HOOK
At 06:14 Singapore time, a four-line item crossed the wires: AirTrunk has agreed to sell a Johor data centre campus for approximately US$1 billion, "as AI selloff bites." Four data points. No buyer. No megawatts. No lease schedule. No cap rate. And yet the number — a billion dollars — did what numbers always do in a bull market: it manufactured the sensation of information while delivering none of the substance.
I have spent twelve years reading disclosures like this one, and I have learned to treat the ratio of headline size to disclosed detail as a risk metric in itself. When a $1B asset sale arrives with fewer verifiable parameters than a memecoin's tokenomics page, the correct response is not to model the deal. The correct response is to model the gap.
Here is the gap. A data centre's value is not a dollar figure. It is a ratio: dollars per megawatt of critical IT load, adjusted for lease tenor, tenant credit, power cost, and cooling headroom. Strip those four variables out and the "$1B" is not a valuation. It is a marketing artefact — a number engineered to be repeated, not examined.
CONTEXT
AirTrunk is not a startup. It is the largest hyperscale data centre platform in Asia-Pacific by committed capacity, with campuses across Sydney, Melbourne, Tokyo, Osaka, Hong Kong, and Singapore — and a Johor footprint that exists for one structural reason: Singapore ran out of room.
In 2019, Singapore imposed a moratorium on new data centre construction, citing land, water, and power constraints. The moratorium was lifted in 2022, but with conditions — a capped allocation of roughly 300MW for new builds, strict PUE ceilings, and a preference for designs that could credibly claim a sustainability credential. The result was a spillover corridor across the Johor Strait: land that is cheaper, power that is more available, and a 2–4ms round-trip latency to Singapore's interconnection hubs that is, for most workloads, functionally indistinguishable from being in-country.
That is the asset class AirTrunk is selling into. And the ownership context matters enormously. In 2024, Blackstone and CPP Investments closed the acquisition of AirTrunk at roughly A$24 billion — about US$16 billion. That is not a distressed seller. That is a sponsor that bought a platform, marked it, and is now doing what sponsors do: recycling a slice of capital back to limited partners to prove the value-creation thesis before the next fundraise.
I have seen this movie before, and I have traded it. In 2021, I watched BAYC floor liquidity thin out in real time while whale wallets moved — not because the art got worse, but because the marginal holder changed. The BAYC crash wasn't about aesthetics. It was about who was left holding the bid. Data centres are not NFTs, but the reflexive dynamic is identical: the headline tells you what changed hands, never who was on the other side or why.
Which brings us to the second context layer, and the one this readership actually cares about: crypto miners have spent three years quietly becoming data centre operators. The compute arbitrage between Bitcoin mining and AI hosting is now the single most important capital allocation decision in the listed crypto complex. And the Johor transaction is a cleaner read on that arbitrage than anything printed in the token markets this quarter.
CORE
Let me do the arithmetic the wire item omitted.
Hyperscale data centre transactions in APAC through 2024 and 2025 have cleared in a range of roughly US$8 million to US$12 million per megawatt of critical IT load, depending on market, lease tenor, and whether the power is secured or merely queued. That range is not a secret — it is embedded in every comparable trade from Equinix, Digital Realty, and ST Telemedia's recent portfolio marks. If the Johor asset is a 100% freehold sale of an operating campus, a $1B price implies something in the region of 80MW to 125MW of critical IT load.
If it is a partial equity stake, a JV interest, or a forward sale of a development site with secured power but no steel in the ground, the implied megawatt figure moves dramatically in either direction. And here is the forensic point: a development site with a secured power interconnection can trade at a premium to an operating asset, because the buyer is purchasing the scarcest input — the electricity — not the concrete.
This is the inversion most retail readers miss. For a decade, data centre value tracked real estate: land basis, construction cost, replacement value, and the cap rate spread to industrial property. Since 2023, value tracks power. A Johor campus with 100MW of firm, contracted, renewable-credentialed supply is worth more than an equivalent campus in a Tier 1 market sitting behind a four-year interconnection queue. The queue is the asset. The building is the afterthought.
Now the cost side. A hyperscale build in Johor runs roughly US$7–9 million per MW all-in — shell, power train, cooling, fit-out, and the substation work that nobody puts in the press release. Add the land and the interconnection cost and you land near the transaction range. That means a $1B sale at 80–125MW is not a distressed mark. It is a flat-to-modest-gain exit at replacement cost, which is precisely what a capital-recycling programme looks like when it is working.
If the sale were genuinely driven by an "AI selloff," we would expect a discount to replacement. We have no evidence of one. We have no evidence of anything. That absence is the story.
Let me be precise about what an AI selloff would actually look like on the ground, because I have been tracking the capex guidance of the four North American hyperscalers — Microsoft, Amazon, Alphabet, Meta — as a real-time leading indicator since my 2025 ETF arbitrage work.
