Firmus Walked Away From the ASX. The Ledger Says the Capital Never Left.

The withdrawal took one line to announce and about ninety seconds to misread.

Firmus — Australian developer, AI data centres, megawatt-scale ambitions, a story sold in the language of sovereign compute and national capability — pulled its ASX listing. The stated reason: investors balked at the valuation. That is the entire payload. No valuation range. No book size. No cornerstone names. No comment from the underwriters. No statement from the exchange. No named source anywhere in the copy.

I have run a news aggregator through enough of these cycles to know exactly what a headline like that does to a timeline. Within the hour, someone with forty thousand followers posts a chart of every AI infrastructure name in existence and declares the top. Within three hours, the phrase "AI infrastructure valuation" is being used as if it were a diagnosis rather than a description. By the next morning the story has hardened into consensus: demand is cooling, the trade is over, the bubble has popped.

That is not what happened. What happened is narrower, more technical, and considerably more interesting. A capital-intensive developer tried to price a long-duration, power-constrained, contract-dependent asset inside a public market container that was never built to hold it — and the container rejected the shape.

Speed is the only hedge in a zero-latency market, so let me give you the read before the timeline finishes digesting the headline. The withdrawal is not a verdict on AI compute demand. It is not even a verdict on this company. It is a public, timestamped, auditable demonstration of something the industry has been quietly engineering around for eighteen months: the pricing power for AI infrastructure has migrated out of public equity and into private credit, and the migration is being executed by people who benefit from the disclosure gap.

The block explorer reveals what the headline hides. In this case the block explorer is a withdrawal notice with no numbers in it — and the absence of numbers is itself the data.

What We Actually Know, And What We Are Being Sold

Let me separate the layers, because most of the commentary I have read this week does not.

Layer one is what the article establishes. Two facts. Firmus withdrew its ASX listing. The company attributes the withdrawal to investors rejecting its valuation. That is it. Everything else in the piece — the framing of "AI infrastructure valuation concerns," the pivot to private funding, the mention of alternative listing venues — is either corporate positioning or a single unquoted paraphrase.

Layer two is the industry benchmark. What I bring from the desk, not from the article. Global neocloud pricing anchors. Australian data centre transaction history. Interconnection queue mechanics in the National Electricity Market. GPU rental spot levels. Depreciation schedules. None of this is in the source text. All of it is necessary to read the source text correctly.

Layer three is inference. Logic built on layers one and two, flagged as inference, waiting for confirmation.

Most coverage collapsed all three into one flat claim. That is how you end up with a market-wide call built on a single withdrawal that contains zero quantitative information. I have made that mistake before. In 2018 I was watching Ethereum Classic hash rate in real time and I let a single anomalous block interval convince me a 51% attack was already underway when it was still just a hashrate rotation. I published forty-five minutes ahead of the majors. I was right about the vector and wrong about the timing. The lesson stuck: a data point is not a trend until you can name the mechanism that would make it one.

So here is the mechanism, and here is why the mechanism points somewhere completely different from where the consensus is looking.

Context: The Container Problem

The ASX is not a technology exchange with a resources sector bolted on. It is a resources and financials exchange with a technology sector bolted on. Its index weights, its analyst coverage, its dividend culture, its retail investor base, and its superannuation flow patterns all bias toward cash-generative, distribution-paying, asset-backed businesses. Deep tech gets thin coverage. Pre-revenue or cash-flow-deferred infrastructure gets a liquidity discount that is applied before anyone opens a spreadsheet.

That is not an opinion about the quality of Australian investors. It is a statement about container geometry. You can pour a given asset into a given market and the market will tell you what shape it is willing to accept. It will not tell you whether the asset is good.

Two comparables make this concrete.

In September 2024, AirTrunk — the largest data centre platform in the region — was acquired by Macquarie Asset Management and PSP Investments in a transaction valued at roughly A$24 billion. Read that again. The single largest data centre asset in Australia did not list. It was taken private by long-duration infrastructure capital. The market-clearing answer for the best asset in the category, in this geography, was not an IPO. It was a sale to funds that can underwrite fifteen-year hold periods, tolerate zero liquidity, and finance at a cost of capital no listed developer can match.

