1From A$1.85 billion to A$43.7 billion in just over a year
Each private round was priced above the last, and every round whose investors were named included a company that did business with Firmus. About eight weeks after Maas Group bought shares on 4 August, the offer price was 2.39 times the A$4.60 a share it paid, adjusted for the proposed share split.
Between September 2025 and August 2026 Firmus raised equity in four rounds with stated valuations, and each was higher than the last. The company’s own releases give the valuations after the money: A$1.85 billion19 in September 2025, about A$6 billion20 in November, US$5.5 billion21 in April 2026 and more than US$10.5 billion22 in August. The press converted the August figure to A$15.5 billion23.
The offer at A$11 valued Firmus above every private round, and each lower price reported in the final days still sat above the valuation stated for the 7 August round.
Firmus equity valuations, September 2025 to the withdrawn offer, A$ billion, log scale
Each private round at its stated valuation after the money, the offer at A$11 a share, and the lower prices reported in the last days of the bookbuild, drawn as one range. US dollar rounds are converted at the Reserve Bank’s rate of 9 October 2026. None of the lower prices was announced or completed.
Firmus raised equity four times in under a year at stated valuations, each higher than the last.
Equity raised by Firmus, September 2025 to October 2026
Each round in the currency its source used, with the valuation as stated. The offer and the lower prices reported in October are shown for comparison; none of the lower prices was announced or completed.
| Date | Round | Amount | Investors named | Valuation, as stated |
|---|---|---|---|---|
| 16 Sep 2025 | Placement19 | A$330m | Ellerston Capital (cornerstone), NVIDIA, Australian institutions and high-net-worth investors | A$1.85bn after the money |
| 14 Nov 2025 | Equity raise (commitments)20 | A$500m | Not named; Morgans sole lead manager | About A$6bn after the money |
| Feb 2026 | Strategic investment25, 26 | A$100m reported; A$110m in Maas Group’s results | Maas Group | About A$6bn, as reported |
| 7 Apr 2026 | Strategic round, subject to closing21 | US$505m | Coatue (lead), NVIDIA | US$5.5bn after the money |
| 4 Aug 2026 | Strategic investment14 | A$300m at A$230 a share | Maas Group | Not stated |
| 7 Aug 2026 | Strategic round (fully committed)22, 23 | US$2bn | Coatue and NVIDIA (follow-on), Blackstone funds (new), Jane Street | More than US$10.5bn after the money; A$15.5bn as converted by the press |
| 1 Oct 2026 | Offer at A$11 a share3, 17, 27 | A$7.1bn base raise (US$4.94bn) | Bank of America, JPMorgan, Morgan Stanley and Morgans as joint lead managers | A$43.7bn, as reported |
| 8 to 9 Oct 2026 | Lower prices reported3, 10, 11, 12, 13 | A$9, A$8.25 and A$5.50 a share | None announced or completed | About A$35.8bn, A$32.8bn and A$21.9bn (derived) |
| 9 Oct 2026 | Offer withdrawn2, 15 |
The offer price set a step beyond the August rounds. Maas Group bought shares at A$23014 each on 4 August, which is A$4.602 after the proposed 50:1 share split. The offer price of A$11 was 2.39 times2 that price, about eight weeks later.
The private rounds were priced by investors who also did business with the company. NVIDIA, which supplies Firmus’s GPUs and was named by CDC as the anchor customer at Melbourne, invested in the September 2025, April 2026 and August 2026 rounds4, 28, 29. Blackstone funds, which lead the US$10 billion30 debt facility agreed in February 2026, came into the August round22. Maas Group, whose subsidiary JLE has received about A$1.1 billion2 of Firmus work orders, invested in February and August14, 26. A price agreed by investors with commercial relationships of this kind tells a reader less about value than a price set by investors without them.
