The machines that bid our batteries

Every five minutes, whatever happens.

A rebid is a change to an offer after its prices are locked, at 12:30 the day before. A battery can move its megawatts between its ten prices, or change how many it has, and AEMO asks for a reason each time. For most units a rebid follows an event, such as a unit tripping or a forecast changing.

The busiest batteries rebid every five minutes whether anything has happened or not. Hazelwood, a 150 MW battery in Victoria, lodged a rebid in almost every five minute cycle its trading day was open. That came to about 471 a day, from just after 12:30 the day before until just before 04:00 at the end. That rhythm held for 27 batteries through September 2026. On a quiet Tuesday that month the fleet lodged 16,251 rebids. On the Wednesday in July when Victoria's price reached $19,070, it lodged 16,492.

In two years the market's batteries went from 7,877 rebids a day to 17,668, and from 59% of every rebid to 73%. Each one rewrites far more than the next five minutes. The median rebid in the largest family changed 16 intervals, starting with the very next one, and one in ten reached 18 hours or more ahead.

Their words give them away.

AEMO doesn't prescribe the words of a rebid reason, so reasons written by people vary and carry the odd typo. The busiest batteries repeat one sentence thousands of times a day with only the numbers changing. That is what software does.

ClearWe sorted every battery rebid since August 2024 by its wording alone. Then we checked those groups against two things that have nothing to do with wording. One is the format of the reference ID each lodgement carries. The other is the second, within the five minute cycle, when the rebid arrives. All three gave the same groups. They sort 45 of the 47 busiest batteries into seven families, and leave two whose words match nobody else's.

Four of the smaller families each centre on one company. AGL's four batteries and Wandoan South, EnergyAustralia's four, Shell Energy's three and CS Energy's two each write in their own words, at their own moments. Wandoan South is registered to a project company of its own, yet its rebids read like AGL's, and AGL dispatches it under contract. A fifth family stamps the time at the start of every reason, and all three of its batteries are run by Neoen.

The two largest families cross owners. We name them for the software that public announcements tie to their members, and the section after next says which batteries those announcements cover.

Two systems, many owners.

The largest family writes one sentence, DUE TO EXTERNAL PRICE FORECAST, or quotes AEMO's own price alerts back to it. Fourteen batteries wrote it in September 2026, adding up to 2,300 MW. They run from the 850 MW Waratah Super Battery in New South Wales to Phillip Island's 5 MW in Victoria. They are registered to fourteen participants, which belong to about eleven owners, since four are project companies of one developer. Their reference IDs share a format no other battery uses. The busy ones lodge within the same forty seconds of every cycle, most between second 207 and second 245.

Arriving together could be a coincidence, since every system reads the same AEMO forecasts. So we also looked at what the rebids did. For each pair of busy batteries we counted how often they lodged within two seconds of each other. Then, in each cycle where both changed their discharge offer, we checked whether they moved it the same way, towards dearer prices or cheaper ones.

SuggestiveInside the largest family, pairs lodged within two seconds of each other 13% to 68% of the time, and moved the same way 60% to 81% of the time. For most pairs from different families those figures fall to 0% to 3% and 45% to 55%, close to a coin toss. We read that as one system bidding all fourteen, most likely one provider's software or one trading service, reading the same forecasts for every battery it bids.

SuggestiveThe second largest family writes Change in forecast SOC, its forecast state of charge, and compares AEMO's forecasts in a fixed format. It spans sixteen batteries and 1,709 MW, from Hornsdale and the Victorian Big Battery to four small sites at Adelaide's water plants. It lodges a little earlier in the cycle and holds together more loosely. It shares its words but not its moves. Most of its pairs with different owners move the same way only 50% to 56% of the time, while Melbourne A1 and A2, which share an owner, do so 86% of the time. That fits one product that each owner runs for itself, rather than one desk.

Who the announcements name.

