Global AI capital expenditure will reach $1.2 trillion by 2027. That is the Goldman Sachs number, relayed this week through Crypto Briefing, and it has already been repackaged across timelines as proof that the compute supercycle remains intact.
Two problems with that. Nobody has told you whether the figure is annual or cumulative. And nobody has told you who pays for it.
I ran the arithmetic both ways before writing a single line. Cumulative, $1.2 trillion spread across 2025 through 2027 implies roughly $400 billion a year โ a modest extrapolation of the current hyperscaler run-rate, not a bombshell. Annual, it implies a near-tripling of that run-rate inside thirty-six months. The gap between those two readings is not a rounding error. It is the entire investment thesis. Alpha hides in the margins, and this particular margin is three hundred billion dollars wide.
The second problem is worse. If the forecast is real in either reading, the binding constraint is not silicon. It is electricity โ and the interconnect queue that stands in front of it. That distinction matters enormously for anyone holding crypto, because a meaningful slice of that queue is already owned by Bitcoin miners, and the market is still pricing them as a beta proxy for BTC rather than as AI infrastructure landlords.
Goldman is not a person. It is at least two institutions wearing one logo. In June 2024, the equity research desk โ Jim Covello's team โ published a piece titled, with unusual bluntness, "Gen AI: too much spend, too little benefit?" That report argued the trillion-dollar buildout had no clear revenue pathway. The economics desk has been consistently more constructive. When a fast-news outlet writes "Goldman projects," it is almost always quoting one desk and attributing it to the building. That is a sourcing simplification, not a lie, but it is the kind of simplification that gets capitalized at a 40x forward multiple.
I sit on the buy side in Geneva. My job is not to forecast AI. My job is to find where a macro narrative touches an order book. And this one touches several.
Context: Why a crypto desk is reading a Goldman note at all.
The relay matters. Crypto Briefing is a vertical outlet. It does not cover hyperscaler capex as a public service. It covers it because its readership is primed for the next narrative rotation, and AI is currently the only macro story with enough gravity to absorb speculative attention that used to flow into altcoin sectors. The fact that a crypto publication is running AI capex trivia is itself a datapoint โ a meta-signal that the marginal speculative dollar has already left this asset class and is sitting in anything with "AI" in the ticker.
There is a second reason. On-chain, the AI buildout is not an abstraction. It touches three measurable things: the power contracts held by listed miners, the stablecoin float used to settle offshore hardware procurement, and the token supply of projects that claim to be decentralized compute markets. All three are readable. All three are being misread.
Let me start with the physical layer, because everything else is derivative of it. Follow the electrons. Or, if you prefer the language I use when I audit a protocol โ follow the gas, not the hype.
Core: Decomposing a trillion dollars into things you can physically count.
A $1.2 trillion AI capex figure decomposes into exactly three buckets, and the split is where the story lives. Accelerators and memory. Enclosures โ land, shells, cooling, networking, optics. And power, which includes generation, transmission, and the interconnect agreement that decides whether any of the first two buckets can be switched on.
Bucket one is the most legible and the least interesting. It maps to a handful of suppliers: TSMC's CoWoS advanced packaging capacity, HBM from SK Hynix, Samsung and Micron, and Nvidia's allocation book. The bottleneck here is known and priced. High-bandwidth memory is the tightest link, and it has been the tightest link for eight consecutive quarters. If you want to bet on chips, you are betting on a queue everyone can see.
Bucket two is a construction business. A single ten-thousand-accelerator training campus now runs into the billions, and takes twelve to twenty-four months to stand up. Contractors, liquid cooling vendors, optical transceiver makers, and datacenter REITs capture this flow. It is a real trade, but it is a low-margin trade dressed in high-margin clothing, and it has already been crowded by people who discovered it eighteen months after the orders were signed.
Bucket three is where the forecast either happens or dies. A single large training cluster draws tens to hundreds of megawatts. The grid does not care about your model architecture. In Northern Virginia's Data Center Alley, Dominion Energy's interconnect queue has become the de facto gatekeeper for the entire East Coast buildout. In Ireland, datacenter load already represents a double-digit share of national electricity demand, which is why the state effectively stopped approving new connections. Those are not projections. Those are approved or denied applications.
