About this paper
Who it is for.John Gowing. He asked the author a direct question: why are the data centres being built? This paper is the answer, carried through to the decision the answer leads to. It is written to board-paper discipline so it can go to the Board of Gowings Bros later without rework, but it is addressed to John, not the Board; the Board comes later, if he decides it should. It assumes no prior knowledge of datacentres. It does assume the reader can price a property deal.
Why it exists.The question has two answers. The market answer: the AI buildout is a race to house computing equipment, and most of the capital lands on the equipment itself, a wasting asset on a three-to-six-year life. The investment answer, the one this paper carries: underneath that race sits the durable asset, the building, the power connection and the land, and it has the shape of the firm's own craft; develop, own for decades, exit to an institution. This paper isolates that asset in Sydney, at boutique scale, and prices it.
What it asks of the reader.One decision, staged. The paper recommends proceeding, in a specific vehicle. But under its own rule no capital commits until six diligence gates close, and today none are closed. The decision it prepares is to authorise the diligence that closes or fails the gates, not to commit AUD 1,260 m today. Sections 0 and 1 carry the recommendation and the numbers; 2 to 5 carry the case; everything after that is context and changes nothing.
The question
The decision this paper prepares is whether to commit AUD 1,260 m over 12 to 18 months to develop and own a 22 → 28 MW premium-tier datacentre in Sydney metro, and in which vehicle: landlord (triple-net) or operator (JV colocation).
The asset.An industrial building with a heavy grid connection and a cooling plant, leased to customers who install and replace their own computers. Gowings owns the durable shell, power connection, cooling plant and land, never the obsolescing computing equipment.
Why the opportunity exists.APRA's CPS 230 operational-resilience standard, in force since July 2025, the Privacy Act reforms staged through 2025 and 2026, and cyber-insurance underwriters that price Australian residency differently from offshore have created a customer class that pays for Australian-located, Australian-controlled space over the US-parented hyperscaler regions. Seven segments: APRA-regulated institutions, legal, medical, mining IP, professional services, family offices and mid-sensitivity government. The case is sized as fill, not market share: roughly 28 MW taken at a service premium of 18 per cent, tested up front by two anchor pre-commitments of about 60 per cent of capacity.
Why this size.The envelope is set by the customer list, not by ambition. The research behind Section 3 and Appendix A2 reads the seven segments above as the full population of Australian customers who will actually pay a premium to hold their data in a sovereign environment, and 22 → 28 MW is what that population fills, with the two anchor pre-commitments as the up-front proof. That is the selling proposition working as intended: sovereignty is the product, the premium is real only inside this customer class, and a facility sized past it dilutes into commodity space it cannot win on price. It is also why the paper does not chase customers of the Firmus type: gigawatt-scale AI-compute businesses that buy on cheapest power and price per GPU-hour, pay no sovereignty premium, and bring the equipment-obsolescence risk this paper refuses. A premium sovereign landlord and a merchant AI barn are different businesses; this paper prices only the first.
Why Gowings, why now.Gowings has developed and owned durable Sydney property for a century and a half; this is the same shape: develop, own for thirty years, exit to an institution. CPS 230 is now in force, sub-30 MW grid connections remain available in the established Sydney corridors without the gigawatt-scale queue, and two-to-three-year transformer lead times make a head start hard to copy.
The recommendation follows.A boutique premium-tier datacentre for Sydney metro · landlord or operator · May 2026
Proceed. Develop a boutique premium-tier datacentre in the Sydney metropolitan area at 22 megawatts of energised capacity, expandable to 28 megawatts. Hold the asset in Vehicle B (Operator / JV colocation) rather than Vehicle A (Landlord / triple-net). Same building, same power connection, same land. Different commercial structure on the floor.
Recommendation quoted verbatim from Section 1 of the investment memo (HSv3). This recommendation is the adviser's. The decision is yours.