2What A$11 required
In the model’s central case, A$11 needed each megawatt to earn about US$30.2 million a year. No cell in the grid falls to the US$15.5 million that the offer documents’ own projects imply, even at the projects’ own margin.
The model asks what Firmus would have had to earn for its discounted free cash flow to equal the enterprise value implied by each price. It takes the capacity build from the offer summary as reported: 46 MW in operation, 865 MW across five developments, and the timetable given for each site5. It takes the capital cost of servers and buildings from the same reports, replaces servers every four1 years, and solves for the revenue each megawatt must earn.
Many combinations of capacity, pricing, margin and discount rate reach the same value, so exhibit 4 shows a grid. Each cell is the revenue per megawatt a year at which the discounted cash flows equal the price, for one cash margin and one discount rate. None of the cells is a forecast.
Every cell needs more revenue per megawatt than the US$15.5 million a year that the offer documents’ projects imply.
Revenue per megawatt a year consistent with each price, US$ million, as at 12 October 2026
Consistent with the price under the stated model. Equivalent critical IT megawatts; contracted capacity profile; debt and cash treated as neutral. Capacity, timetable and capital cost are the offer documents’ as reported. Shading compares each cell with the contract rates that the listed GPU cloud operators Nebius and IREN report in their filings.
| Cash margin | ||||||
|---|---|---|---|---|---|---|
| Discount rate | 50% | 59% | 60% | 70% | 80% | 91% |
| 8% | 32.7 | 27.7 | 27.2 | 23.3 | 20.4 | 17.9 |
| 10% | 36.3 | 30.7 | 30.2 | 25.9 | 22.6 | 19.9 |
| 12% | 40.0 | 33.9 | 33.3 | 28.5 | 24.9 | 21.9 |
| 14% | 43.8 | 37.1 | 36.4 | 31.2 | 27.3 | 24.0 |
| Discount rate | ||||
|---|---|---|---|---|
| Cash margin | 8% | 10% | 12% | 14% |
| 50% | 32.7 | 36.3 | 40.0 | 43.8 |
| 59% | 27.7 | 30.7 | 33.9 | 37.1 |
| 60% | 27.2 | 30.2 | 33.3 | 36.4 |
| 70% | 23.3 | 25.9 | 28.5 | 31.2 |
| 80% | 20.4 | 22.6 | 24.9 | 27.3 |
| 91% | 17.9 | 19.9 | 21.9 | 24.0 |
- At or below Nebius’s 2026 base of about US$12m
- Above that base, up to US$20m
- Within the US$20m to US$25m of recent large contracts
- Above any contract rate reported
- The central cell, 60% at 10%
For comparison, US$ million per MW a year. The offer documents’ project revenue: 15.51, 5, 9. The revenue that only recovers the build, with the equity worth nothing, in the central cell: 22.51. Contract rates reported by listed operators: Nebius, 2026 base, 126; Nebius, Q2 2026 contracts, 20 to 256; IREN, recent three-year contracts, more than 207; IREN, Microsoft contract, 9.78.
| Price a share | Reported timetable, 911 MW | Slower timetable, 911 MW | Reported timetable, at the announced 1,600 MW | Facility fully drawn | Per GPU-hour, central |
|---|---|---|---|---|---|
| A$11 | 30.2 | 30.8 | 27.1 | 32.8 | US$7.88 |
| A$9 | 28.8 | 29.2 | 26.3 | 31.4 | US$7.51 |
| A$8.25 | 28.3 | 28.6 | 26.0 | 30.8 | US$7.38 |
| A$5.50 | 26.4 | 26.5 | 24.9 | 28.9 | US$6.87 |
- At or below Nebius’s 2026 base of about US$12m
- Above that base, up to US$20m
- Within the US$20m to US$25m of recent large contracts
- Above any contract rate reported
A megawatt of revenue can also be read as a GPU-hour. The model converts at the GPU density Firmus planned for Launceston, with every hour billed4, 5, so at A$11 the central cell needs about US$7.881 for every hour of every GB300-equivalent GPU. That density is about 438 GPUs4, 5 for each megawatt of critical IT. The Launceston density is below the 533 GPUs31 that a megawatt of GB300 NVL72 rack power holds, at 72 GPUs and 135 kW a rack at its rated power, and the rack figure would give a lower price per GPU-hour. For context only, on 12 October the median listed on-demand price for a B300 was US$8.9932 an hour, and three-year reservations were listed about 49%33 below on-demand prices. Firmus’s contracts are mainly take-or-pay agreements whose prices are not public5.
