The Firmus Pause: How AI Infrastructure Learned to Love the Private Market Again

I want to start with a confession and a cup of coffee that went cold in a Rotterdam café in the autumn of 2022. It was three days after the Terra collapse had vaporized a six-figure hole through my portfolio, and I was sitting with a spreadsheet of dead tickers and a notebook where I had written one underlined sentence I could not stop returning to: narratives don't die, they migrate. That line has earned its keep ever since. It explained why the modular blockchain thesis survived the stablecoin massacre. It explained why NFT floor prices decoupled from utility and re-coupled to status. And this week, rereading a three-paragraph flash item about an Australian AI infrastructure firm called Firmus postponing its IPO and quietly testing the water on a private round, that same sentence surfaced again like a tide I thought I had already mapped. The interesting thing is not that a company delayed a listing — companies do that every quarter, and the ones that never make the news are the ones that quietly vanish. The interesting thing is which companies are now doing it, and why a retreat is being framed as prudence when the mechanics underneath look a lot more like a calculated arbitrage between two versions of the same story.

Let me be honest about the epistemic ground I am standing on. Firmus is not a name that appears in my earlier research files, and the flash item that prompted this piece is remarkably thin — a postponement, a pivot toward private capital, and almost nothing else. No valuation. No raise size. No named cause. So treat everything that follows as a framework rather than a verdict, a set of questions I would put to the company if I had ten minutes and a term sheet. That is the discipline I learned the hard way in 2017, when I burned €150,000 chasing community coins on the strength of a story I had confused for a thesis.

To understand Firmus, you first have to understand the strange new species it belongs to. The AI infrastructure layer — sometimes called "neoclouds," sometimes "GPU-as-a-Service," occasionally dressed up as "AI hyperscaler" by marketing teams with a straight face — has become the most capital-hungry corner of the technology economy in under three years. These companies do not train models. They do not fine-tune weights. They buy the physical substrate that everyone else's intelligence runs on: racks of NVIDIA accelerators, long-term power contracts, land near substations, and the cooling and networking to keep the whole thing from cooking itself. The business is, in essence, the business of being the shovel in a gold rush that may or may not still be accelerating.

The reference point everyone uses is CoreWeave, which listed in March 2025 after a pricing process that visibly strained against its own private-market expectations. Its debut was less a coronation than an audit — public investors looked at the capital intensity, the customer concentration, the depreciation assumptions baked into the GPU fleet, and demanded a discount that private backers had never asked for. Since then the stock has swung hard enough to remind anyone paying attention that the public market applies a different discount rate to the same assets. That single data point is the lens through which the Firmus news becomes legible.

There is a second, subtler current running beneath all of this, and it is the reason a crypto-focused outlet like Crypto Briefing is the one carrying the story. Between 2024 and 2025, an entire cohort of Bitcoin miners pivoted — sometimes gracefully, sometimes desperately — into AI data centers. The logic was seductive: they already owned power contracts, industrial land, and cooling infrastructure, and their existing business was being slowly strangled by halving economics and volatile hash prices. Converting a mining shed into a GPU colocation facility looked like a natural upgrade. Whether it actually is one depends on assets, team, and depreciation baggage that the market has learned to price with skepticism.

So when I read that Firmus postponed an IPO and is considering private funding, I do not read it as a company in crisis. I read it as a company making a very specific bet about which market will believe its story at a higher price. The postponement is not a cancellation, and that distinction matters enormously. A cancellation usually signals something broke — a restatement, a failed audit, a customer walking. A postponement signals a timing decision. And when a postponement is paired with an active hunt for private capital, it signals something sharper still: the company believes its private valuation is higher than what the public market would currently pay, and it would rather raise from believers than be repriced by skeptics.

That is the core mechanism, and it deserves to be stated plainly. Private markets and public markets do not merely value the same asset differently; they value it on different clocks. Private capital can tolerate a J-curve — years of heavy capital expenditure, negative free cash flow, and delayed profitability — because its investors are locked in, illiquid, and often emotionally or strategically committed to the thesis. Public capital cannot. The moment a company lists, it is subjected to a quarterly rhythm that punishes the exact profile AI infrastructure companies naturally have: front-loaded capex, back-loaded revenue, and enormous uncertainty about whether the demand curve holds. The gap between the private fiction and the public truth is not a scandal. It is a structural feature, and Firmus is simply choosing to live on the comfortable side of it.