Today's ask: authorise diligence under the Section 4 rule. Capital commits only when the gates close; today zero of six are closed.
| Measure | A · Landlord | B · Operator (JV colo) |
|---|---|---|
| Energised capacity | 22 → 28 MW | 22 → 28 MW |
| Total deployed capital | AUD 1,260 m | AUD 1,260 m |
| Stabilised revenue | AUD 149 m/yr | AUD 210 m/yr |
| Stabilised NOI / EBITDA to Gowings | AUD 116 m/yr | AUD 111 m/yr |
| Operating margin | ~78% | ~53% |
| Unlevered stabilised yield / ROIC | ~9.2% | ~8.8% |
| Levered stabilised cash-on-cash | ~12.5% | ~10.6% |
| Full-hold levered IRR, 30 yr | ~8.8% | ~8.1% |
| Simple payback | ~11 yrs | ~11 yrs |
Source: underwriting model, Summary tab. Vehicle A returns calibrated to AI_Datacentres_v3.docx (May 2026). Vehicle B central inputs: A$800/kW/month colo price, 78% stabilised utilisation, 16x exit EV/EBITDA, 80/20 Gowings/operator JV split; all flagged as analyst assumptions on the model's Sources & Notes tab.
Vehicle A wins the central case on the operating lines. Vehicle B is recommended for the margin it owns and the upside it keeps; Section 2 carries the trade-off.
The three return measures are defined where each vehicle is priced (Section 2) and in the glossary (6.F).
Go / no-go rule · six diligence gates (Section 4)
Today: 0 of 6 gates closed. No capital commits until they close.
The rule is pre-committed before evidence arrives. That is the discipline that keeps partial answers from being rationalised into a yes.
The two vehicles, in property language
What it means for the decision: the choice is not whether to build. It is which vehicle holds the building, because the vehicle decides who earns the service margin and who carries operating risk.
The physical asset is identical in both vehicles. Read it as an industrial property development: a purpose-built facility on industrial-zoned land, a high-capacity electricity connection, and cooling plant. Two assets sit on the site and must not be confused. The building, the power connection, the cooling plant and the land are the durable asset; the land and grid connection rights are effectively perpetual, and Gowings owns them under either vehicle. The customers' computing equipment (the 'compute') has a three-to-six-year economic life and is replaced repeatedly. Under either vehicle as recommended, the customers own and replace it. The obsolescence risk sits with them, not with Gowings. The neocloud thesis (CoreWeave, Lambda, GPU-as-a-service), which puts the compute on the owner's balance sheet, is excluded entirely.
Context · The asset, at four scales, and what ages: Appendix A1
| Capex line | Central | Share | Range |
|---|---|---|---|
| Power infrastructure | 330 | 26.2% | 280–380 |
| Shell, civil & envelope | 175 | 13.9% | 145–215 |
| Cooling plant | 165 | 13.1% | 140–200 |
| Soft costs & capitalised interest | 155 | 12.3% | 130–200 |
| Contingency | 105 | 8.3% | 90–140 |
| Grid connection contribution | 90 | 7.1% | 70–120 |
| Land & site acquisition | 80 | 6.3% | 60–110 |
| Mechanical & electrical balance | 75 | 6.0% | 60–95 |
| Network & fibre | 50 | 4.0% | 40–70 |
| Security & compliance fitout | 35 | 2.8% | 25–50 |
| Total deployed capital | 1,260 | 100.0% | 1,040–1,580 |
Source: model, Development Budget tab, cells C6:C16 (ranges columns E and F; shares Assumptions rows 44 to 53). Row labels are the model's own. The AUD 330 m power-infrastructure line is the transformer and power-train capex carried by the OEM row in Section 5. Stored six-decimal shares sum the central column to 1,260.001; the model's own tie-out check rounds the difference to zero and reads TRUE. Displayed figures tie to 1,260 exactly. Variance flagged, not adjusted (G31 ledger).