Vendors and owners announce these choices, so we went looking. Fluence says its bidding software Mosaic bids Waratah, Orana and Ulinda Park, Latrobe Valley, Queanbeyan and Terang. It says its earlier IQ software would bid Hazelwood. All seven are in the largest family, so we call it the Mosaic family. The other seven write the same words, in the same format, at the same second, and no announcement we found ties them to anything else. In its first weeks, on 4 September 2025, Ulinda Park even named Fluence's product in the reason it gave for a testing rebid.

Tesla's Autobidder is named for Hornsdale, the Victorian Big Battery, Bouldercombe and Lake Bonney. That is four of the second family's sixteen batteries, the Victorian Big Battery among them, so we call it the Autobidder family. The rest write in the same words, but no announcement names their software.

Neoen says software it built itself dispatches Capital Battery and the later stages of Western Downs. Those write in the time stamped family, which we call the Neoen family, while Western Downs 1 writes in the Autobidder family. No announcement we found puts a battery in the wrong family.

Two cautions. Tesla built some of the Mosaic family's batteries, Limondale among them, and who built a battery says nothing about what bids it. And the announcements say what each owner chose at the time, not what runs today. The rebids say that in September 2026 the same families were still writing the same words at the same second, so the two agree. The word autobidder on its own proves nothing, since batteries on both systems use it for their software when it fails.

Changing systems.

Batteries do change families. A new battery spends its first weeks on commissioning wording, often naming the test it's running. Then it moves into a family, usually when it starts trading in earnest. Eleven of the Mosaic family's fourteen batteries joined it this way after August 2024.

SuggestiveA battery can also change family within a few minutes. At 09:51 Brisbane time on 21 July 2026, Bulgana lodged a rebid that was time stamped, the way the Neoen family still writes. At 09:56 its reference ID took a new format. At 10:01 its reason was in the Autobidder family's words, and it has written them ever since. The registered participant didn't change, but the owner had, a year earlier. HMC Capital completed its purchase of Neoen's Victorian assets, Bulgana among them, on 1 August 2025. We read the switch as the battery moving to another bidding system. AEMO's data doesn't say why.

Two fair bars.

So most busy batteries are bid by a handful of systems, and two of them write most of the offers. Does the choice show up in what a battery earns? Answering that fairly is harder than it looks. Batteries sit in different regions, come in different sizes, sign different contracts and move the very prices they are paid. So we scored 45 batteries over the year to September 2026 against two bars.

The first is a ceiling. For every battery and every trading day, we worked out the most it could have earned from energy had it known every price in advance. The best possible day keeps to the battery's own offered limits. It leaves room for the FCAS it was enabled for, the services that keep the grid's frequency steady. It loses energy the way the battery's own records say it does, and it discharges no more than the battery really did. Over the year the median battery earned 69% of that. Eight in ten earned between 52% and 77%.

Here is one ordinary day. On 3 November 2025 Greenbank, in Queensland, charged through a morning of negative prices and discharged into the afternoon peak. It earned 73% of the best possible, close to the year's median. The best possible would have saved more of its energy for the two dearest hours.

A plain rule anyone could run.

A perfect day is a poor bar on its own. Nobody knows every price in advance, and the score it gives swings with luck. A battery's score against it in the first half of the year barely predicted its score in the second half.

So the second bar knows only what was public. Every five minutes, the plain rule reads AEMO's latest five minute and predispatch forecasts as they stood when the battery's offer locked. It plans the rest of the day on them, commits to the next five minutes, and is paid the price that actually happened. It never sees the day ahead. It may discharge as much as the battery averaged over the previous four weeks, and it aims to end the day holding what it started with. Anyone with AEMO's data could run it.

ClearOf the 45 batteries, 34 earned more than the rule over the year, by 8% at the median and 9% across the fleet. Unlike the score against a perfect day, this one lasts. Batteries that led the rule in the first half of the year mostly led it in the second, with a rank correlation of 0.58, against 0.18 for the best possible.

Where offers earn their keep.

