letter, or a regulator/utility record — and that document is linked.
Attributed to the company but read second-hand, or carried over from an earlier compile whose original document is not linked. News can put a figure here; it can never make one confirmed.
Derived rather than disclosed. Shown separately from confirmed figures, with the derivation stated, and never included in a confirmed total.
The company has not published this figure. It is stored as null and is never treated as zero.
Only confirmed values count toward a "sourced" percentage, and a value can only be confirmed against a primary document — an SEC filing, an official investor-relations release, a shareholder letter, or a regulator or utility record. News can help discover an event, but can never make a capacity value confirmed.
A record's verification date is when the underlying evidence was genuinely reviewed. It is never set to the current date because a page rendered. Figures never checked against a document show not verified.
Owns the accelerators and sells finished compute. Captures far more revenue per megawatt, and carries the chip capital cost and obsolescence risk to match.
Leases powered, cooled space; the tenant supplies its own accelerators where the lease says so. Earns a few million per megawatt per year rather than tens, with correspondingly lower capital cost.
A powered-shell landlord earning a few million per megawatt per year and a full-stack operator earning many times that are not one better than the other — they sell different products with different capital bases and different risks. A combined revenue-per-megawatt league table would be meaningless, so T2C does not publish one.
Analyst price targets are attributable opinions published by third-party research firms. They are not T2C forecasts, may use different horizons and assumptions, and can change without warning. Consensus statistics use the latest available target from each contributing firm, de-duplicated so an aggregator repeating the same research action cannot count twice. The median is shown as the primary figure because the mean is easily distorted by a single outlier.
Where a research note or provider does not state a target horizon, T2C displays "Horizon not stated". A target date is never manufactured by adding twelve months.
Current availability. The connected market-data plan grants a rating distribution but not price targets, per-firm rating actions or target history. T2C therefore does not display analyst price targets at all, rather than showing an unattributed or placeholder figure. Live provider capability is reported on the data health panel; there is no analyst-target view to open, because there is no analyst-target data to show.
Historical event reactions measure observed share-price movements around past announcements. They do not prove causation and are not predictions of how the share price will react to a future event.
The methodology, implemented and tested, is: Announcements are mapped to the next trading session when they land before the open, the same session when they land during market hours, and the next session when they land after the close or on a non-trading day. Weekends and market holidays are skipped using the trading calendar implied by the price series itself, so a Saturday announcement is never measured against a session that did not trade. Benchmark-adjusted return is the simple difference between the share return and the benchmark return over the same sessions. It is not a beta-adjusted or risk-model alpha. The default benchmark is QQQ, using Adjusted daily closes. Summary statistics are withheld below a sample of 3 comparable events, because a median of two observations is noise presented as a finding.
Current availability. Daily historical candles are not granted by the connected plan, so no event study can currently run. The engine is in place and unit-tested against reference cases; it produces results as soon as a price-history source is connected.
The Scenarios view runs a lognormal random walk on assumptions you choose. It is not options-implied probability, it takes no view on any company, and it is not advice. The median outcome is the 50th percentile — half of modelled outcomes finish above it — and is deliberately not labelled "most likely price", which would imply a mode.
Touch probability exceeds finish-above probability because of the reflection principle for Brownian motion: paths that cross the target and fall back still count as a touch, and each such path can be mapped to a corresponding path that finishes above.
Default volatility: Close-to-close log-return standard deviation on Adjusted daily closes, annualised by sqrt(252) trading days, over a selectable lookback of 30, 60, 90, 252 trading days. When historical closes are unavailable from the data provider, the calculator falls back to a per-ticker typical volatility and says so beside the input. A fallback is never presented as a measured figure.
All deadline arithmetic is date-only and UTC-based, so a deadline always displays as the date chosen regardless of the reader's timezone.
T2C is an information tool. Nothing here is a recommendation to buy or sell anything, and no financial outcome described on this site is guaranteed or certain. Figures are compiled from public filings, may lag, and may contain errors — the corrections log below exists because they sometimes do.
T2C's default is to print Not disclosed and stop. That is correct, and it is also — across the tracked set — unhelpful: nobody can judge scale from a column of blanks. So where a figure can be derived from figures the company itself published, it is derived, and marked as an assumption rather than a disclosure.