| Years | Stage | Source |
|---|---|---|
| pre-Y0 | Diligence on the six gates; closes or fails them; no capital commits | Paper §1 rule card and §4 |
| Y0 | Gates closed; final investment decision; the AUD 1,260 m outlay is booked | Model, Cash Flow & Returns C11 |
| Y0 on | Build; the 2–3 yr transformer lead is the binding constraint | Paper §0 and §5; Development Budget power note |
| Y1 | Energisation at 22 MW on the model clock; first revenue at 7.5 MW utilised | Assumptions C6; Cash Flow & Returns D6, D7 |
| Y1–Y4 | Ramp, 4 years central (7.5 MW, 12 MW, 16 MW, 22 MW) | Assumptions C23 and rows 36 to 39 |
| Y5 | Stabilised: 28 MW basis; 78% utilisation on Vehicle B | Assumptions B40, C88 |
| Y6 · Y12 · Y18 · Y28 | Refresh events, AUD 20 / 35 / 65 / 60 m | Assumptions rows 58 to 61; §6.D |
| Y10–Y15 | Exit window canvassed at gate Q6 | Paper §4 Q6 (year ten or fifteen) |
| Y30 | Full-hold horizon; terminal value in the 30 year IRR | Assumptions C32; Cash Flow & Returns |
Clock: model years, Y0 the outlay year. The model books all development capital at Y0 and first revenue in Y1; it carries no explicit construction period, so the build row is the paper's stated lead-time constraint laid on the model's axis. A physical build at the 2–3 yr transformer lead sits between FID and energisation and would push every later milestone out by the build time; the model does not carry that offset. Variance flagged, not corrected (G31 ledger).
Vehicle A · Landlord (triple-net) · baseline
Gowings leases the entire facility to a specialist colocation operator under a long-tenor triple-net lease. Triple-net (NNN) means the tenant pays rates, insurance and maintenance on top of rent. The tenant operates the floor and captures the colocation service margin from end customers; Gowings collects a near-bond-like rent. This is the firm's traditional craft: site selection, capital cost discipline, lease structuring, exit timing.
| Line | Amount | Source |
|---|---|---|
| Contracted rent (28 MW × AUD 5.31 m/MW/yr) | 149 | Model: Stabilised P&L |
| Operating expenses (22% of revenue) | (33) | Model: Stabilised P&L |
| Net operating income (NOI) | 116 | Model: Stabilised P&L |
The AUD 5.31 m/MW/yr anchor rent is the AUD 4.5 m hyperscaler-Sydney baseline plus the 18% service-tier premium. The premium is Vehicle A's load-bearing assumption (Section 3).
| Measure | Central | Basis |
|---|---|---|
| Unlevered stabilised yield | ~9.2% | NOI / deployed capital |
| Levered stabilised cash-on-cash | ~12.5% | (NOI less interest) / equity · 55% LTV at 6.5% · debt AUD 693 m, equity AUD 567 m, interest ~AUD 45 m/yr |
| Full-hold levered IRR, 30 yr | ~8.8% | DCF incl. ramp, refresh, exit at a 6.5% cap rate (flagged analyst assumption) |
| Simple payback | ~11 yrs | Deployed capital / NOI |
A cap rate is stabilised NOI divided by value: 6.5% values the property at about 15.4 times NOI. The yield and cash-on-cash do not depend on it; only the 30-year IRR does. The IRR is shown for completeness and is not the lead measure: it sits below the cash-on-cash because it carries the ramp years, the refresh capex and the time-discounting of the outlay.
Vehicle B · Operator (JV colocation) · recommended
Gowings owns the same building and additionally owns the operating company that sells the colocation service. Customers contract with the Gowings operating entity for power, space, cooling and connectivity. An established operator runs the floor under a joint-venture management contract: a 3% management fee on revenue as cost-recovery, plus a 20% share of EBITDA. Gowings retains 80% of EBITDA. Gowings does not have the operator skill set in-house; the JV is the structure that buys it while keeping the asset and the majority of the margin.
| Line | Amount | Note |
|---|---|---|
| Colocation revenue (28 MW × 78% utilisation × AUD 9.60 m/MW/yr) | 210 | A$800/kW/month central |
| Power purchase (37.8 MW × 8,760 h × A$130/MWh) | (43) | PUE 1.35 |
| Cooling, staffing, security, service, maintenance, insurance | (22) | Analyst assumptions |
| Total operating expenses, pre-JV fee | (65) | ~31% of revenue |
| Operator JV management fee (3% of revenue) | (6) | JV term |
| EBITDA before JV split | 138 | |
| Operator EBITDA share (20%) | (28) | JV term |
| EBITDA to Gowings | 111 | ~53% margin |
PUE (power usage effectiveness) is total facility power divided by IT-equipment power; 1.35 sits at the NABERS 5-star regulatory floor of 1.4 in force from July 2025. The ~53% Gowings margin sits between NEXTDC's ~51% (FY25, retail-colo profile) and AirTrunk's ~65% (wholesale profile). Source: model, Operator (JV colo) tab.