The offer documents’ own project figures point lower. Summed across the disclosed projects, 12-month contracted revenue works out at about US$15.5 million1 a megawatt a year on the 912 MW pipeline, or about US$4.04 a GPU-hour5, 9. The central cell at A$11 needs almost twice1 that rate.
The project sums also carry a higher margin than the central cell5. At a 91%1 margin, A$11 needs about US$19.9 million a megawatt a year at a 10% discount rate. No cell at that margin falls to the project figure: the lowest, at A$5.50 discounted at 8%1, is US$16.1 million.
In the central cell about 74%1 of the value lies beyond the tenth year of the forecast. Most of the price therefore rested on what the business would earn after 20361, from servers that had not yet been bought.
Most of the requirement pays for the build itself. With the equity worth nothing, the model still needs about US$22.5 million1 a megawatt a year to recover the servers, their replacement and the buildings at a 10% discount rate. That is 75%1 of the requirement at A$11 and 85% at A$5.50. So the lower prices reported in the final days lowered the requirement only to about US$28.8 million1 at A$9, US$28.3 million at A$8.25 and US$26.4 million at A$5.50.
The grid treats the company as free of net debt, because no balance sheet was lodged. If the US$10 billion facility were fully drawn, the enterprise value at the same share price would be higher, and the central requirement rises to about US$32.8 million1, 30. The drawn balance has not been disclosed5.
At A$11 each megawatt in operation carried about A$950 million of enterprise value, against A$71.7 million for each active megawatt at CoreWeave.
Value per megawatt, A$ million, at 9 October 2026
Firmus at each price, against the operating capacity and against the pipeline. Comparables on their own stated bases. CoreWeave is an enterprise value, from its filings and closing price; CDC is an equity valuation, because its net debt was not found; AirTrunk is a 2024 transaction. Australian dollars at the Reserve Bank’s rate.
Firmus at the offer and at the lowest reported price
- Firmus at A$11, per operating MW1, 5EV, debt and cash neutral, over 46 MW operatingA$950m
- Firmus at A$11, per pipeline MW1, 9EV, debt and cash neutral, over the 912 MW pipelineA$47.9m
- Firmus at A$5.50, per operating MW1, 5EV, debt and cash neutral, over 46 MW operatingA$475m
- Firmus at A$5.50, per pipeline MW1, 9EV, debt and cash neutral, over the 912 MW pipelineA$24.0m
Listed GPU cloud operator
Data centre landlords
- AirTrunk, per committed MW (2024 transaction)1, 36EV of more than A$24bn over more than 800 MW committedA$30.0m
- CDC, equity per deployed MW1, 37, 38Equity valuation, 30 Sep 2026, over 350 MW deployed; net debt not foundA$52.7m
- CDC, equity per contracted MW1, 37, 38Equity valuation, 30 Sep 2026, over 1.1 GW contracted; net debt not foundA$16.8m
Exhibit 5 puts the same prices on a per-megawatt basis. Against the 46 MW in operation, A$11 is about A$950 million1 of enterprise value per megawatt. Against the 912 MW pipeline it is about A$47.9 million1. CoreWeave, a listed GPU cloud operator, was valued at about A$71.7 million1 per megawatt of active power and A$29.1 million per megawatt contracted. The two figures for Firmus frame the question the price turned on: how much of the pipeline would operate, and when.
