A new IPO on the ASX

The AFR featured some comments on Firmus:

I’d be the last to hold it against someone for making a buck or talking their book.

But there are some issues here worth expanding on, not least for what they reveal about the semiconductor space and neoclouds generally.

Semiconductors and AI-related companies are our largest sector exposures at the moment, so we’re not exactly AI bears.

Nvidia relationship

Much has been made of the Nvidia relationship whose support was critical to both their debt and pre-IPO equity raises. Nvidia is first and foremost a supplier.

It’s reported they invested $500 million in their last round. But Firmus has earmarked ~US$30 billion for server capex, separate from construction.

Nvidia’s participation reads more like a discount on their order, taken in equity, along with a substantial revenue share.

Nvidia will be a significant net receiver of cash from Firmus, not a payer in to the company.

There is a very useful divisor in AI-related investment that explains much of the performance divergence in AI-related equities: distinguishing between the cash spenders and the cash receivers.

Some companies are raising large amounts of debt and equity to lavish on datacenters, others are receiving that spend.

Our current approach is to invest in the receivers, not the spenders.

Nvidia’s support indicates a weakness that is slowly but steadily approaching in the horizon: hyperscalers are vertically integrating and building out their own custom-designed chips.

Google (TPUs), Microsoft (Maia), Meta (MTIA), and OpenAI (Jalapeño) all boast increases in efficiency, speed, but most of all cost.

Nvidia’s support of a number of neoclouds should be seen in the context of their need to diversify their customer base, and the risk that their biggest customers increasingly move onto their own designs. This also explains why hyperscalers are happy to sign short term deals with the neoclouds - it gives them the time and flexibility to shift later.

Commercially Nvidia’s moves make sense. It’s sensible for them to invest (say) a billion in equity to unlock in tens of billions of orders, or write a revenue guarantee to secure debt on someone else’s balance sheet to purchase many billions more of their products.

But their involvement should be seen in this context, not as a purely commercial partner or investor.

There is also a growing threat from new entrants that are specialised in inference, rather than the general purpose GPUs that Nvidia sells that can both train and generate tokens.

Previous technological barriers are falling fast, as state-of-the-art models are very well-suited for this. OpenAI’s chip was reportedly designed with the help of their own artificial intelligence. These will all be soon competing with neoclouds locked into Nvidia’s ecosystem.

There are plenty of semiconductors that benefit from or are at least neutral to this industry shift, such as those that help with the design and manufacture of custom chips.

And most semiconductor companies are upstream of design, so should win regardless. Which is also part of our semiconductor thesis: focus on companies that will win regardless of market share plays out in chip design.

Firmus is effectively short innovation, and will have a very nervous few years as faster, more efficient, and more cost effective chips are announced and rolled out, while their debt remains outstanding on the current generation being purchased.

5 year old GPUs are holding their value, so shouldn’t depreciation schedules be longer?

This misreads the situation, as it assumes the last five years will be like the next.

Five years ago AI was a comparatively niche industry. There was a fraction of the capital available today to invest in hardware innovation.

Today, the wave of investment across hyperscalers and new VC-backed entrants is historic, and the fruits of this will land in the next 1-3 years.

Some new designs are in the early stages of commercial roll-out, so have zero impact on today’s GPU rental or purchase market pricing that investors are extending into the future. But before Firmus’s debt is repaid on the chips they are buying, they almost certainly will.

From first principles, an inference customised chip, designed specifically for the new workflows (token generation) should structurally outperform the general purpose GPUs being funded here.

And it’s likely that within a few years entire models are burned onto chips, which would make them more efficient and faster again. Taalas is one example, and their roll-out will be well-supported now it’s announced they will be acquired by AMD.

A recent historic analogy can be found in the same way custom chips completely took over bitcoin mining. The shift was almost total.

A real estate company can genuinely rely on refinancing the debt on an asset. The asset is more or less the same. Refinancing old GPUs when new designs are available is an entirely different prospect, and that’s one of the main issues here.

Understand the customer

So why would hyperscalers sign up with neoclouds? They are, after all, making their own substantial capex investments.

There are two obvious reasons, economic and strategic.

Economically, neoclouds run at a loss, so economic value is accruing to their customers.

And leasing keeps datacenter debt and capex off their own balance sheets.

Strategically, having access to the leasing market means firstly they can flex their own supply according to demand.

But most importantly they can rent capacity today and be free to invest in newer generations as they roll out over the coming years, which includes the freedom to shift more compute onto their own designs as those programs progress. Nvidia’s 75% gross margin ensures there are plenty of economics on the table.

Valuation

There is a reported claim that Firmus is cheaper than peers. But I don’t see how you can seriously make that case.

