Hook
A company named Firmus has postponed its initial public offering and is now weighing a private funding round. That is the entire public record. No headline valuation. No named underwriter. No stated reason. Crypto Briefing carried it as a brief, and the market — such as it is — moved on without a second thought.
I didn't. The absence of a number is the number. When an issuer cancels a listing date, the proximate cause is almost never that it cannot list. It is that the clearing price the public market would set sits below the mark its existing investors need to defend. That spread has a name in every capital cycle, and in AI infrastructure in 2026 it has a direction: it is widening.
Consider what a postponement actually costs. Another audit cycle. Another set of legal fees. Another quarter of management time spent on roadshow choreography that never reaches the stage. And, most expensively, a visible gap where a debut should have been — a gap that every competitor, every prospective customer, and every employee holding options will notice. Boards do not absorb that cost for entertainment. They absorb it because the alternative is worse: printing a public valuation that resets the reference price for every future round, every option grant, and every comparable company in the sector. A postponement is a defensive trade. It says the issuer would rather be illiquid at a high mark than liquid at a lower one.
That is a strange thing to say about a company selling compute in the middle of the largest capital expenditure boom in the history of semiconductors. So let's take it seriously.
Context
To read Firmus, you have to read the cohort it belongs to. The last three years produced a distinct class of company — variously called neoclouds, GPU-as-a-service providers, or AI infrastructure operators — whose product is not a model but the physical capacity to train and serve one. CoreWeave listed in March 2025 and became the sector's public reference point, partly because its pricing process was instructive: the deal cleared below the levels implied by its private marks, and the stock has since traded with the volatility of a commodity producer rather than the multiple of a software company. Crusoe, Nebius, Lambda, and a dozen smaller names followed the same arc — raise aggressively, sign long-dated contracts, convert electricity into revenue, and hope the multiple holds.
There is another fact about this cohort, and it is its lineage. A meaningful share of AI infrastructure capacity was built by, or in partnership with, companies that began in Bitcoin mining. The reason is not sentiment; it is asset adjacency. A mining operator already owns the scarce inputs an AI data center needs — interconnected power, land, transformer capacity, a workforce that understands continuous uptime — and needs only to swap ASICs for accelerators. That pivot has been one of the defining trades of the cycle. It is also the reason a crypto-native outlet would carry a story about a data center listing at all. When a crypto trade publication reports on an AI infrastructure IPO, the crypto tell is usually in the capital structure, not the headline.
The lineage cuts both ways. It supplies a shortcut to hard assets, and it imports a discount. Public investors have spent two cycles learning to distrust this sector's accounting — the depreciation schedules, the related-party contracts, the fair-value marks on tokens held as treasury. A mining-to-AI conversion asks that audience to believe the same balance sheet is now a different business. Sometimes it is. The market prices the possibility that it isn't.
There is a macro layer beneath this. The last decade rewarded duration — the willingness to pay today for cash flows that arrive far in the future — because money was cheap and patience was subsidized. AI infrastructure is the purest duration trade on the board: enormous capex now, uncertain revenue later, and a terminal value that depends on assumptions nobody can yet verify. When the cost of patience rises, the first assets to be repriced are the ones with the longest fuse. That is the weather Firmus is listing into.
Core
Here is the mechanism I keep returning to, and it is the reason I think this postponement matters more as a symptom than as an event.
Public and private markets price the same asset on different clocks. A private round is a point-in-time transaction, negotiated between a small set of sophisticated buyers and a motivated seller, marked to the last round and then left alone. It has no daily quote, no redemption, no short interest, and no analyst publishing a target the morning after. Its price is a belief held in place by a lockup. A public price is a continuous auction that discounts terminal value — it asks not what the company will earn next year but what the entire stream is worth after the capex cycle turns, the accelerators depreciate, and the marginal buyer of compute has other options. Those two clocks rarely disagree loudly. When they do, the private clock runs slow.
That is the two-price problem. Private marks on AI infrastructure have been set by investors who are implicitly long the buildout; public marks are set by investors who must price its endgame. A company caught between those clocks has three options: list and take the haircut, stay private and hope the public market catches up, or raise a private bridge to buy time. Firmus appears to have chosen the second and third.
Why this is more than a corporate finance footnote: AI infrastructure has an unusually rigid capital rhythm. These businesses do not scale smoothly; they scale in tranches tied to power. You do not add ten percent more capacity — you energize another building, which means another interconnection agreement, another multi-year power purchase, another GPU allocation, another set of cooling and switchgear contracts. Each tranche is a step function in capex, and each step must be pre-funded. A financing delay is therefore not a cosmetic scheduling slip. It is a re-timing of the entire capex ladder, and in a sector where competitive position is determined by who locks scarce inputs first, re-timing is the same thing as falling behind.
This is where I want to be precise about what 'scarce input' means, because the market's shorthand — GPU count — is misleading. Based on my own modeling work on liquidity congestion during high-volume swaps, I've learned to distrust the headline metric and look for the binding constraint underneath it. In compute, the binding constraint is power: megawatts of energized, contracted, deliverable capacity. Accelerators are orderable; transformers, interconnection queues, and power purchase agreements are not. A data center's valuation increasingly tracks its secured megawatts and the cost of the electrons feeding them, not the number of chips in the rack. The unit of account in AI infrastructure is the megawatt, and everything else is a conversion factor.
That reframing changes how you read a postponement. If Firmus's competitive story rests on a differentiated power position — low-cost renewables, a favorable jurisdiction, a long-dated interconnection — then delaying a listing costs it time but not its moat. If its power position is generic, the delay costs it both, because generic capacity competes on price, price competes on cost of capital, and cost of capital is exactly what a failed listing raises. There is a feedback loop here that is easy to miss: a weaker financing position makes you a less attractive counterparty for the very long-dated contracts that would strengthen your financing position. The loop turns in both directions, and it turns fast.