3Profitability
The 2029 target of US$5.8 billion is close to a sum of project margins before corporate costs. On a four-year server life the same projects earn about US$3.8 billion of EBIT a year.
The offer documents, as reported, describe a company at the start of its build. In the year to 30 June 2026 Firmus earned US$50.8 million of operating revenue and lost US$68.2 million18. The draft prospectus forecast a pro forma loss of $77 million18, 39 for the half year to December 2026.
The forecast that carried the price was for 2029: EBIT of US$5.8 billion18 on revenue of US$12.9 billion, provided seven data centres were operating by the end of 2028. About 85%18 of it was expected from the developments in Malaysia and Indonesia, which were still under construction.
The 2029 figures are close to a sum of project-level numbers. Startup Daily’s addition of the disclosed projects gives 12-month contracted revenue of US$14.15 billion5, EBITDA of US$12.90 billion and EBIT of US$5.82 billion. Those are margins of 91%5 and 41%, before corporate costs, over periods that do not coincide. CoreWeave’s adjusted EBITDA margin in the June 2026 quarter was 59%34, after its corporate costs.
On a four-year server life the projects’ EBIT falls from US$5.82 billion to about US$3.8 billion a year.
From project revenue to EBIT, US$ billion a year, as reported and on a four-year server life
The reported project sums, as added by Startup Daily, and the same projects with servers depreciated over four years against the server budgets. Before corporate costs. The project periods do not coincide, so the bridge is arithmetic on reported totals. The 2029 target is the pathfinder’s, as reported by The Nightly.
The gap between project EBITDA and EBIT is depreciation and amortisation, about US$7.08 billion5 a year. Set against the US$35.97 billion5 of server budgets disclosed for the sites, it is consistent with servers written off over about 5.11 years. On a four-year life, the replacement cycle the model uses, the same projects would earn EBIT of about US$3.8 billion1 a year.
At the projects’ own price of US$4.04 a GPU-hour, the pipeline reaches the 2029 target only with every hour billed and servers lasting about six years.
Annual EBIT at full build, US$ billion, by price per GPU-hour, share of capacity billed and server life
The pipeline at full build, before corporate costs, from the project sums as reported and the server budgets. The third column is the price the reported project revenue implies; the others are 60%, 80% and 120% of it. The model’s other cash costs are set to reproduce the reported project revenue and EBITDA sums, which cover different twelve-month periods. Brackets mark a loss. Listed GPU rental prices are cited for context only32, 40, 41, 42, 43, 44, 45.
| Price per GPU-hour | ||||
|---|---|---|---|---|
| Server life | US$2.43 | US$3.23 | US$4.04 | US$4.85 |
| 3 years | (4.8) | (2.0) | 0.8 | 3.7 |
| 4 years | (1.8) | 1.0 | 3.8 | 6.7 |
| 5 years | (0.03) | 2.8 | 5.6 | 8.5 |
| 6 years | 1.2 | 4.0 | 6.8 | 9.7 |
| Server life, years | ||||
|---|---|---|---|---|
| Price per GPU-hour | 3 | 4 | 5 | 6 |
| US$2.43 | (4.8) | (1.8) | (0.03) | 1.2 |
| US$3.23 | (2.0) | 1.0 | 2.8 | 4.0 |
| US$4.04 | 0.8 | 3.8 | 5.6 | 6.8 |
| US$4.85 | 3.7 | 6.7 | 8.5 | 9.7 |
- At or above the 2029 target of US$5.8bn
- Positive, below the target
- Negative, shown in brackets
Taking the reported project sums as they stand, at the project price of about US$4.04 the pipeline reaches the 2029 target only with every hour billed and servers lasting about six years, when it earns US$6.8 billion1. On a five-year life it earns US$5.6 billion1, just short of the target, and on a four-year life US$3.8 billion (exhibit 7).