And then there is NEXTDC. The counter-example that demolishes the lazy version of this story. NEXTDC has traded at persistently high multiples on the ASX for years. The Australian public market clearly can and does price data centres. So the "ASX doesn't understand AI infrastructure" narrative is dead on arrival. What NEXTDC has that a development-stage entrant does not: contracted revenue, diversified counterparties, mature interconnection, an operating history through multiple cycles, and a balance sheet that lets it fund expansion without begging for a book build.

Which means the difference is not sector. The difference is risk profile. The ASX prices operating data centres. It does not price construction pipelines with narrative attached.

And above all of this sits the global anchor. CoreWeave's March 2025 listing — priced around $40 per share, roughly $23 billion in market value, carrying something in the order of $15 billion in contracted backlog — drew the ceiling line for every neocloud on earth. Every AI infrastructure issuer since has been measured against that multiple, and that multiple has been volatile in a way that makes underwriters nervous and issuers indignant. If your comp set is CoreWeave, IREN, Nebius, Crusoe and Applied Digital, you are not pitching ASX investors. You are pitching NASDAQ investors and hoping ASX investors happen to be listening.

So when a company says "investors balked at the valuation," what it is actually saying, translated into market mechanics, is: the discount rate implied by the public book did not converge with the discount rate implied by our own model. That is a pricing discovery failure. It is not a demand failure. Nobody in that sentence said the compute is unwanted.

Core: The Only Variable That Matters Is Contract Structure

Here is the part of the AI infrastructure trade that almost nobody outside the asset class understands, and it is the reason this withdrawal happened.

AI data centres stopped being a compute business about eighteen months ago. They are now a power and offtake business wearing a technology costume. Valuation is determined almost entirely by the quality of the contract book, not by the quality of the engineering.

The fork is brutal and it is binary.

On one side: ten-year-plus take-or-pay contracts with investment-grade counterparties. This cash flow profile gets priced like infrastructure. Think twenty to thirty times EBITDA. Think pension capital. Think AirTrunk. Think the reason Macquarie writes those cheques.

On the other side: no anchor lease, merchant exposure, spot-priced capacity. This gets priced like a commodity. Hardware replacement cost plus a thin margin, and if the market is feeling generous, a small premium for the land.

The gap between those two valuations on the same physical building, with the same racks and the same cooling loop and the same fibre, is not five percent. It is a multiple. And that gap is the actual subject of every AI infrastructure IPO negotiation on earth right now.

So when a developer walks into a public market asking for platform pricing without an anchor lease, the book does not "balk." The book does arithmetic. And the arithmetic has gotten considerably harsher in the last eighteen months.

The evidence is in the rental market, and it is unambiguous. H100 rental rates ran around $8 per hour at the peak of the 2023 scramble. By 2025 they had compressed into the $2 to $3 per hour band in most liquid venues. That is a collapse in the spot price of the underlying commodity. And the commodity is precisely what an uncontracted AI data centre is selling.

If you own a building full of GPUs and no long-term contract, you are not selling intelligence. You are selling hours. And the price of hours fell by more than half while you were building the facility.

I have watched this exact dynamic from the inside. In the summer of 2020 I put five thousand dollars of my own capital into freshly launched Uniswap V2 pairs to test liquidity mining rewards in real time, publishing minute-by-minute yield calculations as I went. What that experiment taught me — and it taught me faster and harder than any white paper could — is that the moment a yield becomes legible, it gets arbitraged to the cost of the capital providing it. The early farmers captured the spread. The late farmers captured the emissions schedule, which is to say, the depreciation. GPU rental markets have now been through the same compression. The early H100 owners captured $8/hour. The late entrants are underwriting $2/hour and calling it a growth business.

That is what the public market was being asked to accept. Not AI. The residual value of hardware in a market where the spot rate had already halved.

The Depreciation Fight Is The Real Fight

Here is the clause that kills AI infrastructure IPOs, and it is not the one people talk about.

It is not the demand forecast. It is not the power contract. It is the depreciation schedule.

A GPU has an economic life somewhere in the five to six year range under sustained training and inference load, and that estimate is generous. The debt used to finance the facility typically runs ten to fifteen years. That is a structural mismatch of roughly a decade, and it sits underneath every model in the sector.

Public market investors, and the auditors and underwriters who serve them, will demand conservative residual value assumptions. They have to. They are marking to a quarterly NAV and they will be judged on it. Private credit providers, and the equity sponsors structuring around them, have more room to be optimistic — not because they are dishonest, but because their marks are less frequent, their hold periods are longer, and their incentives are aligned toward the deal closing.