Four rules make that safe rather than sloppy:
estimated
confidence, which every aggregate on this site already excludes.Estimates are drawn in amber and labelled. Lime is reserved for evidenced progress; an estimate does not get to borrow the colour that means confirmed. Today 6 figures across the site are derived this way.
| Rule | What is assumed |
|---|---|
| Critical IT from gross utility power | Critical IT load is taken as gross utility power divided by 1.4. Gross power is measured at the utility connection, before conversion, cooling and distribution losses. The divisor is a T2C assumption reflecting typical modern data-centre overheads, not a company figure. |
| Billing from customer acceptance | Capacity formally accepted by a customer is taken as billing. Acceptance is normally the milestone that starts the revenue clock, but the commercial effect is contract-specific and the company has not separately confirmed that billing has commenced. |
| Site capacity from the contract that names it | Where a company discloses megawatts for a contract and names the site serving it, that figure is attributed to the site. It is the contracted capacity, which may differ from the site's ultimate build. |
| Revenue rate from the company's own contracts | Revenue per MW per year is the capacity-weighted mean of this company's own disclosed contracts. Only the company's own agreements are used — never a peer average, which would describe a different business. |
The gross-to-critical-IT divisor of 1.4 is the single assumption that
moves the most numbers. It is a T2C judgement about typical modern data-centre overheads, not a
company disclosure, and it is defined in one place — src/lib/estimate.js — so
changing it moves every derived figure on the site together.
Every other figure on this site is disclosed by a company and cited to a document. The Reality Score is not: it is constructed by T2C from those disclosures. That makes it the one number here that you should check the workings of, so the workings are below.
It is a weighted mean of four factors, each derived from sourced records. No input is hand-assigned, and no factor is scored by opinion.
| Factor | Weight | What it measures |
|---|---|---|
| Promise delivery | 35% | Of the delivery milestones this company both guided and reached, the share that landed on target, early, or inside the guided window. Missing a window is the only outcome scored as a miss. |
| Evidence quality | 30% | The share of this company's published figures and passed gates that rest on a primary document rather than a second-hand report. It measures how checkable the record is, not how good it is. |
| Timeline stability | 20% | The share of guided milestones whose target has not moved since it was first announced. A target that shifts is not a missed deadline, but it is a weaker promise than one that holds. |
| Financing | 15% | Whether capital is committed and available across the company's tracked projects. Not disclosed by most operators at project level, in which case this factor is withheld rather than assumed. |
The weights are a judgement — delivering what you promised matters most, and how checkable the
record is matters next, because an unevidenced record cannot be audited at all. Publishing them
is the honest way to hold a judgement. They are defined in one place,
src/lib/score.js, and changing them changes every score on the site at once.
The score is not a rating, a recommendation, or a prediction. It summarises how well a company has delivered against its own published promises, on the evidence available today.
Values are not edited silently. When a figure changes, the change is recorded and published here.
That figure is the capacity the Finland facility is expected to reach when fully deployed, not capacity energised today. Moved to planned project capacity; current energised MW is now not disclosed. This was a material misclassification.
2.2 GW is the total development pipeline. The company splits it into 648 MW secured and 1,513 MW planned or in development. Secured power corrected to 648 MW.
Both figures come from the Form 10-Q capacity table, which the record now cites alone. The values themselves were correct and are unchanged. The 10-Q also supplied better definitions and three figures previously missing: 341 MW currently energised (gross), 430 MW secured growth capacity, and a total pipeline of 2,161 MW rather than the rounded 2.2 GW. Panther Creek (60 MW) and Scrubgrass (63 MW) are energised but sit under no energy service agreement, and Keel therefore excludes them from its own secured total.
No filing reports zero megawatts — the company simply has no announced lease. Changed to not disclosed with the label "No announced lease", because a zero implies a measured figure.
Replaced with 1,410 MW of contracted critical IT across five named campuses as at 31 May 2026, with all five campus leases itemised.
480 MW is IREN's wider 2026 gross AI Cloud capacity target across multiple sites. Reclassified as a target; company construction MW set to not disclosed pending a comparable official total.
Removed. No primary source in the audit set states an exact fleet count, and a GPU count is not a megawatt measure in any case.
Both retained as dated historical disclosures. Current figures are 839 MW contracted minimum, 102 MW revenue-generating and 336 MW under construction. The Abernathy disposal is recorded so sold capacity does not remain in owned totals.
Removed from both. The claim is not explicitly supported by the cited primary sources.
Every figure now carries a power basis and a value status. Aggregates only combine matching bases and actual/minimum values; targets, pipeline and potential are excluded and reported separately.
Corrected to the Q2 2026 disclosure — 3.7 GW contracted and 1.5 GW active at 30 June — then updated to approximately 4.2 GW contracted per the 11 August earnings call. Both steps are in the ledger.