4The record against each requirement
The record supports the contracts as reported and the first 46 MW in operation. It shows each dated target set since 2025 whose date has passed as missed, and no total has been disclosed for the debt that would fund the build.
Capacity
Firmus reported 46 MW in operation across Singapore and Melbourne, including about 42 MW5 at CDC’s Brooklyn campus serving Meta. Five developments would add 865 MW: Launceston 84 MW5, Wesley Vale 23 MW, Batam 360 MW, and two sites in Kuala Lumpur of 192 MW and 206 MW. With the operating sites they come to 911 MW5, against the 912 MW that Firmus calls contracted and that Bloomberg’s reading of the investor documents calls a pipeline9, 46. The basis of these megawatt figures is not stated. Long Reach in Tasmania and the two South Australian sites sit outside them5.
Of the 912 MW that Firmus calls contracted, 46 MW was in operation.
Capacity by stage, megawatts, as reported to 12 October 2026, and the five developments with their disclosed costs
Megawatts as each source states them, from the offer documents as reported by Startup Daily and Bloomberg, from Firmus, and from CDC’s chief executive as reported by Rampart. The bases are not stated and are not converted. The Southgate figure was the Australian programme announced with CDC in October 2025.
| Site | MW, as stated | Status | Target | Disclosed capital cost |
|---|---|---|---|---|
| Launceston, Tasmania5, 10 | 84 | Under construction | Early 2027 | US$3.48bn |
| Wesley Vale, Tasmania5, 10 | 23 | Development approval pending | 2027 | US$0.90bn |
| Batam, Indonesia5, 10 | 360 | Under construction in leased space | Phased from early 2027 | US$14.46bn |
| KVTP, Kuala Lumpur5, 10 | 192 | Planned | 2027 | US$8.38bn |
| SVTP, Kuala Lumpur5, 10 | 206 | Planned | 2027 to 2028 | US$8.98bn |
| Total5, 10 | 865 | US$36.20bn |
| Basis | US$ | What it covers |
|---|---|---|
| Pathfinder prospectus, as reported | US$37.7 billion18 | The cost of building the five development sites. Exhibit 10 uses this total, because it is the offer documents’ own figure for the programme. |
| Sum of the site costs, as tabled | US$36.20bn5, 10 | The five rows above, as Startup Daily tabled them10. The table does not name the document it was taken from. |
| Server budgets, as reported | US$35.97 billion5 | Servers alone, at every disclosed site, including Melbourne, which is in operation. It covers a different scope from the two totals above. |
The two totals for the five sites differ, and the record does not say why. The pathfinder total is higher than the sum of the sites by US$1.50 billion10, 18. On the sum of the sites, the residual in exhibit 10 would be about A$32.7 billion1 rather than A$34.9 billion. The server budgets, which include Melbourne, are within US$1.73 billion5, 18 of the pathfinder total, so servers rather than buildings carry most of the programme’s cost.
Delivery against announced dates
Of the dated targets Firmus and its partners have set since 2025 whose dates have passed, the record shows each one missed (exhibit 9). Earlier, its Singapore cloud opened in 2023, roughly when it said it would49, 50. Melbourne, the one large site in operation, was being handed to its customer about six weeks after its first-quarter target51, 52. The Launceston AI factory, announced for 2026, now targets early 20275, 53, 54. The South Australian development application due at the end of September 2026 had not been lodged by 12 October, and Firmus was still seeking Crown Sponsorship for both sites55, 56. The Wesley Vale application awaited approval when the offer documents were sent10.
On the public record, every dated target set since 2025 whose date has passed was missed.