| Measure | Central | Basis |
|---|---|---|
| Unlevered stabilised yield (ROIC) | ~8.8% | EBITDA to Gowings / deployed capital |
| Levered stabilised cash-on-cash | ~10.6% | (EBITDA less interest) / equity · 50% LTV at 7.0% · debt AUD 630 m, equity AUD 630 m |
| Full-hold levered IRR, 30 yr | ~8.1% | DCF incl. ramp, refresh, EV/EBITDA terminal value |
| Simple payback | ~11 yrs | Deployed capital / EBITDA to Gowings |
| Exit multiple, central | 16x | EV/EBITDA · boutique discount to NEXTDC trading (60.9x) and the AirTrunk precedent (~22x) |
EV/EBITDA is enterprise value divided by EBITDA, the operating-business counterpart of a property cap rate. Note the basis shift: Vehicle A exits on a cap rate, Vehicle B on an EV/EBITDA multiple; the two are not directly comparable (model flag B1).
Why Vehicle B, not Vehicle A
Vehicle B captures the service-tier margin that Vehicle A cedes to the tenant. The customer pays a colocation rate that bundles rent, power pass-through, cooling, remote-hands service and connectivity: typically two to three times the underlying NNN rent. In Vehicle A that bundled rate is the tenant's revenue line, not Gowings'.
The trade-off is honest. Vehicle A delivers the higher bond-like operating return at the central case (12.5% cash-on-cash against 10.6%). Vehicle B carries operating-business risk in exchange for the right to the service margin and a structurally higher upside if utilisation and pricing are sustained: at the upside corner (90% utilisation, A$1,000/kW/month) Vehicle B's levered cash-on-cash is about 22%, well above Vehicle A's ceiling.
The decision is also strategic, not only financial. Vehicle B moves Gowings from property investor to infrastructure operator. That change in commercial identity is the part of this recommendation that warrants the most scrutiny, and it is why the gates in Section 4 are dominated by the operator-partner question.
Context · The pooling arbitrage the operator sells: Appendix A2
What must be true
What it means for the decision: each vehicle rests on one number. If that number cannot be evidenced under diligence, the recommendation changes. This section names the number, the level that retires the case, and the level that kills it.
| Vehicle | Assumption | Central | Case retires | Case fails | Gate that closes it |
|---|---|---|---|---|---|
| A · Landlord | Rent premium over hyperscaler-Sydney | +18% | below +10% | at +5% | Premium evidenced in signed contracts, not pitch decks |
| B · Operator | Stabilised utilisation at central pricing | 78% | below ~70% | below ~60% | Q2 operator · Q3 anchor pre-let · Q4 JV terms |
Utilisation is the share of sellable capacity actually filled by customers. Thresholds per memo Section 6 and the model's operator sensitivity grid: at A$800/kW/month the levered cash-on-cash crosses below the 8% clear threshold at about 70% utilisation and below the 4% fail threshold at about 58%. The model's Summary note rounds the same grid to ~65% and ~55%; the stricter reading is used here.
Context · Why pooled utilisation is the operator’s craft: Appendix A2
Vehicle B's secondary load-bearing assumption is the power margin. Power purchase is roughly two-thirds of the cash opex stack. A sustained 30% move in wholesale power above the central A$130/MWh costs about AUD 13 m/yr of EBITDA unless passed through to customers via the colocation contract. Gate Q5 exists for exactly this clause.
