The $725 Billion Liquidity Vacuum: AI's Capex Supercycle and the Quiet War for the Grid
Over the past four quarters, Amazon, Microsoft, and Alphabet have committed a collective $725 billion to artificial intelligence infrastructure. My eye is on the horizon, not the hourly candle, but even from a distance this number demands a reckoning. It exceeds the entire market capitalization of crypto at its 2023 trough. It rivals the annual economic output of a mid-sized developed nation. And it is being channeled through just three corporate balance sheets into one of the most constrained physical bottlenecks on the planet: the electricity grid.
In the United States, data center interconnection queues already stretch three to five years at major transmission operators like PJM and ERCOT, and transformer lead times extend two to four years. When three companies decide to move this volume of capital into power purchase agreements, silicon procurement, and buildings that do not yet exist, they do not merely participate in the AI trade. They become the trade.
The figure resists easy framing. $725 billion, spent over three to four years and depreciated on a five-year schedule, produces an annual earnings drag of $100 to $150 billion across three companies — more than the net income of most S&P 500 constituents. The conventional read treats this as demand. The honest reading is structural: the AI build-out has shifted from optional acceleration to mandatory fixed cost, a mortgage on the balance sheet that must be serviced regardless of the weather.
Beneath the headline number, composition matters more than total. A growing share of the spending is migrating away from NVIDIA's general-purpose GPUs toward custom silicon: Google's TPU, Amazon's Trainium, Microsoft's Maia. The hyperscalers are integrating vertically because the economics of AI at this scale no longer tolerate a single supplier's margin. I have seen this script before. It is the same arc that Bitcoin mining traced a decade ago, when general-purpose GPUs gave way to FPGAs, and FPGAs to application-specific miners. The computing world is recursively discovering that specialization is the only way to survive scale. In 2026, the hyperscalers have become the ASIC makers of their own gold rush.
There is a recurring rhythm to infrastructure cycles that market participants conveniently forget. The fiber optic build-out of the late 1990s was overbuilt by roughly 70 percent nationwide, and it took nearly a decade for the capacity to find its price. The shale boom of the 2010s taught global energy markets that oversupply arrives precisely when the investment thesis is most confident. The ICO wave of 2017 taught my own industry that capital committed to adoption curves that have not yet matured does not merely underperform; it destroys value in a highly visible way. This $725 billion is the same animal, running at a faster speed. It does not mean the investment thesis is wrong. It means the gap between capital deployment and genuine revenue resonance is the place where market structure gets tested.
For those who watch liquidity cycles for a living, the transmission channels into crypto are threefold.
The first channel is financial, and the least discussed. When three of the largest equities on earth commit to a fixed capex trajectory of this magnitude, they compress their own free cash flow for the foreseeable future. That compression matters because digital assets still trade as high-beta satellites to the liquidity breadth of the US equity complex. A margin disappointment in megacap tech does not stay contained. It ripples outward into every risk asset sharing the same marginal buyer. The correlation is not ideological; it is mechanical. The same institutional allocator that rotates into bitcoin is usually the one holding Microsoft and Alphabet, and when the diversification overlay tightens, everything with volatility gets sold first.
The second channel is physical, and it may prove more important. AI data centers demand baseload power around the clock, with virtually zero tolerance for interruption. Bitcoin miners, by contrast, are the most flexible electrical loads ever constructed; they can curtail within milliseconds when grid operators call. From my audit work modeling the interplay between compute and power markets, the pattern is unmistakable. The hyperscaler build-out has absorbed the slack in wholesale electricity markets that miners once consumed, pushing mining operations toward the stranded assets — flare gas fields, curtailed hydro, remote wind — that hyperscalers cannot chase because their customers live in cities and their data centers cannot wander. This is often framed as competition, and superficially it is. But it is better understood as the market's clumsy attempt at a division of labor. The hyperscaler is the massive, immobile buyer of reliability. The miner is the agile, distributed buyer of last resort. In the years ahead, that flexibility will become an asset class of its own. The miner that survives this cycle will not be the one that hoarded the cheapest watts; it will be the one that sold the grid something AI cannot deliver — optionality.
