Goldman Sachs did not publish a warning about AI. It published a position. The note โ three sentences, no gigawatts, no terawatt-hours, no timeline, four of its five data points traced back to the same sell-side source โ landed on a crypto desk before it landed on an energy desk. That routing tells you more than the text does. When a research shop starts repeating "power delays" to an audience of Bitcoin miners, the trade is not about whether AI grows. The trade is about who already holds the interconnection rights.
Here is the number that should have been in the headline. PJM's capacity auction cleared at roughly $29 per MW-day for the 2024/25 delivery year. It cleared near $270 for 2025/26. It cleared near $329 for 2026/27. That is a tenfold repricing of firm capacity in two auction cycles. No model needs to forecast AI demand to see what that does to the balance sheet of anyone sitting on a signed interconnect agreement. The Goldman sentence is downstream noise. The auction print is the order book.
I have traded through enough cycles to be bored by "X will be constrained by Y" research. The constraint is almost never the thing named in the title. It is the thing the title assumes you cannot buy. So let me reframe the entire discussion: the scarce asset in this cycle is not electricity, and it is not GPUs. It is the approved megawatt behind a substation that already exists. Everything else is a forecast. That one is a deed.
Context: What the Note Actually Says, and What It Carefully Does Not
Strip the marketing. The claim is that power availability, not chip supply, is becoming the binding constraint on AI data center buildout. On the surface this is true and independently verified by four separate sources that do not talk to each other.
The IEA's Energy and AI work stream puts global data center electricity consumption at roughly 415 TWh in 2024 โ about 1.5% of world electricity โ rising toward 945 TWh by 2030, with AI-optimized facilities as the dominant increment. Lawrence Berkeley National Laboratory's demand study has US data centers moving from about 176 TWh in 2023, roughly 4.4% of national load, to a range of 325 to 580 TWh by 2028, which is 6.7% to 12%. Those two datasets alone justify a serious conversation about grid planning.
But the note stops at "power delays." It does not decompose the bottleneck, and that omission is the whole story. There are three distinct chokepoints, and they have three different solutions and three different clocks.
Generation is one. Interconnection and transmission is two. Equipment delivery is three. Generation is the slowest to fix but the most politically legible. Interconnection is the fastest to politicize and the slowest to administratively clear. Equipment โ transformers, switchgear, HVDC, heavy-duty gas turbines โ is a hard physical ceiling that no policy memo can compress.
LBNL's Queued Up series is the document the note should have cited and did not. The US interconnection queue holds something on the order of 2,600 GW of nameplate capacity waiting for approval โ a figure that dwarfs national peak load. The median time from interconnection request to commercial operation stretched from under two years in the 2000-2007 window to roughly five years for 2018-2023 entrants. That is the real number. Not "AI uses a lot of power." The real number is that a project filed today has a median commercial operation date somewhere around the end of the decade.
Then the equipment clock. Large power transformers in some classes are quoting 120 to 210 weeks, per Wood Mackenzie's tracking. Heavy-duty gas turbine order books extend toward 2028 and beyond, and a new combined-cycle unit from purchase order to first megawatt routinely takes five-plus years. Read those two sentences together and the conclusion writes itself: a generation project greenlit this morning cannot deliver compute before 2028 at the earliest, and probably later.
That is the entire thesis. It took me four hundred words to say what the note compressed into one adjective โ "delays." The compression is not an accident. A decomposed bottleneck invites a specific trade. A vague "delay" invites a sector rotation.
Now the part the note never touches, and the reason a crypto desk cared: cost structure. Run the numbers on a 100 MW IT load at PUE 1.2, $80 per MWh, 80% utilization. Annual electricity spend lands near $67 million. The GPU capital behind that load โ call it 800 GB200 NVL72 racks, roughly 57,000 accelerators โ is a multi-billion-dollar line item depreciating over three years. The power bill is a rounding error against the silicon. Which means the elasticity of demand for electricity, at the margin, for a hyperscaler, is close to zero. They are not shopping for a cheap kilowatt. They are shopping for a guaranteed kilowatt on a schedule. Price is not the constraint. Timing is.