I have watched this movie before, in a different costume. The long arc from 2017's ICO euphoria to the structured liquidity of today is, at its heart, the story of an asset class repeatedly discovering that the audience determines the price. Tokens could raise at absurd valuations in private presales precisely because no continuous market existed to argue with them. The moment a liquid market appeared, the same projects were repriced by people who could exit. AI infrastructure is now living through its own version of that discovery, and the postponement of a single IPO is a small but clean tell.

The deeper question is why this gap exists at all, and here the technical detail matters more than the narrative. AI infrastructure companies carry a set of characteristics that public equity investors are structurally allergic to. The first is capital intensity that borders on the pathological: a single cluster of high-end accelerators can cost more than the entire annual revenue of a mid-sized software company, and the depreciation schedule — typically three to six years — sits awkwardly against contract terms that rarely extend that far with certainty. The second is customer concentration. The neocloud model is built on long-term, take-or-pay style contracts with a handful of hyperscale customers, and while those contracts produce beautiful revenue visibility, they also mean that losing one relationship can gut the top line. Public markets price that concentration as risk; private markets price it as proof of product-market fit.

The third characteristic is the one I find most underrated and most analytically interesting: power, not silicon, is the binding constraint. For years the conversation fixated on GPU supply — who could get allocation, who had the NVIDIA relationship. That framing is now obsolete. The scarce input in 2025 and 2026 is electricity, specifically the megawatts a data center can secure under long-term contract near a grid that can actually deliver them. A company's valuation increasingly rests on how many megawatts it has locked, at what cost, and how quickly it can energize them. This is why I always ask the same question first when I look at a neocloud: show me your power, not your GPUs. A fleet of accelerators sitting in a facility that cannot get grid interconnection is a depreciating liability, not an asset.

Firmus, if the crypto-origin inference holds, may actually have an edge here — or a trap. Miners spent years perfecting the art of securing cheap, sometimes curtailable, power in jurisdictions that welcomed them. That is genuinely valuable in the AI era. But it comes attached to baggage: asset quality that may not meet the redundancy and latency standards enterprise AI customers demand, a team whose instincts were forged in hash-rate optimization rather than SLA management, and a depreciation overhang from equipment that the market knows how to discount. The market has learned to spot a miner in AI clothing, and it applies a haircut accordingly. Whether that haircut is fair is an open question; whether it exists is not.

There is also a geographic variable that the flash item almost certainly underweights. If Firmus is anchored in Australia or the broader Asia-Pacific, its valuation anchor is not the same as a comparable US neocloud. Customer acquisition radius, latency to major AI demand centers, energy cost structures, and the availability of sovereign AI mandates all differ. This is not necessarily a disadvantage. In fact, it may be the single most important strategic fact about the company. The rise of "sovereign AI" — governments and large national enterprises wanting compute that stays within their borders, under their jurisdictions, insulated from geopolitical supply shocks — has created a genuine niche that global players cannot easily serve. A regional infrastructure provider with local power, local land, and local political relationships can build a defensible position that CoreWeave cannot simply bulldoze with a bigger balance sheet.

But sovereignty is a double-edged sword. It tends to come with smaller contract sizes, slower procurement cycles, and customers who are price-sensitive and politically constrained. The revenue is real, but it rarely scales at the rate that public-market growth investors demand. So a sovereign-AI-focused regional provider faces a structural mismatch with public equity expectations: solid, defensible, but not explosive. That mismatch, more than any operational failure, could be the honest reason an IPO got postponed.

Now let me turn to the unit economics, because this is where the story gets uncomfortable and where I think most commentary stops too early. The neocloud business looks deceptively simple: buy GPUs, rent them out, collect the spread. The spread, however, is squeezed from both ends. On the cost side, the accelerator itself depreciates fast and its resale value is hostage to how quickly the next generation arrives — a company holding H100 inventory in 2026 is not holding the same asset it was in 2024. On the revenue side, pricing power is nearly nonexistent for anyone without a genuinely scarce input. Most neoclouds are middlemen in a chain where the upstream (NVIDIA, power providers) and the downstream (hyperscale customers) both have more leverage than they do. In that position, the only durable moat is a resource nobody else can replicate — long-term cheap power in a location that matters — and everything else is a commodity dressed as a platform.

This is precisely the kind of nuance a flash news item cannot carry, and it is why I distrust the framing that Firmus's move signals "AI cooling down." A single company postponing a listing tells you almost nothing about aggregate AI capital expenditure, which is still dominated by the hyperscalers' own guidance — and those numbers have remained aggressive. Microsoft, Google, Amazon, and Meta continue to signal massive infrastructure spending, and their capex guidance is the only variable that genuinely matters for the demand side of this equation. One neocloud's IPO timing is noise against that signal. Confusing the two is the classic error of mistaking a sample for a trend.