Corner cases · Vehicle A (v3 Ch.4 set)
| Scenario | Unlev. yield | Levered c-o-c | Verdict |
|---|---|---|---|
| Central · 45 m/MW · +18% · 60% pre-let · 4 yr ramp | 9.2% | 12.5% | Clears comfortably |
| Corner 1 · capex AUD 60 m/MW | 6.9% | 7.4% | Clears only with wind behind; price discipline required |
| Corner 2 · premium +5% | 8.2% | 10.3% | Fails on premium economics: no real premium over the comparator |
| Corner 3 · ramp 6 yrs, steady state unchanged | 9.2% | 12.5% | Survives with patient capital; the cost is IRR drag in the ramp |
| Corner 4 · 90% pre-let · +25% | 9.8% | 13.7% | Above envelope; interrogate why an operator would concede it |
Source: model, Sensitivity tab (explicit recomputed grids, not Excel what-if tables). Variance flag: v3 cites ~6.5% yield at Corner 2, but its own stated revenue at +5% (AUD 132 m) implies 132 × (1 − 22%) / 1,260 = 8.2%. The recomputed figure is shown; the case still fails at +5%, on premium economics rather than on a yield hurdle. All four v3 source variances are listed in the appendix.
Corner cases · Vehicle B (operator grid)
| Scenario | EBITDA yield | Levered c-o-c | Verdict |
|---|---|---|---|
| Central · 78% utilisation · A$800/kW/mo | 8.8% | 10.6% | Clears |
| Stress · 60% utilisation · A$500/kW/mo | 2.1% | −2.8% | Fails |
| Price compression · 78% utilisation · A$500/kW/mo | 3.9% | 0.9% | Fails |
| Upside · 90% utilisation · A$1,000/kW/mo | 14.5% | 22.0% | Clears |
Source: model, Sensitivity (Operator) tab. Verdict thresholds: clear at 8% levered cash-on-cash or better, marginal above 4%, fail below. Both fail rows combine price and utilisation stress; the anchor pre-let (Q3) and the take-or-pay structure exist to keep the facility out of that corner.
Method · The envelope tested, and the v3 recomputations: Appendix A4
The six gates, and where they stand
What it means for the decision: no capital commits until the gates close. Today zero of six are closed. Under the rule on the decision screen, that means diligence proceeds and financial close does not. Each gate is a question with a named form of evidence that closes it; partial answers do not count.
Q2 · The operator partner
Open · screened, unresolvedWhich established Australian colocation operator runs the floor under the JV management contract, and what is their track record on customer fill, churn and service delivery? This is the single most important gate in Vehicle B. Without it, the JV is a fee paid for capability that does not exist.
The partner screen was run in May 2026 and it did not resolve the gate. The screen filters on partner logic before operator quality: the binding question is not who runs colocation floors well, but for whom a minority-EBITDA management contract on a Gowings-owned asset is mutually rational. A self-funding incumbent has no reason to run Gowings' asset for 20% of EBITDA plus a fee when it can deploy its own capital at 100% equity. On that filter, no single operator clears all four must-haves: the best-credentialed operators are the least willing, and the willing operators are the least credentialed.
Current position: two consortium candidates shortlisted, each pairing an established floor operator with a sovereign sensitive-segment partner. First, Digital Realty Australia with Vault Cloud. Second, DCI Data Centers with AUCloud (AUCyber). The reusable selection rubric in the shortlist paper is the basis for an RFP if both consortia stall.
Evidence that closes it: customer references on three to five comparable facilities; utilisation curves on at least one comparable greenfield over its first five years; financial review of covenant strength; named senior operating personnel committed to this facility.
Q1 · The site and grid
Open · not startedWhich Sydney-metro precinct, with what grid energisation date and what import-export limits at the connecting bus?
Evidence that closes it: written confirmation from the network operator (Endeavour Energy or Ausgrid) on the bus, energisation date and capacity envelope; independent power-systems review; sight of the connection agreement. A site without a specific bus and a credible date is a development project, not an investment.
Q3 · Anchor pre-let
Open · not startedWhich two anchor customers pre-commit to a meaningful share of capacity at signing, from which segments, on what take-or-pay, and at what colocation rate? Take-or-pay means the customer pays for contracted capacity whether or not it is used; it is the cheapest hedge against the lease-up period.
