Is the Firmus IPO worth the hype?
The Firmus IPO has Nvidia, OpenAI and a hyperscaler on board, but at the reported $50bn, ASX investors would pay roughly three times the valuation of the last private round.
Marcus Today straight talk
Australia’s SpaceX. This is the float everyone will want a piece of, and that is exactly why we would be careful. Firmus has the partners any AI start-up would want – Nvidia, OpenAI and a hyperscaler reported as Meta – but only two of its seven sites are running and it has never published a set of accounts. At about $50bn, IPO buyers would pay roughly three times the valuation Coatue, Blackstone and Jane Street paid in August, for a price that already assumes everything is built on time. We think fair value is $32–45bn, and we would rather read the contracted revenue in the prospectus on 8 October before deciding whether the float is worth chasing. At first impressions, our fair value is 10–36% below the float price.
In brief
Firmus plans to raise $7bn (US$5bn) of new money on the ASX, the second-largest Australian IPO after Telstra. It builds and runs Nvidia-based ‘AI factories’ in Australia and South-East Asia, with more than 900MW contracted and OpenAI and a global hyperscaler (reported as Meta) as anchor customers. Only two of seven sites are operating and no accounts have been published. The roadshow reportedly pitches about 12x a US$5bn FY28 EBIT (earnings before interest and tax) target, which would imply about $85bn; at $50bn the multiple is about 7x.
We estimate fair value at $32–45bn, with a central estimate of about $38bn. The range is 10–36% below the roughly $50bn being reported from the roadshow. In our opinion a $50bn price already assumes Firmus delivers everything on time. We would look for the prospectus to show contracted revenue that supports at least $40bn before treating the float as fair value.
Key reasons
| 1 |
Blue-chip customers and partners. OpenAI anchors two Malaysian sites, a hyperscaler (reported as Meta) has contracted about 55,000 GB300 GPUs in Melbourne and Tasmania, and Nvidia is supplier, shareholder and credit backer. |
| 2 |
Nvidia’s floor under Batam. The 360MW Batam campus carries an eight-year revenue-share and credit-support arrangement, reportedly with an underwritten floor price. Few peers have that. |
| 3 |
The price is well ahead of the last round. Sophisticated investors bought in at a US$10.5bn (~$15bn) valuation in August. A $50bn float would hand them a paper gain of more than 200% in about three months. |
| 4 |
The EBIT target needs a lot to go right. US$5bn of FY28 EBIT implies a 28–43% margin. CoreWeave’s adjusted operating margin was about 5% in Q2 2026. |
| 5 |
It is a contract book, not yet an earnings stream. Five of seven sites are still to be built within 24 months, and the prospectus will be the first set of accounts. |
What would change the view
A prospectus showing contracted, take-or-pay revenue that supports at least $40bn, or a float price struck well below $50bn, would make the offer more attractive. Build delays, a smaller contract book than reported or heavy selling by pre-IPO holders would make it less so.
Independent research for Australian investors
marcustoday.com.au – General information only, not financial advice
Firmus is a seven-year-old Sydney-based builder and operator of ‘AI factories’ – liquid-cooled data centres packed with Nvidia GPUs that it rents to AI labs and hyperscalers. It started in Launceston, Tasmania, running bitcoin mining on cheap hydro power with its own immersion-cooling kit, then pivoted to AI compute. Its GPU cloud arm, Firmus AI Cloud (formerly Sustainable Metal Cloud), runs from Singapore on the company’s own HyperCube racks, made in New South Wales. Its private valuation has gone from $1.85bn in September 2025 to more than US$10.5bn (about $15bn) in August 2026.
| Co-CEOs | Oliver Curtis, Tim Rosenfield | Founded | 2019 (Firmus Grid) |
| Co-founder | Jonathan Levee | HQ | Sydney, NSW |
| New directors | Hatton, Bartlett, Shuttleworth | Listing | 22/10/2026 (ASX) |
| General counsel | Lachlan Pfeiffer | Sites | 7 (2 operating) |
Key shareholders (disclosed or reported): Nvidia, Coatue, Blackstone, Jane Street, Maas Group (ASX: MGH), Regal Funds Management, Ellerston Capital, Archibald Capital, Tectonic, Alex Waislitz and the Pratt family, plus the founders. Maas Group is the only listed Australian way to own a piece of Firmus before the float: it holds a stake worth $400m at cost and an $855m contract to supply Power Cube electrical units.