Nebius is doing a run-rate of $2 billion with $600 million of GAAP losses. Market cap US$66 billion.

Coreweave is over $10 billion with over $2.5 billion of GAAP losses. This excludes capex. Market cap $49 billion.

Firmus did A$50m revenue last year and is being valued at US$30 billion.

Using 12 month numbers to 30 June 2026, which are dated but at least publicly available:

Note Firmus has not yet listed, information is still scarce, including the details around their contracts. These companies report differently and have slightly different business models, but are broadly in the business of leasing GPU capacity. I’ll post any corrections to the website.

It may seem unfair to use FY2026 numbers, but all these companies had the same opportunity set over the last few years, all have significant revenue ‘backlog’ and expansion plans, and all compete in the leasing market against each other.

And on a run-rate basis, Firmus has one of the worst margins of the neoclouds we look at.

The valuation is also being justified in reference to a proposed 2028 EBIT multiple.

But privately, investment bankers would be the first to explain why you shouldn’t use multiple analysis in highly leveraged project financing.

And even if taken at face value, this is an estimate of EBIT ‘as complete’. Taking a snapshot of EBIT in the future ignores the fact these are depreciating assets, interest will ensure the total repayments will significantly exceed the cost of the assets, and most importantly: the entire capital base will need to be regularly replenished simply to maintain a steady state.

There is also a very high initial debt burden which must be repaid. Valuation on a snapshot EBIT multiple conveniently leaves this out.

There are recent real estate bankruptcies where ‘as complete’ valuations were used to secure finance, and you can read the AFR to see how that turned out.

But at least real estate kind of stays there once it’s built.

In this case, assets must be replenished

EBIT can’t be used to pay debt, it must come from cash flow after all expenses.

Even at $10 billion revenue, Coreweave is losing $2.5 billion, so there are no scale effects in the industry clear as yet. So Firmus may have to raise and capex this money all over again just to stand still.

There’s a greater risk here. Usually I’d say a tech business should IPO sooner rather than later, as aside from other advantages, being listed gives you a few extra rolls of the dice when things go wrong.

In this case, if the value of chips falls below the debt, which is possible given all this innovation and Nvidia’s annual release cadence, new equity would go to repaying lenders rather than purchasing new assets, which is the kind of scenario where equity raises fail.

[As an aside Coreweave is trading roughly breakeven on an EBIT basis… which is how these neoclouds like to present themselves, but they clocked $660m of interest last quarter, which is where the run-rate losses come from. But as an equity investor how can you strip out that cost?]

This is different to say the best software companies, where each cohort generates new and growing revenue every year, so the lifetime return on investment spend can be very high. GPUs depreciate, and a new generation will have to be financed, bought and installed in the mid term.

Emerging markets

And much of the basis of this ‘EBIT’ is in emerging markets, which have their own challenges.

Many years ago I was living in London and there was an amusing episode that has some parallels.

Nathaniel Rothschild, to much fanfare, announced the opportunity of a lifetime, and raised over a billion US dollars in a widely publicised, widely discussed IPO. Needless to say, money flowed in and little came back.

Amusingly, he claimed to have smoking gun evidence of misbehaviour by his emerging market partners, but the company alleged this was obtained by email hacking, which is of course illegal.

This was promptly denied, and it all became very messy, but it was fascinating to watch it play out in real time from excitement, to capital rise, to IPO (technically more of a SPAC), to embarrassing failure. A later investigation found at least some of the allegations were true, but the egg was well and truly scrambled by then.

Rothschild and his management team did some of the right things, investing £100 million of their own money with UK levels of governance, but even with that, and all the constant reporting and public scrutiny, most of the money never came back.

It’s easy to invest in foreign countries because the state systems can work to encourage money to flow in. It is only after the money has flowed that patronage networks appear and demand their cut, as more than one ASX-listed African miner can attest. And lest we get arrogant in Australia, we certainly have our own version of patronage networks which must be properly fed.

In short, we think this should be properly valued as project equity, buying highly depreciating assets, in a space where rapid innovation ensures a high risk of technological obsolescence. Nvidia itself now has an annual cadence of new generations, and the current set of chip-related startups is impressive.

$7 billion of project equity is being sold at a valuation of $43.7 billion, substantially greater than peers, who are themselves losing money with major long term question marks over their viability.

So (in my opinion) there are far better ways to invest in AI, not least by focusing on:

1. The receivers, not the spenders, of all this cash,

2. The beneficiaries, not the losers, of ongoing technical innovation, and

3. Doing all this in a quantitatively risk-managed way, as one thing we can be sure of, there will be many twists and turns to come.

But now I’m talking my book too ;)

Michael