Targets and outcomes, Firmus and its partners, 2023 to October 2026
Each row is a dated target given by Firmus or a partner, directly or as reported, against the first public evidence of what happened, each with its own notes. A quarter or half is read as its last month and a year as December.
| Commitment | Target, as stated | What happened | Gap |
|---|---|---|---|
| Melbourne cluster of 18,500 GB300 GPUs live | Q1 2026; online by April 202647, 51 | Being handed to a customer on 11 May 2026; operating by 8 September 202646, 52 | About six weeks to handover |
| Launceston AI factory, Stage 1, 90 MW | By 2026; operational by mid-202653, 54 | Building complete about November 2026, compute commissioned in Q1 2027; operations from early 20275, 52 | About nine months, if early 2027 holds |
| First stages, up to 150 MW and 54,000 GB300 GPUs | Mid-202647 | Melbourne only, about 42 MW, operating by September 20265, 46 | Tasmanian share outstanding |
| Tasmanian campus of about 36,800 GB300 GPUs complete | Second half of 2026; late 20264, 51 | Compute now due in Q1 202752 | At least one quarter, if Q1 2027 holds |
| Project Southgate with CDC | Up to 1.6 GW through 202847 | Ended with about 42 MW deployed48, 57 | Partnership ended |
| Tailem Bend development application | Filed at the end of September 202655 | Not lodged; Crown Sponsorship still sought on 12 October 202656 | Running |
| SMC zone in Singapore (SIN01) live | Second half of 202349 | Live in Singapore in 202350 | On time, approximately |
| SMC zones beyond Singapore (India, Thailand, Europe) | 2023, then the first half of 202449, 58, 59 | No operating zone outside Singapore confirmed; India absent from the September 2026 footprint46 | Not delivered, as far as the record shows |
| ASX listing | Trading from 22 October, then 23 October 202627, 60 | Book closed without a price on 8 October; application withdrawn on 9 October 20262, 3 | Withdrawn |
Partners
The Southgate programme with CDC was announced in October 2025 with a potential of up to 1.6 GW47 through 2028. It ended at about 42 MW48. The two companies describe the ending differently. Firmus says they “mutually agreed earlier this year” not to proceed57. CDC’s chief executive said CDC was not planning a 1.6 GW roll-out because Firmus “made other choices” about developing its own sites, and a CDC spokesperson said CDC now treats Firmus as a customer5, 61. Firmus keeps the Melbourne hall and will not expand it to the 150 MW61 once planned.
Power
In Tasmania, Firmus holds a three-year contract with Aurora Energy for 104 MW52 at Launceston, executed in March 2026 at market rates62. Its request for about 444 MW63 of long-term supply for its three Tasmanian sites awaits a decision by Hydro Tasmania, expected by the end of 2026. In South Australia it has a 12-year agreement with Gunvor for 600 MW64 of firm supply, but no development application56. No Firmus power arrangement for Batam or Malaysia is public.
Funding
The five developments were reported to cost US$37.7 billion18 in the pathfinder, about A$54.0 billion1 at the model’s exchange rate. Exhibit 8 sets that total beside the other two costs in the record. Against it, the public record shows equity raised in the year to August 2026 of more than US$3 billion22 and the US$10 billion facility, whose drawn balance and terms are not public30. Taken at face value, the equity and the facility leave A$34.9 billion1 of the programme to be funded from other sources, such as further equity, project debt, customer prepayments or vendor finance (exhibit 10). The offer’s base raise of A$7.1 billion17 would have met part of it, and was withdrawn2. The figure sets the programme’s gross cost against the funding on the public record, and leaves out operating cash flow and the timing of spending. The offer documents reportedly said management expected adequate equity funding for the development sites for 12 months after the offer5.
The equity raised and the facility at its limit leave A$34.9 billion of the programme to be funded from other sources.
The five-site programme against funding on the public record, A$ billion
Gross programme arithmetic. The cost is the pathfinder’s total for the five sites, as reported, at the model’s exchange rate; exhibit 8 sets it beside the other costs in the record. The facility is shown at its limit because no drawn balance is public. The offer’s raise is shown but was not completed.