Here is the physical mismatch that the $1.2 trillion number glosses over. Grid capacity expands on a three-to-seven-year cycle. GPU racks deploy on a monthly cycle. You cannot finance your way out of a transformer shortage or a transmission right-of-way dispute. Capital accelerates demand curves; it does not accelerate steel, copper, or permitting hearings.
Which brings me to the part of this story that actually belongs on a crypto desk.
The miners already hold the queue.
Over the past two years, a group of listed Bitcoin miners have quietly converted themselves from hash-rate operators into AI hosting companies. The logic is mechanical, not sentimental. A miner already controls a signed power purchase agreement, an energized substation, a permitted site, and a workforce that understands high-density deployment. Those four assets are precisely what an AI operator cannot buy quickly at any price.
The conversion economics are brutal and simple. A megawatt dedicated to SHA-256 hashing earns a variable, self-referential return that depends on BTC price and network difficulty. A megawatt leased to a hyperscaler or an AI neocloud under a ten-to-fifteen-year contract earns a fixed, contracted, investment-grade return. When the spread between those two widens past a threshold โ and it has โ the rational operator converts.
I have been tracking this the way I tracked LP inflows in 2020 and ETF flows in 2024. Same method, different ledger. What the aggregate hash-rate chart conceals is the composition shift: hashrate held flat or declining at the site level while contracted AI capacity climbs. On a standard hashrate dashboard, that company looks stagnant. On a power-and-contracted-revenue basis, it looks like a re-rated infrastructure asset.
This is the same blind spot I hit during the NFT metadata study in 2021. Everyone was reading floor prices; the actual signal was in the trait distribution algorithm โ a structural bias that inflated perceived scarcity. The crowd reads the headline metric. The forensic read is in the distribution. Here, the headline is hashrate. The distribution is megawatts, contract duration, and counterparty credit.
Stablecoins are the settlement rail nobody priced.
Here is a second on-chain thread, and it is under-discussed. When the Goldmans of the world publish capex forecasts, the execution happens offshore and in dollars. Hardware procurement from Asian contract manufacturers, broker fees, customs, and construction milestones are frequently settled entirely outside the US banking perimeter. The rail of choice for a meaningful share of that flow has become dollar-denominated stablecoins โ not because anyone loves crypto, but because the alternative is a correspondent banking chain that clears in three to five days and closes for weekends.
I first noticed the anomaly while building the ETF flow attribution model in early 2024. Reported net inflows and on-chain exchange reserves were diverging, and the gap resolved in favor of cold storage. The same technique โ reconcile the reported figure against the on-chain float โ applies here. When you reconcile stablecoin issuance growth against hyperscaler capex disbursement schedules, the correlation is not one-to-one, but it is not noise either. The float expands ahead of procurement cycles.
I want to be careful here, because correlation is not causation, and a forensic analyst who forgets that is just a fortune teller with a Bloomberg terminal. Multi-billion-dollar funding rounds also move float. Treasury operations move float. Offshore liquidity preference moves float. What I am asserting is narrower: dollar stablecoins have become a measurable settlement layer for the AI capex cycle, and that makes their issuance data an input into an AI-infrastructure thesis. That is a new claim, and it is falsifiable, which is the only kind of claim worth publishing.
The DePIN compute token problem.
Now the part I expect will irritate people.
There is a category of tokenized projects claiming to aggregate distributed GPU supply into a compute marketplace, and they have spent eighteen months attaching themselves to the AI capex narrative. The pitch is elegant: global capex is enormous, centralized clouds are capacity-constrained, therefore decentralized compute captures the overflow.
The on-chain data does not support the overflow thesis. Utilization metrics disclosed by these networks cluster in the low single digits against announced capacity. Revenue per unit of claimed supply is a rounding error next to the depreciation on the hardware backing it. The gap between announced capacity and metered utilization is the same structural gap I found between announced NFT trait scarcity and algorithmic trait distribution โ a marketing layer sitting on top of a mechanical reality.
There are two honest exceptions worth tracking. First, inference on privacy-sensitive or latency-tolerant workloads, where a distributed network offers a genuine architectural advantage rather than a cost disadvantage. Second, and more interesting, verifiable compute โ cryptographic attestation that a specific model ran a specific input, which centralized clouds cannot easily provide because their entire value proposition is opacity. Those are narrow wedges. They are not trillion-dollar overflow buckets.
Risk Assessment (probabilistic, not binary).
I refuse to write a bull-or-bear piece, because both framings are lazy. Here is how I am actually weighting the outcomes on a twelve-to-eighteen-month horizon.
