Is the NEM’s battery fleet starting to compete with itself?

Dan Lee recently asked on this site whether a “battery correlation penalty” is taking shape, drawing the parallel to the solar correlation penalty WattClarity named back in 2018: as more storage connects, competition for the same charging and discharging times grows, and the spreads that make storage money get competed away. Sound in theory, hard to check per unit, since the data hasn’t existed until recently. That piece worked from a single season and flagged spreads as worth watching over a longer window. This is that check: three years, per unit; not a final verdict, but a step past the single-season story.

We expected the story to be discharge-side, in the fading evening peak, instead the surprise was the midday lows, where batteries appear to be pushing up the cost of their own charging.

Key findings

• On ordinary days (no spikes), intraday two-hour spreads have roughly halved since their late-2024 peak, across all four mainland regions.
• Both ends of the day’s value have moved, but the pattern isn’t uniform. In Queensland and Victoria the floor crossed from negative to positive, in South Australia and NSW it moved little.
• On price-event days, the spread itself holds up regardless of how much storage is online, but per-unit capture of that spread shrinks as more storage competes for it, especially in Victoria.
• Events still pay; but they pay less per battery as the fleet grows.
• Fleet charge bids now sit right on the clearing price in most regions, in QLD and VIC the fleet is bidding positive prices to charge, up from a discount a year ago.
• The per-unit picture can be murky, depending on whether you count nameplate or what units actually offer.
• Capacity growth is tangled up with everything else that changed over this window. The trends are suggestive, but more data is needed before calling it definitively

What was analysed

Two measures, kept purposefully separate to avoid contaminating each other.

Price spread

The first is the electricity spot price, per region and per day. We take the average of the highest two hours of five-minute prices and subtract the average of the cheapest two hours. A simple metric, but it reflects the ideal hours of arbitrage activity, and so represents the gross spread a two-hour battery could earn on one cycle with perfect timing.

Percent of Perfect

The second measure involves the batteries themselves. For each unit on each day, a linear program works out the most energy-arbitrage revenue it could have earned over that day’s actual prices, given its energy and power limits and the availability it signalled to the market. “Perfect foresight” here means the optimizer knows that day’s prices in advance; it is not a claim that the battery could have earned this in real time. It is also energy-arbitrage only, with FCAS excluded, so it is a ceiling on the arbitrage component, not on a unit’s total achievable revenue. The optimiser treats each battery on the same terms to avoid biasing the comparison.

Price events

Days with any five-minute price over $300/MWh are treated as ‘event’ days and held aside. The $300 line is roughly where outcomes start to be driven by cap contracts and hedging rather than by day-to-day arbitrage, so setting these days aside keeps the spread measure focused on ordinary trading.

Price spread has roughly halved across every mainland region

Spreads were moderate through 2023, widened through 2024 to a peak late that year, and have fallen since. The drop runs through 2025 and into 2026, which is also when the BESS fleet grew fastest.

The chart below shows the size of the fall: from peaks of $200 to $280/MWh in late 2024 down to roughly $90 to $105/MWh by early 2026. The direction is consistent across all four mainland regions, despite different fleets, different climates and, in South Australia’s case, a much noisier path.

 

 Intraday 2-hour price spread by region, non-event day mean, 2023 to 2026

 

That consistency across regions is suggestive of a shared driver, but a shared downward trend is also exactly what you would expect if spread and offered capacity are both just trending over the same three years, regardless of whether one is causing the other. The regressions below should be interpreted with that in mind.

Regressing the daily spread against offered BESS capacity in each region returns a negative relationship everywhere: correlations of around -0.40 in NSW, -0.49 in Queensland, -0.37 in South Australia and -0.53 in Victoria. On their own these numbers are consistent with the correlation-penalty story, more offered capacity, smaller spread, but they don’t establish it. Both series trend over the same three years, and a level correlation between two trending series can look strong even with no causal link.

To get past that, we ran the same relationship on month-on-month changes rather than levels: does a month where offered capacity grows more than usual also see the spread fall by more than usual? In all four regions, no. The first-difference relationship is flat and not statistically significant anywhere. That result, is not just a caveat: with around fifteen monthly observations per region it’s a low-powered test, so this doesn’t rule out the mechanism. However it means the level correlations above are best read as consistent with the story rather than evidence for it. The case for “the fleet is doing this to itself” currently rests more on the rising-floor pattern below, and on market mechanisms, than on these regressions.

The floor is moving higher, and crossing zero in two regions

A version of this has been doing the rounds recently: a chart of the average Q1 price shape over the day, showing the evening peak softening and the midday dip lifting as more storage comes online. The same pattern shows up across all four mainland regions in the data, on ordinary (non-event) days.

 

Average Q1 intraday price shape by region, ordinary days only (event days excluded), 2023 to 2026. Darker lines are more recent years. The evening peak softens and the midday trough lifts as the fleet grows.