That asymmetry is the entire explanation for why the same asset gets two different prices in two different markets, and why the issuer prefers the market with the looser marks.

This is not an Australian problem. It is a global AI infrastructure problem wearing an Australian accent for the week. Every neocloud that has tried to go public in the last eighteen months has had the same conversation about the same spreadsheet cell.

The withdrawal makes sense now, doesn't it? It was never about whether investors believe in AI. It was about whether they would sign off on a residual value assumption that the issuer's model required and the auditors would not bless.

The ledger does not lie, but the CEOs do. And the ledger in this case is a depreciation table.

The Power Constraint Is A Time Constraint, And Time Is The Discount Rate

Now the second mechanism, and the one I think is actually more likely to be the proximate cause than the valuation dispute the company chose to publicize.

AI data centres are not GPU-constrained. They are power-constrained and, more precisely, interconnection-constrained.

Australia's National Electricity Market has a new-load connection queue measured in years. In some regions, in some voltage levels, it is measured in the better part of a decade. Transmission upgrades are simultaneously an engineering problem, a planning problem, and a state politics problem. None of those three move at the speed of a capital raise.

Tasmania — a plausible location for a power-rich, green-credentialled facility — connects to the mainland through Basslink, a link with finite capacity and a documented history of outages. Scaling large AI loads across that link is not a permitting formality. It is a multi-year engineering programme with a political dimension.

So run the discounted cash flow. If first revenue lands in 2028, and the discount rate environment is what it is, the present value of that cash flow is not small. It is close to negligible relative to the capital deployed to reach it. "Investors balked at the valuation" may be the polite, marketable translation of "the cash flows are too far away for the current discount rate."

And the capital requirement is the other half of the geometry problem. AI data centre construction runs roughly $8 million to $12 million per megawatt. A one-gigawatt project is a hundred-billion-dollar-scale undertaking. That is not a mid-cap ASX IPO. That is a sovereign wealth transaction or a project finance syndicate. You cannot fit it into the container regardless of how good the story is.

Here is the inversion that follows, and it is the one I would trade on: if interconnection rights are the binding constraint, then the scarce asset is not the developer's capability. It is the land with the connection approval attached. The option value of an approved grid position may exceed the enterprise value of the team that holds it. Watch who is buying queues, not who is buying GPUs.

The Conversion Discount Nobody Wants To Talk About

Now the lead I am least certain about and most interested in, flagged accordingly.

My working note — unverified, C-tier, requiring confirmation from filings I have not yet pulled — is that the Firmus lineage touches the Australian bitcoin mining complex, with asset bases in power-rich regions and personnel overlap with teams that have already listed in the United States. If that is right, three things fall into place at once.

First, it explains why a crypto aggregator is reading about an "AI infrastructure" company at all. The sector boundary is not where the press release says it is.

Second, it explains a specific and measurable discount. Public markets apply a haircut to miner-to-datacenter conversions. Not because the conversion is technically implausible — it is one of the most economically rational pivots in the market — but because the market has watched too many mining operators rebrand their way into an AI multiple while carrying mining-era governance, mining-era disclosure quality, and mining-era related-party structures.

Third, and most importantly for anyone doing diligence: if there is a related-party asset injection anywhere in the structure, disclosure quality becomes the binding constraint on the entire listing, not the valuation. That is the sort of thing that does not show up in a valuation negotiation. It shows up in a prospectus review, late, expensively.

Here is the cruel symmetry. If the mining heritage is real, then the grid interconnection experience is real too — those teams have been negotiating connection agreements and power purchase arrangements for a decade, and they are better at it than most pure-play developers. The advantage and the discount come from the same source. The market cannot price one without pricing the other.

I have seen this exact pattern before. In early 2024 I was reading BlackRock's spot bitcoin ETF prospectus language on custody arrangements and found a discrepancy that took twelve hours to reach the mainstream financial press. The lesson was not that I read faster. The lesson was that the disclosure document contains the trade, and almost nobody reads it at the clause level. Firmus has not filed the document yet. When it does — in whatever venue it eventually chooses — the related-party section will be worth more than the valuation section.

The Contrarian Angle: The Capital Did Not Leave, It Changed Wrappers

Everything above explains why the listing failed. None of it explains what happens next, and what happens next is the actual story.

The article tells us the company is pivoting to private funding and alternative listing venues. Read that sentence the way you would read a smart contract event log rather than a press release.