Much of that lead is what an offer is for. The rule commits its megawatts on a forecast before the price is known. An offer sets a price for each band, so the battery only runs when the price clears it. Across the fleet, most of the extra money came from ordinary intervals between $100 and $300, worth 7% of what the rule earned. Spikes above $1,000 added 3%. Ordinary prices also hold most of what the batteries missed. Of the whole gap to the best possible, 66% sits in intervals between $100 and $300.

ClearSouth Australia is the exception, and its batteries led the most. All twelve we scored there beat the rule, by 23% across the region's fleet. In the other regions the fleet led by 3% to 6%. The median South Australian battery did even better, by half again, but that median is lifted by SA Water's four small batteries, which beat the rule by 59% to 67%.

SuggestiveSouth Australia's batteries didn't get closer to a perfect day. Their median score against it, 69%, is about the same as in New South Wales, 69%. The rule just does badly there. It earned only 46% of the best possible in South Australia, against 72% in Queensland. Spikes made the difference. In South Australia, intervals above $1,000 added 16% to what the rule earned, the most of any band. In every other region the batteries did no better than the rule in spikes, and their lead came from ordinary prices, worth 7% to 8%. South Australia's price also swings the most, with a standard deviation of $474 a megawatt hour against $88 to $263 elsewhere. A rule that commits on the forecast pays for every miss. An offer that waits for the price doesn't.

The lead belongs to the offer, whoever writes it. The four batteries that rebid least, Eraring, Supernode 1, Swanbank and Tarong, each lodged fewer than 50 rebids a day against hundreds for most. All four beat the rule too, and within regions, batteries that rebid more often did slightly worse. That comes with a caution. All four are among the seven largest batteries we scored, and three are run by generators with trading desks of their own.

Mosaic or Autobidder.

Now the question the first half of this article leaves open. If the two big systems bid batteries differently, their families should separate in their scores. Each battery has to be compared with others in its own region, because region alone explains 38% of the differences in how batteries do against the rule.

InconclusiveWithin regions the two families sit close together, and they even swap places between the bars. Against the plain rule, the Mosaic family's eight scored batteries sat about 7% above their regions' medians, and the Autobidder family's thirteen about level with theirs. Against the best possible the order flips, with the Autobidder family about 2% above and the Mosaic family level. Neither gap is beyond chance. Shuffle the two families' labels at random and a gap as wide turns up 30% of the time on the first bar and 68% on the second.

The closest things to natural tests agree. Western Downs 1 and 2 share an owner and a site but write in different families, and they landed at 100% and 97% of what the rule earned. Hornsdale and Blyth share an owner and a region and also write in different families. They landed at 129% and 112%. So one pair landed together and one landed 17 points apart, and both matched or beat the rule. Across all the families, the family a battery writes in explains 30% of what region leaves, and a random shuffle explains as much 11% of the time. The owner does no better, at 40%, matched by a shuffle 22% of the time. With 43 batteries, and most families and owners holding only a few, a year can't tell them apart.

One more thing tilts the ceiling. When a big battery discharges more, the price falls, because it pushes out the dearest megawatts that would otherwise have run. A perfect day ignores that. We priced the best possible day again on the offers the rest of the market made, and for the median battery it shrank by 9%, more for the biggest. That lifts the median score from 69% to 77% of the best possible. The rule pays for it too, and priced the same way 39 of the 45 batteries beat it. None of this moves the two big families apart.

The job sets the score.

ClearThe lowest scores tell the same story from the other end. Leave out the two batteries held for the network and one pattern stands out. The three batteries that did the most regulation were Bulgana, Bouldercombe and Phillip Island. Regulation is the FCAS service that follows AEMO's signal up and down every few seconds. On average they had 27%, 21% and 20% of their megawatts enabled for it, against 2% for the median battery. They are also three of the four lowest scores against the rule, which chance alone would rarely produce, and the pattern holds month by month. Regulation asks for a battery kept near the middle of its range, ready to move either way, and the energy it leaves untraded counts against it here. It was traded for something else. FCAS made up 58%, 44% and 45% of what these three earned, against 4% for the median battery.