Evidence that closes it: signed heads of agreement on both anchors before financial close, jointly covering at least 40% of sellable capacity at central pricing or better; take-or-pay with limited step-down; cross-default between the two anchor contracts; a contingent third anchor under exclusivity.
Q4 · JV terms
Open · not startedWhat is the EBITDA split, the management fee structure, the governance of capex decisions, and the operator-exit mechanism over the 30-year hold?
Evidence that closes it: a term sheet at or tighter than 80/20 (Gowings/operator); management fee at no more than 3% of revenue; capex decisions over a threshold reserved to Gowings; a buyout right that lets Gowings replace the operator without disrupting customer contracts; part of the operator's EBITDA share conditional on utilisation and customer-satisfaction targets.
Q5 · The power contract
Open · not startedWhat is the structure of the wholesale power contract, and how is wholesale-price volatility shared between the operator and the customers?
Evidence that closes it: a long-tenor power purchase agreement (10 years or more) with a credible counterparty, its floor and ceiling within the central A$130/MWh assumption; customer contracts that pass wholesale moves above an agreed level (typically A$150/MWh) through to customers via a power-band escalator. Without this clause, a sustained electricity-market shock destroys Vehicle B's EBITDA.
Q6 · Exit
Open · not startedWhich natural buyer absorbs the asset at year ten or fifteen, on what basis (EV/EBITDA multiple or property cap rate), and at what level?
Evidence that closes it: indicative valuation ranges from three or four credible buyers across both buyer sets: digital-infrastructure funds for the operating company, and listed property trusts and super funds for the underlying property if the JV is unwound at exit. Recent comparable transactions. The cap-rate-versus-EV/EBITDA basis shift (model flag B1) is the gating subtlety.
| Gate | Question | Status |
|---|---|---|
| Q1 | Site and grid | Open |
| Q2 | Operator partner | Open · screened, unresolved |
| Q3 | Anchor pre-let | Open |
| Q4 | JV terms | Open |
| Q5 | Power contract | Open |
| Q6 | Exit | Open |
| Closed: 0 of 6 · rule verdict today | Diligence proceeds; no financial close |
Counterparties and precedents, named
What it means for the decision: a case of this size is only as strong as its named counterparties. Every material name is below. If a name cannot be produced under diligence, the relevant gate stays open.
Operating comparables
| Comparable | EBITDA margin | EV/EBITDA | PUE | Read with care |
|---|---|---|---|---|
| NEXTDC (ASX: NXT) | ~51% (FY25) | 60.9x LTM | 1.30 | Growth multiple on a 133.9 MW forward order book; not a stabilised level and not an underwriting anchor |
| AirTrunk (private; Blackstone/CPP) | ~65% | ~22x precedent | 1.35 | Sep 2024 acquisition at ~A$24 bn EV; strategic-buyer premium; wholesale margin a boutique cannot match |
| Vehicle B central case | ~53% | 16x assumed exit | 1.35 | Boutique discount to both; sensitivity 10–22x |
Sources: NEXTDC FY25 results (28 Aug 2025; revenue A$427.2 m, EBITDA A$216.7 m); Finbox NXT EV/EBITDA LTM (Dec 2025); Blackstone AirTrunk announcement (4 Sep 2024; 20–23.5x range per industry estimates); AirTrunk Sustainability Report 2024. A 30-year hold should not underwrite to peak trading or peak precedent; 16x sits between and below both.
Named counterparties by role
| Role | Names | Where they bind |
|---|---|---|
| Grid connection | Endeavour Energy · Ausgrid | Gate Q1: written bus, date and envelope confirmation |
| Transformer and power-train OEMs | Hitachi Energy · GE Vernova · Siemens Energy · ABB | Capex line AUD 330 m; 130–350 MVA units on 2–3 year lead times |
| JV consortium candidates | Digital Realty Australia + Vault Cloud · DCI Data Centers + AUCloud (AUCyber) | Gate Q2: shortlisted May 2026, unresolved |
| Exit buyer sets | Super funds · listed property trusts (REITs) · digital-infrastructure funds | Gate Q6: property buyers if the JV unwinds; infrastructure buyers for the operating company |
| Calibration incumbents | Macquarie Data Centres · Canberra Data Centres · Equinix | Anchor the AUD 4.5 m/MW/yr baseline rent and the A$800/kW/month retail-colo price |
Sources: memo Sections 8–9 and Appendix B; operating-partner shortlist paper (May 2026); v3 supply-chain chapter for the OEM set.