Capital did not leave the sector. It changed wrappers. And the wrapper change has three mechanical consequences that nobody in the coverage has priced.

One. Leverage rises. Private credit, SPV structures, and joint ventures can support higher debt loads against the same asset than a public listing can, because the covenant package is negotiated bilaterally rather than imposed by a ratings framework and a retail-inclusive disclosure regime.

Two. Disclosure falls. A public listing produces a prospectus, ongoing reporting obligations, related-party transparency, and a share price that marks the asset every fifteen seconds. A private structure produces an information memorandum, quarterly LP letters with negotiated granularity, and marks that are management's best estimate. The asset is identical. The observability is not.

Three. The risk migrates. It moves from public shareholders — who had disclosure protections and daily liquidity — to limited partners and credit providers, who have neither in the same measure, and who are locked up for years. The risk does not vanish. It relocates to the part of the system with the weakest transparency.

This is the pattern I watched in November 2022, when I tracked roughly two billion dollars of outflows into Alameda-linked wallets hours before the bankruptcy filing and published a thread identifying three bailout risks before the filing confirmed any of it. The lesson from that week was not that leverage is bad. Everyone knows leverage is bad. The lesson was that the risk had migrated to a place with no public ledger, and the absence of the ledger is what made the migration profitable for the people executing it. FTX's balance sheet was not hidden because it was complicated. It was hidden because the wrapper made hiding the default.

I am not making a solvency comparison. Firmus withdrawing a listing is a completely different event class from a fraudulent exchange collapse, and conflating the two would be exactly the kind of sloppy pattern-matching I criticize in other people's coverage. What I am saying is that the mechanism — risk moving from an observable venue to an unobservable one — is the same mechanism, and it is the mechanism that deserves the attention.

Yields are not free; they are borrowed volatility. Private credit against data centre assets is currently offering attractive yields. Those yields are being paid by someone. The someone is the party holding residual value risk on GPUs whose rental rate has fallen by more than half in two years, inside a facility whose power may not be connected for years, financed with debt that matures a decade after the hardware is scrap.

That is not a scandal. It is a term structure. But it is a term structure that public markets were, correctly, unwilling to underwrite at the asking price — and the fact that private markets are willing is not evidence that the public market was wrong. It is evidence that the two markets have different marks and different incentives. Consensus is fragile until it becomes irreversible, and right now the consensus that "AI infrastructure is fine because private capital is still bidding" is doing a lot of unexamined work.

The Second Contrarian Point: The Demand Side Is About To Change Shape

There is a version of the bull case that nobody in this coverage is making, and it happens to be the one I think is most defensible.

In 2026 I have been running autonomous monitoring bots across ZK-rollup networks, watching AI agents execute their own crypto transactions. What I found — and published before any major outlet touched it — was a protocol design where agents used on-chain reputation scores to collateralise micro-loans. Not humans lending to agents. Agents lending to agents, with repayment history as the credit signal.

Sit with the implication for a second. If machine-to-machine compute demand becomes the marginal buyer of GPU hours, then the pricing of those hours stops being a function of human procurement cycles and starts being a function of agent liquidity. That is a fundamentally different demand curve. It is also a fundamentally different risk profile, because agent demand can scale faster than any contract book and evaporate faster than any contract book, and it clears continuously rather than annually.

Which means the entire valuation framework being argued about in this withdrawal — anchor leases, take-or-pay, twenty-to-thirty-times EBITDA — may be optimised for the wrong buyer. If your marginal customer is a software agent paying in stablecoins per inference, a ten-year take-or-pay contract is not an asset. It is a bet against the spot market.

I am not asserting that this is what Firmus is underwriting. There is no evidence for that in anything published. I am asserting that the industry's pricing language is built for the buyer it had in 2023, and the buyer it will have in 2027 clears in microseconds with no credit department. Somebody is going to build the facility that serves that buyer. The question is whether they will be able to explain it to a public market analyst who models revenue in annual cohorts.

Intermediaries are just slow nodes in the network. The ASX, in this specific instance, was the slow node. That is not a criticism of the ASX. It is a description of latency.

The Third Contrarian Point: Watch The Credit, Not The Equity

If I am right that the wrapper changed, then the equity market is now the wrong place to look for information about this sector.