ClearBouldercombe shows it best, because its job changed. Tesla bids it under a revenue sharing deal with its owner. From October to February, FCAS was 34% to 75% of its revenue, regulation took about 33% of its megawatts, and it trailed the rule in all five months. From March its regulation averaged 12%, it discharged 71% more energy a month, and it beat the rule in six of the seven months to September. Its score against the rule went from 74% to 106%. Over the same months Queensland's other eight batteries moved by a median of 4 points, and Bouldercombe moved by 31, the most of any. The same battery and the same software family, given a different job, earned a different score.

SuggestiveBulgana's battery was built to keep supply steady for a glasshouse beside its wind farm, and the glasshouse's owner had collapsed by the time the battery switched on. In our year it was run mainly for regulation on both the systems that bid it, with 28% of its megawatts enabled before its change in July 2026 and 24% after. Its score against the rule went from 29% in the three months before to 36% in the two after, while Victoria's other batteries held near 98%. Changing system barely moved its score, because its job didn't change. Phillip Island was built for AusNet to support the island's supply through peak holiday periods. That role doesn't show in the records, so we can't say how much of its year it shaped.

SuggestiveThe next two lowest do little regulation, and the records point elsewhere. AEMO marks Capital Battery as locally constrained in 11% of intervals, against 2% for the median battery. On its most constrained fifth of days it fell 17% short of the rule, against 3% on the rest. Our bars ignore the network inside a region, so on those days they ask for megawatts it may not have been able to send. Wandoan South is dispatched by AGL under a 15 year contract. On average it went into each day's twelve dearest intervals fuller than any other Queensland battery, at 55% of its range against a median of 42%. That looks like a retailer keeping energy back as cover for its customers. Its contracts are private, so we can't confirm it.

What else could explain it, and what we can't see.

• A shared trigger. Every system reacts to the same AEMO publications, so some lining up is expected. But each family keeps its own second, and pairs from different families rarely lodge within two seconds of each other.
• A shared lodging tool. Several owners lodge through tools that stamp the same reference format, which is why we never sort by references alone. The Mosaic family's format appears nowhere else.
• A shared owner further up. Some participants may share an owner the registration list doesn't show. Even so, the Mosaic family's participants include companies as unlike as a metering business and a hydro developer.
• Contracts with the network. Most of the Waratah Super Battery is held ready to protect the network, and the Victorian Big Battery holds most of its capacity for the network through summer. Their offers still show those megawatts, so their scores say little about their bidders. We mark both and leave them out of the comparisons.
• Hedges. A battery that has sold cap contracts may hold energy back for a spike that never comes. That costs it against both bars here and pays it elsewhere.
• FCAS and wear. We score energy only and hold each battery's FCAS as it was. A bidder that gives up energy to earn more in FCAS, or cycles less to spare its cells, looks worse here than it is.

What it means.

For a trader, families matter more than owners. About six in ten megawatts of the busiest batteries write in the words of the Mosaic or Autobidder families. The registration list shows 14 participants behind the Mosaic family, yet in the market they behave like one bidder. When the forecast it reads moves, 12 busy batteries in four regions rewrite their offers within the same forty seconds, and mostly in the same direction.

SuggestiveIt also means a fault travels. At 1:26am on 24 September 2026 Blyth lodged a rebid giving its reason as an autobidding error. Seven minutes later Capital Battery, in another region, gave the same reason. Both write in the Neoen family, so we read it as one system failing once.

For an owner, the job and the region matter more than the vendor. Most offers beat a plain rule anyone could run, and the batteries that led kept leading. The lead was largest in South Australia, where prices are hardest to call. A battery held for regulation, a network contract or a retailer's cover scores lower for reasons that have nothing to do with its software.

Sharing software isn't wrong. Owners buy bidding software the way they buy anything else, and plenty of them have chosen the same product. A year of scores didn't show that the choice between the two big systems changed what their batteries earned.

In supagrid.

• Open any battery and point at any five minutes. Its offer sits beside what it charged and discharged, with the time it was lodged, which rebid of the day it was and the reason given.
• Ask the Analyst what a battery was offering through a price spike and what each rebid changed.