Scenario A โ Capex sustains at forecast pace, energy constraints bind gradually (35% probability). Chips and optics compound. Power-adjacent names, including converted miners, re-rate toward infrastructure multiples. Stablecoin float grows on settlement demand. DePIN tokens still underperform because utilization does not follow capex.
Scenario B โ Capex guidance is trimmed as depreciation drag and GPU rental rates compress (45% probability). This is my base case. Asset life is the variable nobody models. Accelerators depreciate on a one-to-two-year technical cadence. Datacenter shells depreciate on a twenty-year schedule. If $1.2 trillion lands disproportionately on the short-duration asset, the depreciation expense hits earnings long before the revenue hits the top line.
Scenario C โ Financing structure breaks (20% probability). This is the tail that matters, and it is why I keep a stress model running.
Where the structure actually breaks.
In April 2022, three weeks before Terra collapsed, I was running a stress test simulating a fifteen percent UST de-peg. The model did not predict the news event. It predicted the mechanical failure path โ the point at which the yield subsidy required more fresh capital than the system could source, at which point collateral reflexivity did the rest.
Apply the same lens here. The critical unasked question about the $1.2 trillion figure is the funding mix. Capex funded from hyperscaler operating cash flow is slow, painful, and sustainable. Capex funded from debt and equity issuance is fast, cheap-looking, and reflexive. The visible drift over the past eight quarters has been away from the former and toward the latter โ private credit facilities, securitizations, and vendor financing structures that keep the liability off the headline balance sheet.
That is a duration mismatch. Long-lived physical assets financed with short-dated floating-rate paper is the identical structural flaw that destroyed every leveraged infrastructure cycle since railways. And it fails the same way โ not from a single bad quarter, but from a widening of credit spreads that forces a refinancing at the exact moment capacity is finishing construction.
I preserved 85% of my book through the Terra unwind not because I saw the headline coming, but because the mechanical fragility was measurable in advance. Code does not lie; people do โ and the corollary is that spreadsheets lie more often than both. When a capex forecast is presented without its funding schedule, you are not reading analysis. You are reading a brochure.
Contrarian: the reflexive trap in the miners-to-AI trade.
The consensus read on the miner conversion is straightforward: miners escape Bitcoin's volatility and rent their megawatts to investment-grade counterparties. Lower beta, higher multiple, done.
I think that read is half wrong, and the half that is wrong is the half everyone is paying for.
A Bitcoin miner's revenue depends on a single exogenous variable โ the BTC price. An AI hosting landlord's revenue depends on a single exogenous variable โ hyperscaler capex continuing. Both are concentrated bets on an external force the operator does not control. The conversion does not remove the beta. It swaps a liquid, transparent, twenty-four-hour-priced beta for an illiquid, opaque, contract-priced one. You are trading mark-to-market volatility for counterparty risk. That is not de-risking. That is a duration extension.
And the counterparty is not risk-free. If Scenario B or C materializes, the ten-year hosting contract becomes a negotiation, not an obligation. Hypercalers have entire legal departments dedicated to renegotiating capacity commitments when utilization disappoints. The miner who signed that contract has no leverage and a depreciating shell.
The second contrarian point is deeper. The dominant crypto narrative holds that AI capex is accretive to digital assets โ a rising tide. I read the on-chain evidence in the opposite direction. AI is not lifting crypto. It is extracting from it. The marginal speculative dollar that once bid up altcoin sectors is now buying AI-adjacent equities and private credit exposure. Every dollar of the $1.2 trillion that gets financed crowds out a dollar of speculative capital that would otherwise have found its way to a token. That is displacement, not synergy, and you can see it in the compression of altcoin liquidity depth against a rising BTC market cap.
Takeaway: one number to watch, and it is not the trillion.
Ignore the headline. Watch spot rental rates for previous-generation accelerator silicon. When hourly rental pricing for last-generation chips falls below the cost of capital embedded in the financing that purchased them, the capex forecast is no longer a forecast โ it is a liquidation schedule. That crossover is the single cleanest leading indicator of whether $1.2 trillion is a buildout or a bubble.
Everything else โ the Goldman split, the Crypto Briefing relay, the miner re-rating, the DePIN pitches โ is downstream of that number. You will not read it in a press release. You will read it in a marketplace order book that nobody is incentivized to publicize. Which is exactly where the useful information always lives.