Capital intensity
The disclosed construction budget for the two Tasmanian sites is about US$5.9 million5 per megawatt, against US$14 million to US$16 million8 for IREN’s liquid-cooled build at Childress. The pathfinder claimed a construction cost of US$6 million18 per megawatt of computing power sold, against an industry average of US$11.3 million. Server budgets run at US$32 million to US$42 million5 per megawatt, against about US$29 million8 for IREN’s GB300 systems. The building cost is low and the server cost is high relative to IREN, and the servers are where most of the capital sits.
5What the record supports, and what it does not yet show
Under the stated model, every price tested needed more revenue per megawatt than the offer documents’ projects imply. The record supports the contracted demand as reported. It does not establish the delivery pace, the margins or the funding the price assumed, and the private rounds are weak evidence of value.
This section gives inashanu’s view of the evidence. Each view is stated with the facts it rests on. None is a view on whether to buy, sell or hold any security.
- Contracted demandSupported by the record, as reported
- The offer documents reported US$72.8 billion18 of revenue secured under contracts, mainly take-or-pay, and Meta, OpenAI, NVIDIA and Fireworks have each been named as customers4, 5, 28, 46. No contract has been published, and their terms, termination rights and credit support are not public.
- Delivery at the pace the price assumedNot established by the public record
- The price assumed the build followed the reported timetable, taking Firmus from 46 MW in operation to the full pipeline by the end of 202810, 46. Of the dated targets set since 2025 whose dates have passed, the record shows none met on time. The CDC partnership that was to carry the Australian programme has ended, and the development approval for Wesley Vale was still pending (exhibit 9).
- Margins at the level the target assumedNot established by the public record
- The 2029 target is close to a sum of project margins before corporate costs: 91% at EBITDA, against 59% for CoreWeave after its costs. On a four-year server life the same projects earn about US$3.8 billion of EBIT, against US$5.82 billion as reported. Even at the project margin, every price in exhibit 4 needs more revenue per megawatt than the projects provide.
- Funding for the buildNot established by the public record
- No total was disclosed for the debt that would fund a programme costing US$37.7 billion, against more than US$3 billion of equity raised in the year to August 2026. The facility’s drawn balance and terms are not public.
- The private rounds as evidence of valueWeak evidence
- Every round whose investors were named included the company’s GPU supplier, its lender or its contractor: NVIDIA in September 2025 and in April and August 2026, Blackstone funds in August, and Maas Group in February and August14, 22, 26, 29. When public investors were asked to price the plan, the reports of lower prices did not produce an offer the board would accept.
The requirement against the record
On the evidence available when the offer was made, the public record did not establish what A$11 required. Under the model, every price reported, at every margin and discount rate in exhibit 4, needed more revenue per megawatt than the offer documents’ own projects provide. Delivery at the pace assumed had not been shown, and the margins and the funding rested on plans that had yet to produce results. The book did not clear at the lower prices reported either. The contracts are substantial, and the record will change as the sites are built and begin to bill.
6What it means for Australia’s AI infrastructure pipeline
Of the Australian programme announced at up to 1.6 GW through 2028, about 42 MW operates, and 88% of the capacity Firmus had under development was in Indonesia and Malaysia.
Firmus is one case of the gap between announced and operating capacity that inashanu’s Australian data centre study measures. Its Australian programme was announced at up to 1.6 GW through 2028, and about 42 MW of it operates, at Melbourne47, 48. Of the capacity under development, 88%5 was in Indonesia and Malaysia. The study’s register follows each Australian project from announcement through approval, power and construction to operation. Read the Australian data centre study
Since publication, 15 October 2026
Firmus withdrew its offer on 9 October2. Shares in listed GPU cloud operators had fallen in the days before the withdrawal and fell further after it, most sharply on 14 October65, 66, 67. That day CoreWeave closed at US$65.7865, about 20%35 below the US$82.08 close of 9 October used in exhibit 5. Nebius and IREN closed 16%66, 68 and 13%67, 69 below their closes of 9 October.