A genuine AI capex contraction shows up in a specific sequence. Hyperscaler capex guidance gets revised down for two consecutive quarters. Pre-leasing rates on new wholesale campuses fall below 60%. Power purchase agreements get renegotiated, deferred, or walked. And only then do operating assets trade at a discount to replacement cost, because operating assets are the last thing to reprice — they carry contracted cash flows that lag the sentiment cycle by four to six quarters.
None of those conditions were met when this transaction surfaced. What was true is that AI-adjacent equities had a rough quarter, that a handful of listed data centre REITs de-rated on rate expectations, and that crypto media — which now covers anything with the word "compute" in it — needed a narrative to hang a $1B number on. "As AI selloff bites" is not reporting. It is a caption. And captions are the cheapest form of causality ever invented.
The crypto convergence is where this gets commercially interesting.
Consider the listed Bitcoin miners. Through 2024 and 2025, the marginal miner's economics deteriorated as hashprice compressed and the post-halving block subsidy reset the revenue base. But the same miners hold something AI operators desperately need: energised sites, interconnection rights, and transformers already in the ground.
Transformer lead times in North America run 18 to 36 months. Interconnection queues in PJM and ERCOT run up to four years. That is the moat, and it is a moat denominated in copper and queue position, not in hashrate. Hashrate is rentable. A 230kV interconnection agreement is not.
So the trade is straightforward. A miner with 500MW of energised load can either mine Bitcoin at a variable, volatile, capital-light return, or convert a portion of that load to AI hosting at a contracted, investment-grade, 10-to-15-year return. The conversion capex is real — liquid cooling retrofits, higher-density power distribution, network upgrades, and the staffing to run a facility that cannot tolerate a five-minute outage. But the revenue multiple expansion is enormous, which is why Core Scientific, IREN, TeraWulf, and Hut 8 all re-rated on hosting contracts while their hashprice exposure stayed flat.
Yield farming isn't dead; it just moved into the power contract. The 2020 version of the trade was chasing an APY printed by a smart contract with anonymous counterparties. The 2025 version is chasing a cap rate printed by a 15-year lease with an investment-grade hyperscaler. The mechanics of yield aggregation are identical. The counterparty risk is inverted, and the capital intensity is roughly a thousand times higher.
I audited auto-compounding vaults in 2020 and calculated that manual rebalancing lagged automated strategies by about 15% on an annualised basis. The lesson I took from that was not "automation wins." It was that yield is a function of who bears the operational burden. In DeFi, the protocol bore it. In data centres, the operator bears it — and that is exactly why the asset commands a premium to replacement cost in tight markets. You are not buying megawatts. You are buying the removal of an operational headache from a hyperscaler's balance sheet, priced at whatever the hyperscaler's internal cost of that headache happens to be.
Now the cooling layer, because it is the most under-discussed variable in every data centre headline I read.
A GB200-class rack draws roughly 120 to 140 kilowatts. A traditional air-cooled hall was engineered for 8 to 15 kilowatts per rack. That is not an incremental upgrade. It is a different building. Retrofitting a legacy hall for liquid cooling runs US$1.5–2 million per MW on top of the base asset, and it requires overhead space, coolant distribution units, and a facility layout that most 2020-vintage designs simply do not have.
An asset without liquid cooling headroom is a stranded asset wearing a data centre costume. It will still lease — but to general-purpose enterprise workloads at a fraction of the rent per megawatt that AI training and inference commands. If the Johor campus was designed with liquid cooling provision, the $1B price is arguably cheap. If it was not, the buyer is purchasing a depreciation schedule. The wire item disclosed neither, which means the single most consequential technical fact about the transaction is invisible.
Latency deserves the same scrutiny. The 2–4ms round trip between Johor and Singapore is irrelevant for model training, which is batch-oriented and latency-insensitive. It is decisive for inference serving financial workloads, real-time risk, and anything touching a matching engine. That is not a hypothetical for me. My 2025 work on ETF arbitrage between TradFi custody and decentralised liquidity pools was entirely a latency business — mapping settlement differences to find a $150,000 annualised edge. In that world, four milliseconds is not a rounding error. It is the entire trade.
So the Johor campus has two distinct tenant pools with two distinct rent ceilings: AI training tenants who care about power and cooling, and latency-sensitive financial and cloud tenants who care about proximity to Singapore's interconnection fabric. An asset that can serve both commands a blended rent that an asset serving only one cannot. Again, undisclosed.
Tenant concentration is the next blind spot. Wholesale colocation is a business of very few, very large leases. A single 30MW commitment from one hyperscaler can represent the majority of a campus's net operating income. That is wonderful for cash flow stability and catastrophic for renegotiation leverage at renewal. The difference between a shell-and-core lease and a fitted lease is the difference between owning a warehouse and owning a warehouse plus somebody else's capital expenditure. Cap rates on single-tenant hyperscale assets compress to 5–6% precisely because the credit is investment grade — and blow out to 8–9% the moment that credit is a tier-two cloud provider nobody has heard of.