And yet I would be dishonest if I pretended the postponement carries no information at all. It carries one piece, and it is worth extracting carefully. If AI infrastructure companies increasingly prefer private capital to public listings, that tells us the public window for this specific asset class may be narrowing — not because AI is dying, but because public investors have grown disciplined about capital intensity at exactly the moment private investors remain enchanted by it. That divergence is the real story, and it is a divergence that historically resolves in one of two ways: either private valuations eventually correct downward to meet reality, or a wave of consolidation absorbs the mispriced middle.

Here is where I want to push against the comfortable narrative, because the comfortable narrative is that postponing an IPO is a sign of weakness. I think that reading is lazy, and it misses the arbitrage. Consider what a company actually does when it postpones a listing while raising privately. It keeps its story in the hands of investors who signed up for the story. It avoids the moment when a liquid market gets to vote on whether the story survives contact with a discount rate. It buys time — time for revenue to catch up to valuation, time for contracts to mature, time for power to energize, time for the next hyperscaler deal to land. In a market where the cost of being publicly wrong is permanent and the cost of being privately patient is merely expensive, postponing is not retreat. It is optionality.

The contrarian angle goes further, and it is the part I find genuinely counterintuitive. Everyone assumes the private market is the easier audience because it is more credulous. I would argue the opposite is now true for AI infrastructure. The private investors funding these companies are increasingly the same crypto-native and crossover funds who watched 2022 dismantle the last generation of over-levered, narrative-driven infrastructure bets. They are not naive. They are simply locked in, and lock-in changes behavior: an investor who cannot exit must convince themselves the thesis is right, which produces a kind of self-reinforcing conviction that a public market — full of daily exits and second-guessing — can never sustain. The private market is not more rational about AI infrastructure. It is more committed. And commitment, in the short run, looks identical to being right.

That is why I think the most probable unstated truth behind the Firmus news is a valuation gap, not a business failure. The company or its backers believe the private mark is defensible and the public mark is not yet there. So they wait, raise a bridge if needed, and keep the fiction intact a little longer. This is rational for them. It is also, quietly, a warning for anyone who thinks private AI infrastructure marks represent realized value. They represent a negotiated belief, and beliefs can be renegotiated.

What I would want to see before forming a real view is a short list of things the flash item could not provide. The pre-postponement valuation range, and whether any new private round would be flat, up, or down against the last one — the direction of that number tells you everything about which way the wind is actually blowing. The cash runway, because AI infrastructure burns capital at a rate that turns a comfortable cushion into an emergency in two quarters, and a postponement plus a private raise sometimes masks a liquidity squeeze dressed as strategy. The revenue mix and customer concentration, because a single anchor tenant can be either a lifeline or a liability depending on how the contract is structured. And the power position — megawatts secured, cost per megawatt-hour, and interconnection timeline — because in this business, that is the only number that reliably predicts who is still standing in three years.

I would also watch the comparable set with more attention than the headline. If other neoclouds begin postponing or quietly shelving listings in the coming quarters, Firmus stops being an anecdote and becomes a data point in a genuine repricing. If they do not, and listings proceed at healthy multiples, then Firmus is idiosyncratic — a regional player with regional constraints, doing exactly what a rational regional player should do. The distinction between those two worlds is the difference between a framework and a forecast, and I refuse to collapse them prematurely.

There is a version of the next eighteen months where this all looks obvious in hindsight. AI infrastructure's private valuations were always going to meet the public market's discipline eventually, and the companies that survived would be the ones with real power, real contracts, and real customers rather than the ones with the loudest narratives. The postponement of a single IPO is not that reckoning. But it is the kind of small, unglamorous signal that tends to precede one — the first crack in a facade that everyone insists is load-bearing.

The lesson I keep relearning, from the Rotterdam café to this week's flash item, is that the price of a story depends entirely on who is allowed to argue with it. Give a narrative a liquid market and it must survive daily cross-examination. Keep it private and it survives on conviction alone. Firmus has chosen conviction for now, and there is nothing irrational about that — as long as you remember that conviction is a loan against the future, not a payment from it.

So here is the question I am left holding, and the one I would put to anyone building a position in this space: when the private and public versions of the same AI infrastructure story finally have to reconcile, which one do you believe will turn out to have been the fiction? Because I have a strong suspicion it will not be the one wearing the discount.