The reported $50bn sits above our fair value range
Equity value in A$bn. Only the bull case gets above the reported float value.
|
August 2026 round~$15bn
|
|
Bear case~$27bn
|
|
Our fair value range$32–45bn
|
|
Our central estimate~$38bn
|
|
Probability-weighted~$42bn
|
|
Reported IPO value~$50bn
|
|
Bull case~$65bn
|
Source: Marcus Today estimates; press reports of the roadshow. As at 24/09/2026. $1 = US$0.71.
The valuation has risen about 27-fold in 13 months
Post-money valuation in A$bn at each raise. The IPO figure is reported, not confirmed.
|
Sep 2025 equity$1.85bn
|
|
Nov 2025 equity~$6bn
|
|
Feb 2026 Maas stake~$6bn
|
|
Apr 2026 Coatue round~$8bn
|
|
Aug 2026 strategic round~$15bn
|
|
Oct 2026 IPO (reported)~$50bn
|
Source: Firmus releases and press reports. US$ rounds converted at the rates reported at the time.
Sophisticated money (Coatue, Blackstone, Jane Street, Nvidia) bought in at a US$10.5bn valuation only seven weeks before the float was launched. We think the gap between that round and a $50bn IPO is the central valuation question for IPO buyers.
| 6–7/10/2026 | Institutional bookbuildPrice discovery with institutions. Joint lead managers are Morgan Stanley, Bank of America, JPMorgan and Morgans. | |
| 08/10/2026 | Prospectus lodgedThe first published accounts: contracted revenue, take-or-pay terms, capex still to spend, net debt and any sell-down by existing holders. | |
| 12–19/10/2026 | Retail offer$7bn primary raise, up to $7.5bn with a $500m greenshoe. | |
| 22/10/2026 | ASX listingAt this size, likely fast entry to the ASX 200 and index buying. | |
| Late 2026 | Launceston (Southgate) due84MW IT load and about 36,800 GB300 GPUs for the hyperscaler customer. | |
| Q1 2027 | Batam first capacityThe largest single project: 360MW and US$25–30bn of expected offtake over six years. | |
| Oct 2027 | First escrow release10% of founder shares released after one year; a further 39.9% after two years. |
What we want from the prospectus: contracted revenue and its timing, the take-or-pay terms, remaining capex for the five unbuilt sites, net debt after the float, and whether any existing holders are selling.
Firmus has more than 900MW of contracted capacity, but only two of seven sites are running. In our view it is a contract book, not yet an earnings stream.
How it makes money
Firmus builds the data hall, buys the Nvidia systems (GB300 now, Vera Rubin NVL72 next) and sells GPU capacity on multi-year contracts. This is the ‘neocloud’ model used by CoreWeave and Nebius. Its pitch is lower cost per token: renewable power, immersion and liquid cooling, and factory-built modules that go up faster.
Customers
OpenAI – anchor customer for two new Malaysian AI factories (September 2026). Global hyperscaler (reported as Meta) – a multi-year, multi-billion-dollar contract for about 18,400 GB300 GPUs in Melbourne and about 36,800 in Tasmania (March 2026); media reports name it as Meta. AI-native tenants – Fireworks AI and the multi-tenant Batam campus.
Project Southgate, run with CDC Data Centres and Nvidia, targets 1.6GW across Australia by 2028. Firmus has put the full national program at $73.3bn of capital spend. Supporting deals include a trans-Pacific subsea cable with SUBCO and Maas Group’s $855m electrical contract. Source: company releases and press reports.