Power demand growth of this magnitude has also revived the nuclear conversation in ways that would have seemed improbable a decade ago. Small modular reactor designs once dismissed as uneconomic are now being contracted off the drawing board by hyperscalers who cannot find enough grid capacity to meet their own commitments. The deeper implication for crypto is that the energy mix is shifting in real time; jurisdictions rich in nuclear and stranded renewables become the new compute havens, and mining operations are following the same map.
The third channel is structural, and I want to give it the most weight. A significant portion of this build-out is wrapped inside long-term GPU capacity agreements — structured contracts between cloud providers and AI labs like OpenAI and Anthropic that pre-sell compute years in advance. These arrangements reduce the risk of building capacity nobody will rent. But they transfer the credit risk of the entire enterprise to the weakest balance sheet in the chain: the AI startup. The geometry is darkly familiar. Substitute "GPU capacity agreement" for "liquidity mining contract" and you are looking at the same architecture that produced the DeFi boom of 2021: top-line revenue commitments built on the assumption that end-user demand will keep extending forever. I spent eight months as a junior analyst modeling the sustainability of yield farming protocols, and the lesson that stayed with me is simple: when the yield runs out, the farm consolidates. The same pruning is coming to AI capacity markets, and it will arrive on a familiar schedule — the lag between the announcement of the commitment and the recognition that the revenue to service it has not yet arrived.
From a positioning standpoint, the asymmetry is unusual. The market has already repriced semiconductor suppliers, power equipment manufacturers, and even uranium developers into the AI trade. Bullish consensus is fully embedded in those charts. But the assets that would benefit from the downside scenario — the hedges against an AI revenue miss — remain unloved. I have spent the past three years studying the correlation between bitcoin mining equities, grid-balancing assets, and cloud infrastructure plays, and the reading is consistent: when the AI narrative stumbles, the rotation into energy-flexible and trust-layer assets is sudden, violent, and unforgiving for late entrants.
The consensus framing of this $725 billion is relentlessly bullish — for chips, for megacap tech, for the AI narrative itself. I want to offer a quieter, uncomfortable thesis: this capex cycle may be the largest subsidy ever granted to decentralized computing, precisely because it is so centralized.
Consider the cost curve. By 2028, the hyperscalers will almost certainly have built enormous overcapacity in compute, because corporate planning cycles are structurally biased toward overshooting the demand they are trying to forecast. That overcapacity will eventually be sold at whatever price clears the market. Decentralized AI networks, federated learning protocols, and on-chain inference markets will be the beneficiaries of a subsidy they did not ask for and did not pay for. The centralized build-out is inadvertently funding its own decentralized competitors. The bust, when it comes, was never an end; it was a necessary pruning — of weak hands, of thin value propositions, of businesses built on borrowed narratives. The AI capacity glut will prune the hyperscalers' pricing power and fertilize the soil for something more distributed.
The tide does not ask where we entered before it turns. But it does reward the prepared observer. The quieter question inside all this spending is not whether AI will be profitable, but which internet — the closed one of hyperscale clouds or the open one of verifiable ledgers — will inherit the surplus. If compute becomes cheap and abundant enough, the marginal cost of training collapses, and the moat of the compute aristocracy narrows. The real scarcity in that world will not be silicon. It will be trust. And trust, at least, is a ledger problem.
This is where my own bias surfaces, and I state it plainly. I built my first quantitative risk model for a Bitcoin ETF strategy, then spent the following years auditing AI-generated content on public blockchains, trying to preserve some trace of human provenance in an automated sea. The threads are the same. Every technology cycle produces two classes of winners: those who monetize the infrastructure and those who monetize the trust — or the lack of it. NVIDIA is the infrastructure class. The miners who become grid-flex suppliers are the trust class.
For digital asset markets, the $725 billion is not a distant headline; it is the macro backdrop against which the next cycle will be traded. The depreciation charge alone — $100 to $150 billion annually — is a shadow balance sheet that will discipline every risk appetite decision between now and 2028. When AI revenue growth finally slows below the capex trajectory — and it will, because every capex cycle overshoots — the reckoning will not be contained to three technology companies. It will ripple through the energy complex, the chip supply chain, and every risk asset that borrowed confidence from the AI trade's success.
The ledger does not lie; it merely waits for the same mistake to be repriced. For those of us watching from the edge of the grid, holding a two-decade-old conviction that liquidity moves in cycles, the orientation is simple. The machines are learning. The market should be too. The question is whether we have been reading the right ledger — and whether we will have the patience to wait for the account to be settled.