 

Split the spread into its two ends and the picture is more mixed than the headline trend suggests, but the floor is doing more than you’d expect if this were purely a fuel-costs or weather story.

The peak two hours soften as the fleet grows; the floor moves too, and in three of the four regions by as much or more. Queensland is the cleanest case: its peak fell from about $160/MWh to $110/MWh, while its trough went from around $-31 to $+12, crossing zero. Victoria shows the same crossing, trough from about $-48 to $+1 while its peak fell from $140 to $103. South Australia’s trough rose from about $-76 to $-12, the largest move of any region or either end, but stayed negative, while its peak edged up slightly. NSW is the exception: its trough barely moved (around $26 to $22) while its peak fell from $161 to $113, here the compression looks like a falling peak, not a rising floor.

 

Peak two hours vs trough two hours by region, showing the spread narrowing from both ends

 

This matters because of what can cause each end to move. A softer peak has a long list of explanations: cheaper gas, milder weather, more wind. A rising floor has a shorter one. Lower fuel prices pull the whole curve down, they do not lift it. In Queensland, Victoria and South Australia, something is soaking up the cheap midday surplus that used to go to whoever wanted to charge cheaply.

The largest new flexible load to arrive in the NEM over this period is by far battery charging, which points back at the fleet, even if it is not the only candidate, and even if NSW shows the floor isn’t moving everywhere. Where it is moving, the fleet looks like it may be bidding up its own fuel, raising midday demand into the solar trough it relies on for cheap charging.

This is the opposite direction to the better-known solar story, and worth not confusing with it. A decade of cumulative solar pushed the midday floor down, toward zero and below. What the last eighteen months show is the floor coming back up. Different driver, different timescale, and the recent move is the one that matters for charging cost.

The fleet is bidding up the floor

Everything above is inferred from outcomes. Prices rose at the bottom of the day, and we argued the fleet’s own charging is a plausible cause. That argument is stronger if we look at what batteries actually offer to pay to charge, not just at what the price did.

AEMO publishes the charge (LOAD) bids, so we can see the price at which each unit is willing to buy energy. We keep only the bids near where the market clears, within $200 of that day’s cleared charge bids, and volume-weight them into a fleet charge willingness-to-pay per region.

The reasoning to restrict to the middle of the stack is that charge bids are bimodal: capacity parked at the market floor around $-1000 that a unit will not charge with, and capacity parked near the cap that it must charge regardless of price. Neither end is a price a unit is competing at. Only the bids near the clearing level tell you what it is willing to pay for ‘fuel’.

 

Fleet charge willingness-to-pay by region, marginal bids only (within $200 of that day’s realised trough), non-event days, March 2025 to June 2026

 

The direction is the same as the floor. QLD and VIC move from bidding around $-55 and $-35 to charge, up to positive, around $+20 and $+30, over the window. SA rises hard but stays mostly negative, the same pattern its realised floor shows. In three of four regions the fleet is now willing to pay more to charge at midday than it was a year earlier, and in QLD and VIC it now bids positive prices where it used to require a discount. NSW is flat and shows no trend either way, consistent with it being the exception throughout this piece. One NSW unit runs under a system-security obligation that caps what it offers, the same one flagged earlier; excluding it barely moves the NSW line, so the flatness is real.

 

Fleet charge willingness-to-pay against the realised midday trough, by region, non-event days

 

Put the bids and the realised price (during charging) on the same axis and they sit almost on top of each other. The applied window would allow the fleet bid to land up to $200 either side of the clearing price, however it lands right on it or just below. That indicates a price-setting buyer: the fleet’s willingness to pay is the clearing price. The floor is not something happening to these batteries, it is the price at which they are bidding to buy.

It must be noted the measure is defined relative to the cleared price as mentioned previously, so the two lines moving together is partly built in. Following the price up and pushing it up could be an artefact of the method.

This does not settle direction on its own. But it feeds the floor argument from earlier: the fleet bidding at the clearing price is the demand-side buyer that section pointed to, now visible in what units offer to pay rather than only in the price that resulted.

Two honest limits on the measure itself. The charge-bid tables only go back to March 2025 in our data, so this is sixteen months, not the three years behind the rest of the analysis, a limit of what exists rather than a choice. And it only counts unit-days with bids near the clearing level, so it is conditional on a unit competing for cheap fuel that day, not a measure of the whole fleet’s appetite.

Price events still pay, but less per battery

On event days, the 2h spread itself shows no relationship with how much storage is online, in any region the correlation is essentially zero and not significant. The size of the opportunity on an event day doesn’t depend on how much BESS capacity is in the market. Hence the “events still pay” story.

But the LP-optimal $/MWh, what a unit could actually capture from that opportunity given its own size, does fall as offered capacity grows, in all four regions, and significantly in three of them. Victoria is the standout: r=-0.67, the strongest relationship anywhere in this analysis, non-event included. So the pie isn’t shrinking, but it’s being cut more ways. The opportunity an event day represents is holding up; what any individual battery can extract from it is not, at least not in Victoria, and to a lesser extent elsewhere.