EVs from company results and market data. CY27 revenue consensus: CoreWeave US$26.6bn, Nebius US$12.1bn, IREN US$2.81bn (FY27, June year-end). *NextDC price, market cap ($8.66bn) and consensus target as at 24/09/2026; NextDC multiples on FY27. Firmus uses equity value because net debt after the float isn’t known, at $1 = US$0.71 (the term sheet’s $7bn = US$5bn). Backlog definitions differ between companies. A dash means the figure is not available.
At about US$39m per contracted MW, Firmus would be priced above CoreWeave and IREN (about US$25m) and below only Nebius (about US$64m) – for a company with two of seven sites running and no published revenue.
The EBIT pitch doesn’t reconcile easily. The roadshow reportedly frames Firmus at about 12x a US$5bn FY28 EBIT target, against a reported 17x for CoreWeave. At the reported $50bn (US$35.5bn) the multiple is about 7x, but only if that target is met. We checked that target from the bottom up.
| 1 |
Revenue per MW. Batam’s US$25–30bn over six years on 360MW works out to about US$12.7m per MW a year. Nebius cites US$20–25m. |
| 2 |
Run-rate revenue. 900MW × US$13–20m gives US$12–18bn a year, but only once every site is built and full. |
| 3 |
Margin needed. US$5bn of EBIT on that revenue implies a 28–43% EBIT margin. CoreWeave’s Q2 2026 adjusted operating margin was about 5% (US$128m on about US$2.6bn of revenue), because GPU depreciation absorbs most of its 59% EBITDA margin. |
Consensus on the listed peers supports that caution: all three are forecast to lose money next year even as revenue doubles at CoreWeave and roughly quadruples at the other two. In our view the FY28 target assumes all five unbuilt sites arrive on time and at margins well above the listed leader.
Post-money equity value, at $1 = US$0.71.
As an example, if the offer were priced at $5.00 a share, our base-case value would be about $3.80 and the probability-weighted value about $4.20. We will convert these into a share price once the prospectus sets the offer price and shares on issue.
What could hold the price up after listing
A free float of about 14% ($7bn of $50bn), founder escrow, likely fast entry to the ASX 200 and strong Australian demand for AI exposure. CoreWeave priced below its range in March 2025 at US$40 and now trades at US$86.90. Early trading could run well ahead of fundamentals, in either direction.
|
Build and delivery. Five of seven sites must be delivered within 24 months. Power connections, construction and GPU supply all have to line up, and any slippage pushes out revenue while interest and depreciation keep running. |
|
|
Customer concentration. OpenAI and one hyperscaler probably dominate the contract book. OpenAI’s spending commitments far exceed its revenue, and it doesn’t publish full accounts. |
|
|
Funding and leverage. Southgate alone is a $73.3bn program. A US$10bn debt facility and $7bn of IPO money cover only part of it, so more raisings are likely. |
|
|
Australian policy. Proposed reforms on data-centre energy use and grid connections have hit NextDC’s share price. Firmus’s Tasmanian, Victorian and South Australian sites need the same grid access. |
|
|
GPU obsolescence. Chips are depreciated over roughly six years, but Nvidia launches a new generation every year. Re-contracting prices for older GPUs are unproven. |
|
|
AI capex cycle. Sentiment toward AI infrastructure has cooled at times in 2026. CoreWeave trades about 44% below its peak, and credit markets price meaningful default risk for some neoclouds. |
|
|
Related parties. Nvidia is supplier, shareholder, credit-support provider and revenue-share partner; Maas Group is shareholder and contractor. The prospectus should spell out these terms. |
|
|
Governance and float structure. Co-CEOs, founder escrow that releases in stages from year one, and possible later selling by pre-IPO holders sitting on large paper gains. |
|
|
Disclosure gap. There are no audited financials in the public domain yet. Every number in this report rests on company releases or press reports. |
General information only. This content does not take into account your objectives, financial situation or needs. Consider whether it is appropriate for you and read the relevant PDS or prospectus before acting. Forecasts are possibilities, not guarantees.
Marcus Weekly - Delivered Mondays