On 14 October CoreWeave’s unsecured notes due 2032 traded below 80 cents70 on the dollar, according to TRACE pricing data, a price at which they yield more than 15%71. That is a yield on debt, which ranks ahead of equity, and it is above 14%1, the highest discount rate in exhibit 4. At 14% and the central 60% cash margin, the model needs about US$36.4 million1 a megawatt a year at A$11, or US$9.49 a GPU-hour, and about US$30.4 million at A$5.50, against the US$15.5 million that the projects imply.
Over the same days Ornn’s H100 rental index rose from US$2.6540 a GPU-hour on 11 October to US$3.2172 on 14 October; the project price in this analysis comes from contracted revenue and does not change.
On 9 October Bloomberg reported that Firmus was exploring a private round of US$2 billion to US$3 billion73, 74 with existing investors. If a round of US$3 billion73, 74 closed, about A$4.3 billion at the model’s exchange rate, the A$34.9 billion of the programme that exhibit 10 leaves without funding on the public record would fall to about A$30.6 billion1. The size, price and investors of any round remain a data gap.
On 6 October the cloud provider Lambda was reported to have put back its planned listing to 202775. On 13 October AirTrunk was reported to have postponed the planned Singapore listing of a real estate investment trust that could raise US$1.5 billion76.
7Method and limits
The model is a ten-year discounted cash flow of the operating business, built in Excel with live formulas and published with this analysis. For each price it converts the equity value to an enterprise value on two financing bases, holds the capacity, timetable and capital cost to the reported figures, and solves for the revenue per megawatt at which the discounted free cash flow, with a terminal value, equals that enterprise value.
Servers are replaced every four years. Capacity is expressed as equivalent critical IT megawatts; the bases of the reported figures are not stated, so that conversion is an assumption of the model. Revenue per megawatt is read as a price per GPU-hour at the GPU density Firmus planned for Launceston, with every hour billed4, 5. Australian dollar figures use the Reserve Bank’s rate of US$0.697924 per Australian dollar on 9 October 2026, unless a source gives its own conversion.
Firmus’s own figures from the offer documents are second-hand: no prospectus was lodged, and they rest on press reports of the pathfinder prospectus and of a broker summary18. Where reports conflict, the page shows each one with its source. Bank analysts’ estimates were used only as leads and appear nowhere in the analysis. The workbook lists every input with its source, and every output with its cell.
Data gaps
- No prospectus was lodged, so no figure from the offer documents could be read in a lodged document. Every such figure rests on press reports of the pathfinder prospectus and a broker summary.18
- Total debt, the drawn balance of the US$10bn facility, its terms, and any financing for Batam and Malaysia.5, 30
- The currency and basis of the forecast half-year loss, and whether the FY26 loss is before or after tax.18, 39
- The megawatt basis of each site, and whether all of the 912 MW is under signed contracts.9, 46
- Contract prices, terms, termination rights and credit support for each customer.5
- The periods behind the project sums, corporate costs, and the useful lives the company applies to its servers. Until the periods are known, the model’s cost calibration to those sums, which sets exhibit 7, cannot be checked against a single-year forecast.5
- The terms of the arrangements reported to take effect if the listing slipped past 30 November 2026.13
- The size, price and investors of any private round after the withdrawal.73
- The A$10m difference between the A$100m reported for Maas Group’s first investment and the A$110m in its FY26 results.25, 26
- For the comparables: net debt for CDC and for NEXTDC at 30 June 2026, Nebius’s active power at mid-2026, and a transaction price index for GB200 and GB300.37, 45, 77, 78
Corrections and updates
- Correction, 15 October 2026. The board’s reason for withdrawing the offer is now quoted in full, including the market conditions it cited, with Firmus’s stated next step.
- Update, 15 October 2026. A new subsection, Since publication, covers events to 14 October.