Two-hour or four-hour storage? The answer moved in eighteen months

We sized a fleet against 2023 spreads and the market re-priced underneath us. What we would ask before sizing the next one.

In late 2022 we committed to 940 MWh of storage across seven sites. The duration decision — two hours against four — was made on eighteen months of observed intraday spreads and a forward curve we had reasonable confidence in. We built mostly two-hour. By mid-2024 the economics of that decision had inverted, and not because we had modelled the spreads wrong.

What we measured

Comparing the assumption set at financial close with three years of realised data across the seven sites:

• Average intraday spread came in within 8% of our forecast. The price signal we sized against was broadly right.
• Cycles per year came in 31% below forecast. Not because there was less opportunity, but because more of the opportunity was concentrated in fewer, deeper events.
• Revenue from frequency services fell by roughly two thirds as the market saturated — faster than our low case.
• The share of total revenue from events longer than two hours rose from 11% to 34%.

That last line is the decision. A two-hour asset captures the first two hours of a deep event and then watches the rest. When deep events were rare, that cost little. When they became the main source of margin, it cost a great deal.

Why the forecast was right and the decision was wrong

We forecast the price. We did not forecast the shape of the opportunity, and we had no line in the model for it. Our duration sensitivity ran two-hour against four-hour at the same spread and the same cycle count, which by construction favours the cheaper asset. It could not have produced the other answer.

Our sensitivity analysis varied the price and held the shape fixed. It was a test of how wrong the price could be, presented as a test of the duration decision.

This is the failure worth naming. It was not an information problem — the drivers of longer events, principally the pace of renewable build-out on the same networks, were in our own development pipeline. It was a modelling problem. The model had no variable that could carry the effect.

The four questions we ask now

• What is the duration of the top 5% of events, and what is that trending at? Not the average event. The ones that carry the margin.
• What is being built behind this node in the next five years? Renewable build-out on the same network is the main driver of event depth, and it is knowable from the queue rather than from a price forecast.
• What does the augmentation path cost, and is the site laid out for it? Three of our seven sites can be extended to four hours within the existing footprint and grid connection. Four cannot. That distinction was not a design criterion in 2022 and it is now the first one.
• What breaks the revenue stack we are relying on? Frequency services were a third of our case and saturated in under two years. Anything that can be saturated by other people's capital should be modelled as if it will be.

What we did about the existing fleet

Two sites are being augmented to four hours in 2026 — both of them ones where the layout allowed it, which is the whole reason that question is now on the list. Three will stay at two hours and are performing acceptably at a lower return than underwritten. Two are in a market where the ancillary collapse was sharpest and are the weakest assets we own.

We have not written the duration question down as an answer, because we do not think it has one. Four hours is right today at most of the nodes we look at. What we have written down is the method: size against the shape of the opportunity, not the average of it, and buy the option to change your mind where the site allows it.