Now the tokenisation angle, because this is where the Johor transaction connects to the on-chain economy in a way the wire item never mentioned.
Real-world asset tokenisation has spent two years promising to bring infrastructure yield on-chain. Most of it has been treasury bills with extra steps. But data centre cash flows are structurally ideal for tokenised wrappers: long-duration, contracted, inflation-linked through lease escalators, and denominated in hard currency. A 15-year lease with an investment-grade hyperscaler is, functionally, a corporate bond with a cooling system attached. The only reason it has not been securitised and fractionalised at scale is legal and jurisdictional complexity — not financial engineering.
Data centre ABS issuance has been building quietly through 2024 and 2025, and every successful deal compresses the cap rate on the underlying assets by widening the buyer base. That is the real long-term story here. Not AI sentiment. The gradual conversion of physical compute into a tradeable, yield-bearing instrument.
Which means the buyer of this Johor asset matters more than the price. If the buyer is an infrastructure fund or a sovereign wealth vehicle, the signal is that data centre assets are entering the same institutional securitisation pipeline that swallowed toll roads, airports, and fibre. If the buyer is a hyperscaler buying its own landlord, the signal is vertical integration — and a long-term threat to the wholesale colocation model that AirTrunk built its franchise on. The wire item disclosed neither, which makes the single most important fact about the transaction invisible.
Let me also flag the regulatory layer, because I pivoted toward compliance-heavy analysis after the 2022 Terra collapse and I have not stopped watching it.
Johor's data centre boom was policy-enabled. Malaysia's Investment Development Authority offered tax incentives, the state offered land, and the power utility offered capacity. The Johor-Singapore Special Economic Zone agreement formalised the corridor further, aligning customs, talent mobility, and power planning across the strait. That worked until it didn't.
By 2024, Johor was facing genuine constraints: water for cooling in a region with seasonal stress, grid capacity in the southern corridor, and a political calculus that increasingly favours industrial diversification over raw megawatt growth. The most likely regulatory change over the next 6 to 12 months is not a ban — it is a tightening. Higher PUE requirements, water-use thresholds, local content rules, and a slower approval pipeline for new interconnections.
For an operating asset, that is neutral-to-positive: scarcity raises the value of incumbency and makes an already-permitted campus more valuable, not less. For a development pipeline, it is a direct hit to the option value of future phases. This distinction is the whole ballgame, and it is why the seller's motivation matters more than the buyer's.
And this is the part where the "AI selloff" framing collapses completely. If you believed AI demand was rolling over, you would be selling development land — the highest-beta, most option-heavy part of the portfolio — not operating assets with contracted cash flows. If you believed AI demand was durable but scarce in permitted capacity, you would sell an operating campus at replacement cost, recycle the capital, and redeploy into the next interconnection queue before policy tightens and the queue gets longer. The transaction structure described by the wire item is consistent only with the second interpretation.
CONTRARIAN
Here is the angle nobody published alongside the $1B number.
The phrase "as AI selloff bites" does three things at once, and all three are editorial rather than analytical. It supplies a cause the transaction did not disclose. It implies a direction of travel — downward — that the price does not confirm. And it converts a routine sponsor-level capital recycling event into a macro signal about artificial intelligence, which is a far more clickable subject than a cap rate decomposition.
Speed without precision is just noise; the "$1B" headline is precisely that — noise dressed as signal.
The real signal is the opposite of the caption. A sponsor that bought a platform for US$16 billion eighteen months ago, and is now trimming a single asset at what appears to be replacement cost, is telling you the platform is performing. You do not monetise a slice of a portfolio you believe is about to de-rate. You monetise a slice to establish a mark, seed the next vehicle, and demonstrate to LPs that the value-creation playbook works before you ask them for more capital.
I have watched this reflex before. In 2017, I found an integer overflow in the Parity multi-sig contracts during a casual review and pushed a warning into Telegram channels within minutes of the exploit's discovery. The lesson was not that code breaks — code always breaks. The lesson was that 17 reveals the true cost of trust, and the cost is always paid by whoever is last to read the disclosure. The retail reader who sees "$1B" and "AI selloff" in the same sentence has been handed a conclusion without a single input required to verify it.
Trust no one. Audit everything. Repeat.
TAKEAWAY
Watch three things, and none of them are the price. The buyer's identity — infrastructure fund means financialisation, hyperscaler means vertical integration, and those are opposite trades with opposite implications for every listed colocation operator. The implied dollars per megawatt once the asset's critical IT load is disclosed; that single ratio will tell you whether this was a premium, a mark, or a haircut. And whether the proceeds are redeployed into a new Malaysian interconnection queue within two quarters.
If they are, the "AI selloff" headline was written about a transaction that was actually about the opposite. If they are not — if the capital leaves the region entirely — then the caption was right for the wrong reasons, which is the most dangerous outcome of all. The market rarely punishes you for being wrong. It punishes you for being right too early with someone else's capital.