This is consistent with what WattClarity has noted from the price-setting side: more batteries are involved in setting high prices, but that role is split across more units. Put together, the energy-arbitrage squeeze isn’t confined to the day-to-day cycle, it shows up on event days too, just in per-unit capture rather than in the size of the event itself. That’s a second angle alongside the FCAS saturation already documented since late 2023, not a separate story.

Nameplate vs offered capacity

There is a wrinkle worth being upfront about. Measure opportunity against nameplate, and per-MW arbitrage clearly falls as fleets grow. Measure it against what units actually offer to the market, and some of that washes out.

 

Arbitrage opportunity per MWh vs nameplate fleet energy (MWh), by region

 

 

Arbitrage opportunity per MWh vs offered fleet energy (MWh), by region

 

That is not necessarily a contradiction. The fleet offers well under nameplate on average, and we excluded units still in commissioning using a rolling cycling test. In NSW the gap is widest, partly because one large unit runs under a system-security obligation that caps the capacity it can offer to the energy market. More broadly, as spreads thinned, operators have adjusted what they offer and which markets they offer it into, alongside growth in contractual arrangements.

 

Offered BESS energy by region, raw (top) and as a share of nameplate (bottom).

 

Both NSW and VIC show temporary dips in the bottom panel that don’t appear in the top: NSW’s reflects its capped unit’s availability varying before settling (the excl.-capped-unit line is comparatively steady), and VIC’s reflects a temporary outage at a major unit in late 2024, not a structural constraint. Neither shows up as a dip in raw offered energy, which keeps climbing through both periods.

So per MW of what is actually in the market, returns have held up better than the raw price signal suggests, because the people running these assets are adjusting. Per MW of what has been built, they have not. Which number you care about depends on where you sit. If you run an existing battery, the offered number is closer to your reality. If you are financing a new one into the connection queue, nameplate is the denominator, and you have to consider revenue beyond energy arbitrage.

The big caveat

Fleet capacity went up over this three-year window, but so did many other things. Anything else that drifted over the same period, gas, coal availability, climate, could cause or contribute to similar effects. With this data alone we cannot fully separate capacity growth from a general trend.

One point is worth naming specifically. Behind-the-meter batteries, mostly residential, have grown fast over the same window, and they charge in the same midday trough. Our regression only sees offered grid-scale capacity, so some of what it attributes to the utility fleet may be the behind-the-meter fleet lifting the floor alongside it. That cuts the same way as our argument, more storage soaking up cheap midday energy, but it means the grid-scale number alone is not the whole cause.

What argues for this being real, despite the level correlations and the floor pattern both falling short of a clean causal story: the narrowing of the spread is concentrated in the period of fastest fleet growth, not spread evenly across the years, and in Queensland and Victoria the floor crossed from negative to positive over exactly that period, which a market-wide cost factor would not do on its own even if it can explain a softening peak. NSW and SA don’t fit that pattern as neatly. As more data comes in over the coming year or two, a repeat of the first-difference test will have more observations and more power, and should tell us more clearly whether this firms up or fades.

The charge-bid section is an attempt to get closer to the mechanism, by looking at what the fleet offers to pay rather than only at what the price did. It shows the fleet bidding to charge right at the clearing price, and that price rising, in three of four regions. But the bid measure is anchored to the trough, so it cannot on its own prove the fleet pushed the floor up rather than followed it. The case for pushing rests on there being no other flexible midday demand of that size. Treat it as consistent with the story and closer to the mechanism, not as the thing that settles the debate.

Consequences for the broader market

If this holds up, the arbitrage on offer to a new two-hour battery on an ordinary day is meaningfully thinner than it was eighteen months ago. The cheap-charging half of the business case is quietly eroding, and the spot revenue that still looks strong comes from events nobody can build a business plan around.

Batteries will keep their bread-and-butter arbitrage day to day, but it looks set to shrink as competition grows. Expect more operators to explore alternative commercial arrangements, or to look harder at the value of longer-duration systems, and there’s an early sign of that already in the data: fleet-average offered duration has moved from a wide 0.94-1.69h spread across regions in 2023 to a tighter 1.5-1.7h band by early 2026. The fleet isn’t just growing in capacity, it’s growing in duration.

None of this will surprise those building grid-scale batteries; the squeeze was already long theorised. What is useful for the broader market is seeing what batteries actually earned, and could have earned, rather than what a forecast might have assumed.

The per-unit numbers behind all of this analysis, actual revenue, the optimal benchmark, capture rates and the underlying dispatch, are available for every battery in the NEM. All data is sourced from AEMO. The optimiser and this read on the data are our own, and we welcome feedback on both.

 

---

About our Guest Author

---