Background appendix
Nothing in this appendix changes the decision. It is context, demoted on purpose so it stays out of the decision path.
A. Where this sits among the four global theses
Four theses dominate the global AI-datacentre conversation: the hyperscaler captive build; the neocloud merchant build (excluded: it puts three-to-six-year compute on the owner's balance sheet); the property and infrastructure-fund hold (Vehicle A); and the operator vehicle (Vehicle B). The recommendation is squarely the fourth. At frontier scale an Australian site cannot compete with US Sun Belt cost; the boutique at 22 to 28 MW is defined against that limit, and against the Australian-located hyperscaler regions a well-run local boutique can charge a service premium.
Method · The screening behind this placement: Appendix A3
B. The asset, from chip to hall
Scale, for orientation: this facility's whole 22 to 28 MW is boutique against the hyperscale Sydney campuses (AirTrunk's Eastern Creek site runs 200 MW, its Mascot site 300 MW and expanding), and the frontier merchant build this paper refuses is measured in gigawatts.
| Scale | What it is | Power | Cost (USD) | Life |
|---|---|---|---|---|
| Chip | One GPU accelerator; matrix maths in parallel | ~700 W | 25–40 k | 3–4 yrs economic on training work |
| Server | Eight GPUs in one chassis; liquid-cooled at the chip | ~10 kW | 250–400 k | Turns over with the chips |
| Rack | Eight servers; 64 GPUs per cabinet; cooling binds the floor | ~80 kW | 2.5–3.5 m | Turns over with the chips |
| Hall | Rows of racks; grid feed, cooling water, fibre | 50–200 MW | 0.5–2 bn | Shell, substation and grid right do not age |
The compute layer ages on a three-to-four-year economic life; the shell, the substation and the grid right do not. Everything else exists to keep the chips cool, fed and connected. That split is the whole reason the two-vehicle structure works: the durable perimeter is Gowings' asset, the ageing interior is the customers' problem. Because any one customer's demand is peaky, pooling many customers with different busy periods keeps the building usefully full. That pooling is the operator's craft, and it is the structural reason the facility earns a service-tier rent.
Context · The stack in full: Appendix A1 · The demand it pools: Appendix A2
C. Demand segments the operator sells into (Vehicle B)
| Segment | Illustrative counterparties | Fitting workloads |
|---|---|---|
| Mining IP | BHP, Rio Tinto, Fortescue, Newmont Australia | Geological models, ore-body confidentiality |
| Legal sensitive | Allens, MinterEllison, King & Wood Mallesons, Herbert Smith Freehills | Privileged matter management, M&A data rooms |
| Medical premium | NSW Health Pathology, Garvan Institute, Royal Prince Alfred | Genomic data requiring Australian residency |
| APRA-regulated | Commonwealth Bank, Westpac, NAB, ANZ, QBE, IAG, AustralianSuper | CPS 230 workloads where boutique resilience exceeds defaults |
| Professional services | Deloitte, EY, KPMG, PwC, McKinsey, BCG, Bain, Accenture | Client data under sovereignty clauses |
| Family conglomerate | Sydney real-estate dynasties, family offices | High-net-worth, intergenerational and trust data |
| Government mid | NSW Customer Service, Communities & Justice, Services Australia | PROTECTED-equivalent workloads short of defence enclaves |
Counterparties are illustrative, per memo Appendix B. No single segment justifies the facility; together they support a build sized for them. CPS 230 is an APRA prudential standard on operational resilience in force from July 2025.
D. Refresh profile, 30-year hold
| Event year | AUD m | Cumulative | Scope |
|---|---|---|---|
| 6 | 20 | 20 | Early UPS battery cycle |
| 12 | 35 | 55 | UPS and minor M&E refresh |
| 18 | 65 | 120 | Mid-life transformer and cooling refresh; single largest event |
| 28 | 60 | 180 | Envelope and cooling refresh ahead of terminal sale |
| 180 | ~14% of deployed capital across 30 years |
Land, site and grid connection rights are held at original cost and do not depreciate to zero. The per-event split is a model allocation (flagged); v3 gives the AUD 180 m total. Owner refresh applies to plant, not to customers' compute, which the customers replace on their own cycle.
