What to build, then keeping phase 1 on track.

The situation

Three good projects and one connection to the grid

The developer holds a queue position for 600 MW at its substation and has $2.6bn. Three candidates are on the table: a 200 MW hyperscale campus for an anchor tenant ($2.0bn), 400 MW of battery storage earning merchant arbitrage and capacity value ($520M) and a 250 MW solar project with a 15-year power purchase agreement ($310M). The hurdle is 9%.

The campus's value depends on lease rates through the AI cycle. The scenarios include a recession with a pause in AI capital spending, a demand air pocket in which hyperscalers pause and lease rates reset lower, and a period of high energy prices. All three projects compete for the same interconnect.

What Capibud did

One workspace from option choice to monthly review

1. Built each option from a research model. Margins were calibrated to industry-typical returns and left editable, with capex, schedule and execution risk on shared factors including an AI and data-centre demand cycle.
2. Solved the portfolio under the interconnect. The optimiser chose what to build within the budget and the 600 MW constraint for the best expected NPV with a penalty on the bad year, and scored the usual ranked lists on the same scenarios.
3. Carried the sanction into construction. Phase 1 of the campus was split into six packages: shell and core, long-lead electrical, a 230 kV substation, cooling plant, backup generation and the fit-out of three data halls.
4. Ran the monthly review. At the September 2026 data date, actuals were blended with a reference-class prior, the remaining risks simulated on 2,000 scenarios, and contractor dates checked against Capibud's P80.
5. Applied crews and windows. Shared electrical crews were capped at a 550-head site ceiling, with the year-end shutdown and summer heat derating outdoor work.

Results: what to build

Build the campus and the solar; leave $290M unspent

The recommendation has an expected NPV of $394M on $2.31bn of capex, a P10 of $91M and a 5% chance of a negative portfolio NPV. The remaining 90 MW of interconnect is real but useless: the next project in line needs 400 MW. A 10% larger budget would buy nothing that clears the risk-adjusted bar, so $290M stays unspent on purpose.

The campus itself is the risky part: a 7% chance of destroying value on its own, a P10 of $46M and a P50 IRR of 9.9% against the 9% hurdle. Execution slippage costs about $116M across the recommended pair, almost all of it overrun on the campus, whose P80 commercial operation date is April 2030 against a July 2029 plan.

Results: phase 1

Two packages to escalate, and one shift to add

Programme budget authority is $1,185M. The P80 estimate at completion is $1,218M, and only 42% of scenarios finish within authority. CPI is 0.96 and SPI 0.95. Remaining contingency of $89M covers 73% of the P80 need; five pending change orders worth $43M would use almost half of it if granted.

The finish is driven by the data halls

The fit-out's P80 finish is February 2028, 183 days beyond the August 2027 target, at about $520,000 a day of delay. Crews on the package are at 79% of what the work in hand needs (388 on site against 489) and time-based SPI is 0.85. Capibud's first suggested recovery is a second shift on the fit-out, to be priced and previewed against the crew ceiling before anyone commits.

What it missed and limits

A synthetic workspace, used to show the mechanics

• Margins are assumptions. Each option's margin is calibrated to an industry-typical return and is editable. The ranking between storage and solar is sensitive to it.
• Execution history is generated. Budgets, progress, contractor reports and change orders are synthetic; they show how the review works, not how real campuses perform.
• One run. Figures come from one seeded engine run and move when the synthetic history is regenerated.
• Interconnect is simplified. The constraint is one hard MW limit; real queue positions carry their own timing and upgrade-cost risk.

What this means for you

Lumpy constraints beat ranked lists