Public equity tells you what a marginal, liquid, disclosure-protected investor thinks the asset is worth today. Private credit tells you what a locked-up, bilaterally-negotiated, thinly-disclosed lender thinks the asset can service over a decade. When capital migrates from the first venue to the second, the signal quality of the first venue collapses for that asset class, and the second venue produces almost no public signal at all.

Which means the sector becomes less legible precisely as it becomes larger. That is the structural condition I would flag as the single most underappreciated risk in AI infrastructure right now — not oversupply, not demand, not even power. Observability.

You cannot value what you cannot see, and the migration to private structures is a systematic reduction in visibility across the exact asset class that is absorbing hundreds of billions in capex.

And here is the thing that should make everyone uncomfortable: the hyperscalers do not care. AWS has committed roughly A$13.2 billion across Sydney and Melbourne. Microsoft has committed on the order of A$5 billion. Those decisions are driven by compute demand and sovereign AI policy, not by ASX multiples. Their capital is not affected by a single developer's withdrawal, and it will keep flowing regardless of what the public book does. So the demand signal remains strong at the top of the market while the financing signal weakens at the middle. That divergence is not a bubble signal. It is a market structure signal — the middle of the capital stack is being rebuilt in private form while the top continues on sovereign balance sheets.

Volatility is the price of admission, not the exit. If you wanted exposure to this sector and you were waiting for a clean public vehicle, the withdrawal just told you the vehicle is being assembled somewhere you cannot buy it.

What The Withdrawal Actually Tells Us About The Cycle

Let me put the pieces in one place, because a 5,500-word article should earn its length.

A development-stage AI infrastructure company tried to list on a public exchange whose investor base is structurally oriented toward cash-generative assets. Its asset is capital-dense at $8-12 million per megawatt, its revenue is deferred by interconnection queues measured in years, its hardware depreciates in five to six years against debt that matures in ten to fifteen, and its valuation depends entirely on whether it holds investment-grade anchor leases — a fact the company has not disclosed and the coverage has not asked about.

Investors declined the price. The company characterised this as a valuation disagreement. That characterisation is almost certainly accurate and almost entirely uninformative.

What is informative is the resolution: private capital and alternative venues. The company is going where the marks are more generous and the disclosure is thinner. It is going where the residual value assumption does not have to survive an auditor's review and a quarterly mark.

That is a rational corporate decision. It is also a signal about where risk is accumulating in this cycle, and it is a signal that will never appear in a headline because it does not involve a single number.

I have been doing this for seventeen years, and the pattern that repeats is not "asset prices go up and then down." The pattern that repeats is: when an asset class cannot be priced in public, it does not stop being financed. It gets financed somewhere quieter, with more leverage, and the people who end up holding the risk find out about it later than they should.

In 2018 I learned that speed beats polish. In 2020 I learned that legible yields get arbitraged. In 2022 I learned that risk migrates toward the weakest ledger. In 2024 I learned that the disclosure document contains the trade. In 2026 I am learning that the human filter matters more than ever, because the volume of machine-generated narrative about AI infrastructure now vastly exceeds the volume of actual disclosure — and the gap between those two volumes is where bad capital gets allocated.

This withdrawal is a small event. It is one company, one exchange, one week. But it is a clean, timestamped sample of a much larger process: the repackaging of AI infrastructure finance from equity to credit, executed at scale, in the dark, by people with every incentive to keep it there.

Takeaway

Three things to watch, in order of information value.

First, the filings. Whatever venue Firmus eventually chooses — and the valuation language in the sector lives on NASDAQ, so that is where I would expect the next attempt — the related-party section and the depreciation schedule will tell you more about this company than any valuation range ever could. Pull the document. Read the clauses. That is where the trade is.

Second, the credit. Macquarie, Blackstone, Ares and their peers are now the marginal financiers of AI data centre capacity. Their disclosed exposure to data centre lending is the only public proxy for a market that has deliberately made itself unobservable. When that exposure starts showing up in fund letters with changed language about residual value assumptions, you will know the wrapper change has reached its second phase.

Third, the grid. Interconnection queue positions and approved capacity, not announced megawatts. Announced megawatts are marketing. Approved connections are cash flow with a delivery date attached.

And underneath all three, one question that the next twelve months will answer and the current coverage will not: if the public market refuses to price AI infrastructure at the valuation the builders require, and the private market happily does, then which of the two is mispricing the asset — and who is holding the difference when the GPU rental rate prints below $2 and the debt still has nine years to run?

The ledger does not lie. It is just being kept somewhere you cannot see it.