I want to start with a number most people in this industry scroll straight past.
Five to fifteen kilowatts. That is the power envelope of a conventional data center rack — the steel cage that has carried the internet on its back for twenty years. Hold that figure in your mind.
Now set beside it what is rolling into every AI conversation on the planet. NVIDIA's GB200 NVL72, and the Rubin-generation racks queued behind it, push a single cabinet past one hundred kilowatts. Some liquid-cooled configurations edge toward one hundred and twenty. That is not an upgrade. That is a phase change — the kind that moves a problem out of the realm of engineering and into the realm of physics, where nobody can negotiate.
Last week, Goldman Sachs finally stopped talking about the chips and pointed at the wall behind them. Next-generation AI servers, the bank warned, could trigger localized blackouts.
Not global. Not apocalyptic. Localized. A single word doing an enormous amount of quiet work — and almost everyone who repeated the headline let it slide. I did not. When a bank as disciplined as Goldman chooses that adjective, it is telling you precisely where the failure will surface, and who ends up holding the bill.
Here is why this story landed in a crypto publication at all, and it is not an accident.
It ran through Crypto Briefing. That is not an energy desk. It is not an AI desk. It is a crypto desk — and that single fact tells you more about the shape of the narrative than any number in the wire copy. The real story humming underneath Goldman's warning is the migration of Bitcoin miners into AI data centers, a pivot that has been quietly accelerating through this entire bull market while most people were staring at charts.
Why do miners matter here? Because a mining farm is, physically, almost an AI data center waiting to happen. It already holds the three assets that are hardest to acquire: an energized substation, an interconnection agreement with the local utility, and land zoned for brutal industrial load. What it lacks is revenue that scales. And that is where the math turns both brutal and beautiful at once.
The economics of hash rate are deflationary by design. Block rewards halve. Competition compresses margin toward zero. Compute sold into AI training, by contrast, prices against a demand curve that is close to vertical. So the rational miner does the rational thing: rip out the ASICs, pour in the liquid cooling, and sell the same megawatt to a hyperscaler at a multiple of what mining ever paid.
I have watched this pattern from inside the machine. Back in 2020 I treated five thousand dollars of savings as a laboratory, forking yield strategies on Uniswap and SushiSwap, losing forty percent of my capital to impermanent loss, and publishing every failure in public. What I learned then still holds: in crypto, the story always runs three moves ahead of the spreadsheet. The mining-to-AI pivot is the same pattern at industrial scale — a beautiful thesis, a hidden cost, and a deadline nobody has priced. Hold that thought. It becomes the contrarian turn of this entire piece.
Now the technical heart of it. I have audited enough protocol economics to be allergic to vibes, so let me ground this in the supply chain instead.
The first thing to understand is that the power density jump is not linear. Going from fifteen kilowatts to one hundred and twenty is not a bigger breaker and a fatter cable. It invalidates the entire thermal and electrical architecture of a legacy facility. Air cooling dies at around twenty to thirty kilowatts per rack. Past that, you are forced into liquid — cold-plate first, and for the densest deployments, full immersion. That means new manifolds, new fluid chemistry, new leak-detection, and a building that was never designed to hold hundreds of gallons of dielectric fluid above millions of dollars of silicon.
On the electrical side, the shift is just as severe. Traditional data centers run alternating current up to the rack and convert locally. High-density AI halls move toward high-voltage direct current, busway distribution, and rack-level power shelves. Every one of those components has its own supply chain, its own lead times, and its own single points of failure.
And this is where Goldman's word — localized — earns its keep. The bottleneck is not the server. The server ships. The bottleneck is the grid the server plugs into. The real constraint on AI compute is not silicon. It is the transformer, the interconnection queue, and the substation down the street.
Consider transformer lead times. Before the boom, a large power transformer might arrive in twelve to eighteen months. Today, utilities are quoting three to four years and rising. Interconnection queues in major markets like PJM and ERCOT stretch years deep, backlogged behind solar, storage, and data center requests all fighting for the same copper. A hyperscaler can write a check for a hundred thousand GPUs overnight. It cannot write a check that makes a four-year transformer appear next week.
