Appendix A3 · Method for §6 of the main paper · May 2026
The four theses, screened
The main paper places the recommendation among four global theses and moves on. This note carries the screening behind that placement: three diagnostic questions, the scorecard at central priors, and the rule for reading it. Nothing here changes the decision.
Three questions, four theses
The May 2026 screen asked three questions of each of the four global theses: how durable is enterprise renewal into the next contract cycle (central prior 75 per cent); what share of the labs’ compute commitments is cleared by external customer revenue rather than recycled vendor flows (central prior 25 per cent); and what is a GPU worth at year four (central prior USD 9,000 per H100). The scorecard below is the screen at exactly those priors.
| Thesis | Q1 · Renewal | Q2 · External clearance | Q3 · GPU residual | Read at these priors |
|---|---|---|---|---|
| Hyperscaler captive AWS · Azure · GCP · Meta | Adverse | Mixed | Adverse | Refuse |
| Neocloud merchant CoreWeave · Lambda · Nebius · Crusoe | Mixed | Adverse | Adverse | Refuse |
| REIT / infrastructure fund Digital Realty · Equinix · Brookfield · KKR | Favourable | Mixed | Favourable | Capital, with structural protection |
| Sovereign / strategic Stargate · G42 · HUMAIN · EU sovereign | Favourable | Favourable | Mixed | Capital, with structural protection |
Reading rule, unchanged from the screen: three favourable answers deserve capital; two, capital with structural protection; one, caution and probably refuse; zero, refuse. Cell thresholds and central priors are the May 2026 screening dashboard’s own; the dashboard’s slider form is retired from this pack and the scorecard is shown static at its central defaults.
Three of the screen’s cell reads carry the argument. The neocloud exposure on Q2 is direct: at the May 2026 reading, roughly 60 per cent of CoreWeave’s revenue was Microsoft, the bulk of it OpenAI workload. The REIT row is favourable on Q1 and Q3 for one structural reason: the tenant absorbs demand risk and hardware obsolescence while the landlord holds power, shell and grid connection through the cycle. And the hyperscaler exposure runs through depreciation: extended GPU depreciation schedules are justified only while enterprise renewal holds.
External revenue against the recycled loop
Two kinds of money meet AI-infrastructure capex. External customer revenue, from consumers, developers, enterprises and governments, is the money that ultimately services the capital. Alongside it runs a recycled loop: equity and compute commitments passing from vendor to lab and back again as compute purchases. Microsoft to OpenAI to Azure; Nvidia to OpenAI to Oracle. The loop is real money, but it is not external revenue.
The screening rule carried from the dashboard: any underwriting that treats lab-side compute backlog as if it were external customer revenue should be discounted to the share that external revenue plausibly supports. Q2 above prices that share; the central prior is 25 per cent.
The boutique case does not lean on this question. It is sized as fill from seven named segments (main paper, S0), not as a share of a market forecast, which is why the market-level question appears only here, in method.
The Australian application
REIT and infrastructure fund is the structural fit for Gowings: power, shell and connection rights as the durable asset, with GPU obsolescence sitting with the tenant. Sovereign-strategic applies to the government-mid segment specifically. Hyperscaler-captive is the comparator the service premium is measured against. Neocloud does not apply: it puts three-to-six-year compute on the owner’s balance sheet, and the main paper excludes it entirely.
The screen pre-dates the memo’s two-vehicle structure and its placement holds under it. Both vehicles keep the durable asset with Gowings and the compute with the customers; Vehicle B adds the operating margin on top of the same placement without adopting the neocloud balance sheet.
Main paper · The four theses, compact: §6.A · The two vehicles: §2