E. The four source variances, flagged not corrected
| Flag | Source states | Recomputation |
|---|---|---|
| 1 | Unlevered yield 8.7% (v3 Fig 4.1) | NOI 116 / capital 1,260 = 9.2%. The 8.7% requires ~94.5% revenue realisation (a spot-pricing discount the prose does not reconcile) |
| 2 | 'Levered IRR' 12.7% | On v3's own method this is (NOI less interest) / equity: a stabilised cash-on-cash, not a time-weighted IRR. The model's true 30-year levered IRR is ~8.8% |
| 3 | No exit cap rate specified | The 30-year IRR cannot be computed without one. The model uses a flagged 6.5% analyst assumption; the yield and cash-on-cash do not depend on it |
| 4 | Corner 2 (+5% premium) yield ~6.5% | v3's own revenue at +5% (AUD 132 m) implies 132 × (1 − 22%) / 1,260 = 8.2%. The case still fails at +5%, on premium economics |
None of the four changes the recommendation. Full arithmetic on the model's Sources & Notes tab and in the reconciliation ledger accompanying this site.
Method · The recomputations in context: Appendix A4
F. Compact glossary
| Triple-net (NNN) | Tenant pays rates, insurance and maintenance on top of rent. Vehicle A's lease structure |
| NOI | Rent less operating costs, before financing and tax. Vehicle A's operating profit measure |
| EBITDA | Earnings before interest, tax, depreciation, amortisation. Vehicle B's operating profit measure |
| Cap rate | NOI divided by value; the property market's price for stabilised income. Inverse of a P/E-style multiple |
| EV/EBITDA | Enterprise value over EBITDA; the operating-business counterpart of a cap rate |
| Cash-on-cash | Cash to equity after interest, divided by equity. Single stabilised year |
| IRR | Time-weighted return across the whole hold, including ramp, refresh and exit. Shown, never led |
| PUE | Total facility power over IT power. 1.35 assumed; NABERS 5-star floor is 1.4 from July 2025 |
| Utilisation | Sellable capacity actually filled. Vehicle B's load-bearing assumption |
| Take-or-pay | Pay for contracted capacity whether or not used. The cheapest hedge against slow lease-up |
| Churn | Customer departure rate per year; 8–12% is the Australian retail-colo norm |
| Ramp | The fill-up period from energisation to stabilised utilisation |
G. Sources
Primary: AI_Datacentres_v3.docx (May 2026), seven chapters and four annexes; every Vehicle A figure traces to it. Vehicle B inputs are analyst assumptions flagged B1–B8 on the model's Sources & Notes tab. Underlying v3 sources: Uptime Institute cost and lifecycle benchmarks; JLL Datacenter Outlook 2025–26; AEMO NEM dispatch data; NEXTDC FY25 results and capex disclosure; AirTrunk and Macquarie Data Centres disclosures; APRA CPS 230 materials; OAIC Privacy Act reform papers; Insurance Council of Australia cyber-underwriting bulletins. Vehicle B additional: NEXTDC FY25 Results (28 Aug 2025); Finbox NXT EV/EBITDA LTM (Dec 2025); Blackstone AirTrunk announcement (4 Sep 2024); AirTrunk Sustainability Report 2024. All amounts in 2025 Australian dollars; AUD/USD 0.65.
Appendix · Context and method
Four context notes sit behind this paper. None changes the decision; each explains one layer of it. They are for the reader who wants the method, not the verdict.
| A1 | The asset, at four scales | What ages and what does not, from chip to hall | supports §2 |
| A2 | What the customer buys | GPU-hours, bursty demand, and the pooling that fills a floor | supports §3 |
| A3 | The four theses, screened | The screening method behind the placement in §6 | supports §6 |
| A4 | How the numbers were tested | The envelope tested, and what the model recomputed against the source pack | supports §3 |