This is why the failure is localized. High-voltage transmission is largely resilient — the bulk system is engineered for exactly these swings. The fragility lives at the distribution level: the feeders, the local substations, the neighborhood circuits that were sized for homes and small businesses and are suddenly being asked to feed a campus drawing the equivalent of a small city. When several of those clusters land in the same utility territory within the same two-year window, the local grid does not collapse — it browns out, throttles, and defers. Reliability degrades. That is what Goldman is describing, and it is far more mundane, and far more likely, than the apocalypse the headline implies.
So the honest framing lives somewhere the media skipped. The story is not "AI servers cause blackouts." The story is "the electricity system cannot be summoned on the timeline that capital demands." One is a headline. The other is an investment thesis.
That thesis has a face, and it is the crypto miner turned colocation landlord. When a miner sells its megawatt to an AI tenant, it is effectively selling a scarce permit that took years to obtain. The land is cheap. The interconnect is priceless. And in a world where every new gigawatt of AI capacity waits in a queue, an already-energized site trades at a premium no greenfield project can match.
But there is a second-order effect that almost nobody is modeling, and it cuts against the whole decentralized story this industry tells about itself. The mining-to-AI migration concentrates compute. It takes a globally distributed, permissionless network of hash power and funnels it into a handful of hyperscalers running closed, proprietary training clusters. The physical infrastructure of decentralization is being quietly converted into the physical infrastructure of centralization. We call the resulting entity a data center. We should call it what it is: a moat.
This is where my own work has been heading. In 2022, through the worst of the bear market, I locked myself in a Seattle apartment for six months reading zero-knowledge papers and writing a manifesto on privacy as a human right. What I took from that period is simple: the technology is never neutral, only the incentives are. And the incentives of the AI buildout currently point straight at consolidation — fewer operators, larger campuses, tighter control over the compute that will train the models mediating everyone's daily life.
There is a counter-current, and it is worth naming carefully rather than cheering blindly. Decentralized physical infrastructure networks — DePIN — are trying to route compute through distributed, permissionless supply. On paper this is the answer: coordinate idle GPUs, reward providers, keep the network open. In practice, the same physics bites. A distributed model cannot escape the grid any more than a centralized one can. Every node still draws from a local feeder. Decentralization does not repeal thermodynamics. You cannot fork a substation.
Which brings me to the part that will annoy the maximalists.
The loudest voices in this space treat the mining-to-AI pivot as pure triumph — a heroic second act for an industry the mainstream wrote off. I think that reading is too generous, and slightly dishonest.
Here is the uncomfortable truth. Most of the assets being converted were built on a decentralization narrative. They were financed by token holders who believed in a permissionless network of sovereign miners. And then those same assets are being quietly re-pointed at the most centralized, most gated compute market that has ever existed. That is not a pivot. That is a confession — a quiet admission that open networks could not monetize the hardware fast enough to survive, so the hardware went to work for the closed ones.
None of this makes the miners villains. I would make the same trade. The revenue curve gives you no choice, and a landlord who refuses to rent is not principled — he is just bankrupt. But we should be honest about what is happening. The same entities now writing bullish threads about decentralized compute are, on their balance sheets, betting against it.
And there is a deeper irony worth sitting with. The very constraint Goldman identified — local grid capacity — is an argument for decentralization made in physical terms, not ideological ones. If concentrated compute keeps triggering localized reliability problems, the long-run answer is not a bigger campus. It is more nodes, spread thinner, closer to where energy is actually cheap. The market may rediscover decentralized infrastructure not because it is beautiful, but because the grid forces it to. That is the rare case where the values-based argument and the ruthless economic one arrive at the same address.
So where does that leave the reader holding a portfolio in the middle of a bull market?
The temptation is to chase the obvious names — the miners that already rank among the best-performing equities of this cycle on the strength of the AI hosting story. I understand the impulse. But the crowded trade is the one where everybody already knows the punchline. The less obvious exposure is the unglamorous plumbing: the transformer manufacturers, the grid equipment suppliers, the switchgear and busway vendors, and the regulated utilities whose service territories happen to sit where the data centers are landing. Those businesses do not need a narrative to work. They just need a queue that keeps getting longer.
Compute is becoming the most valuable commodity on earth, and compute runs on electrons. The token model has finally met the load model, and the load model is patient, physical, and indifferent to sentiment. Decentralization is a verb, not a noun — and verbs require power. Most of this industry spent the last cycle arguing about the noun. The next one will be decided by who controls the verb, and where the current actually flows.