Which brings me to unit economics, the part of the story that public diligence punishes hardest. The core tension in GPU-as-a-service is a maturity mismatch between assets and contracts. Accelerators depreciate on a schedule the market now debates openly — three years for some operators, five or six for others, with the difference flowing straight to reported gross margin. Customer contracts, by contrast, are typically multi-year and often structured as take-or-pay, which gives revenue high visibility and low diversity. When the depreciation curve is steep and the customer list is short, reported margin becomes a function of two assumptions rather than a fact about the business. Public investors price assumption risk. Private investors, who sit close enough to renegotiate, often do not. A high-margin neocloud and a low-margin one can be running identical hardware and different accounting.
I've watched this exact dynamic before, in a different asset class. In 2022 I argued that the failure of Terra was not primarily a flaw in its algorithmic mechanism but a toxic correlation between two balance-sheet items that were supposed to be independent. The lesson generalizes: when two variables assumed to be uncorrelated turn out to be the same variable wearing different names, the risk model is not conservative — it is fictional. In AI infrastructure, the assumed-independent variables are 'demand for compute' and 'availability of cheap capital.' They are not independent. The buildout is financed by the expectation of its own continuation.
I keep reaching for an analogy from my own work on restaking, and I think it holds better than the market admits. Restaking isn't a narrative shift in security; it's a repricing of the same collateral against multiple claims. Restaking isn't a narrative shift in security so much as a leverage story with better branding. And restaking isn't a narrative shift in security — it is a reminder that collateral reused is risk compounded. The neocloud balance sheet has quietly become that structure: the same projected cash flow pledged against power contracts, GPU leases, and equity marks at once. The collateral is not a token. It is the future itself, posted three times.
Let me be fair to the other side of the trade, because the bear case on AI infrastructure is lazy when it ignores geography. The global buildout is not monolithic, and neither is the demand. An operator working from Australia or the broader Asia-Pacific faces a different market than a US hyperscaler-facing neocloud — smaller addressable commercial demand, longer latency paths to some workloads, but also a policy tailwind the US cohort does not enjoy: sovereign AI. Governments increasingly want compute that resides inside their jurisdiction, under their legal regime, for reasons of data residency and strategic autonomy. That demand is less price-sensitive and more relationship-driven, and it is precisely the demand a regional operator can serve without competing head-on with CoreWeave. Sovereign AI is the realistic niche for second-tier infrastructure, and it is a niche defined by politics rather than performance.
There is a regulatory layer here too, and it is where my work on the Australian digital asset framework becomes relevant. Listing venues in Australia and Europe now impose disclosure and compliance regimes — think MiCA in Europe and the emerging ASX-adjacent frameworks — that raise the fixed cost of being public. I've argued before that most project-level KYC is theater, that a determined buyer can route around identity checks with a handful of wallets, and that the real effect of compliance regimes is to tax the honest participant while the sophisticated one absorbs the friction as a line item. The same logic applies to listing compliance: the costs are real, they are paid upfront, and they fall hardest on exactly the mid-sized issuers who cannot amortize them across a large float. A postponement is, among other things, a decision to defer that tax until the float can carry it.
Watch the shape of the bridge, if one appears. A private round priced flat to the last mark is a signal of confidence; a down round disguised as a structured note — liquidation preference, ratchet, or a convertible with a valuation cap — is a signal of stress wearing a friendly face. These instruments let a company keep its headline number while quietly transferring value to new money. That structure is not unique to AI infrastructure; it is the standard vocabulary of late-cycle private markets. If Firmus raises on terms, the terms are the story, and the headline valuation is marketing.
Tie the threads and they point at a single conclusion. AI infrastructure is in a capital arms race in which the winner is determined by cost of capital, and cost of capital is determined by which market — private or public — will underwrite the longest duration. The private market has been willing. The public market is becoming less so, because it can see the terminal-value question and is obliged to answer it. Firmus's postponement is the visible edge of that divergence. It is not a verdict on Firmus. It is a reading on the spread.
Contrarian
The consensus reading of a postponed IPO is weakness, and I think that reading is backwards here. A company that defers a listing to protect a private mark is not signaling distress; it is signaling that private capital still prices its future above what public capital will pay today. That is a statement about two markets, not about one company. If the business were deteriorating, the rational move would be to list quickly, take whatever clearing price exists, and hand the terminal risk to public shareholders. Deferral is what you do when you believe the public price is wrong.
The genuinely contrarian point is subtler and less comfortable. The risk in this sector is not a demand collapse — hyperscaler capex guidance remains the strongest single variable in the entire compute chain, and it has not broken. The risk is a financing-structure unwind: a cohort of mid-sized operators who financed long-duration assets with short-duration optimism, and who will discover, one postponement at a time, that the private clock cannot run slow forever. The public market is not being unfair to AI infrastructure. It is being early. The blind spot is the assumption that private and public valuations must eventually converge upward rather than downward.
There is also a disclosure asymmetry worth naming. Private investors in these rounds see the contract book, the customer concentration, the power costs, the depreciation policy — and they price it. Public investors see a prospectus and a roadshow. The public discount is partly a price for that opacity, and no narrative substitutes for the data private holders already have. When a company chooses the private market, you are watching someone choose the room where the information is.
Takeaway
So watch the second postponement. One is a scheduling decision. Two is a pattern. Three is a regime — and the third is where the AI capex story stops being a story about chips and becomes a story about who is willing to hold duration. Firmus is not the signal. Firmus is the first tick.