That single reframe destroys most of the analysis floating around this story. If power were a cost problem, the market would solve it with efficiency and location. Because it is a timing problem, the market solves it by buying whatever already has a queue position. And the largest pool of pre-approved queue positions in North America is not held by utilities, not by real estate, and not by hyperscalers. It is held by Bitcoin miners.
Core: The Order Flow Nobody Is Printing
Let me be precise about what a miner actually owns, because this is where most analysts wave their hands.
A mining facility at scale owns four things that matter: a signed interconnection agreement with a specific MW capacity, an energized substation and step-down transformers already on site, a utility relationship with a track record of actual load, and land with the environmental and zoning work already done. None of that is glamorous. All of it takes years to replicate and cannot be bought at any price on a reasonable timeline in the tightest nodes.
That is why the last two years of miner-to-AI transactions are not "pivots" and not "diversification." They are arbitrage on queue position. When Core Scientific contracted capacity to a hyperscaler, when IREN repriced its sites against AI demand, when Cipher and TeraWulf structured hosting deals, the underlying asset being repriced was never the ASICs. It was the interconnect. The market was paying for time it could not otherwise buy.
I have audited enough token distribution contracts in 2017 to know what a mispriced claim looks like. The tell is always the same: the thing everyone is bidding on is not the thing that is scarce. In the ICO era, everyone bid the token; the scarce thing was code integrity, and almost nobody priced it. Here, everyone bids GPUs and models; the scarce thing is an approved megawatt with a short path to energization, and almost nobody prices it. The pattern rhymes. It always rhymes.
Let me quantify the arbitrage to make it concrete rather than rhetorical. Suppose a miner holds 500 MW of contracted capacity at a site with existing infrastructure, and the all-in cost to energize incremental load is, say, $1.5 million per MW including upgrades. A hyperscaler facing a five-year queue and a 2026-2027 compute target will pay a substantial premium over replacement cost to compress that timeline by two to three years. The premium is not for the electrons. It is for the calendar. Anyone who models these deals as "power purchase at market rate" is missing the entire spread.
Now the second-order effect the note ignores completely: the bottleneck is geographic, not global. The constraint lives in Northern Virginia, in Dublin โ where data centers at one point represented roughly 21% of Ireland's national electricity consumption โ in Singapore, and at specific ERCOT and PJM nodes. It does not live in the world. This distinction matters enormously because a global "power shortage" narrative supports a broad sell-side rotation, while a nodal shortage narrative supports a very specific, very concentrated, very tradeable set of assets. The vague version is a story. The specific version is a position.
And the specific version is already visible in behavior, which is the only signal I trust. Microsoft signed a 20-year PPA with Constellation for the 835 MW Three Mile Island unit, targeting a 2027 restart. Amazon contracted roughly 1,920 MW of long-term supply from Talen at Susquehanna. Those are not press releases. Those are hedges against exactly the timing mismatch I described โ executed by the two counterparties with the best information in the market. When the largest buyers in the world pay for restarts and co-location instead of waiting for the grid, they are telling you the grid will not arrive in time. That is the most expensive, most credible data point in this entire story, and the note does not mention it.
Let me add the equipment layer, because it is where the physical ceiling actually binds. Transformers, switchgear, HVDC converter stations, gas turbines, backup power in the form of UPS and BBU, and liquid cooling CDUs are all quoting longer and pricing higher. That is a textbook seller's market. In a seller's market, the margin migrates to whoever controls delivery slots, not whoever controls demand. So the value chain reallocates: away from chips-models-applications, toward power assets, grid equipment, and interconnection rights. The beneficiaries are named and public โ electrical equipment makers, utilities with firm capacity, nuclear and gas operators, cooling and power distribution. The losers are equally named and less discussed: the mid-tier AI company that rents third-party compute and has zero power negotiating leverage, absorbing both a compute lease markup and an electricity passthrough with no ability to hedge either.
This is the part retail misses. The compute shortage narrative coming out of cloud providers is partly capacity management, not capacity absence. When a hyperscaler is power-constrained, it prioritizes internal high-value workloads โ its own models, its anchor customers โ and the elasticity of supply to everyone else drops. The startup does not experience "the grid is slow." The startup experiences "our reserved capacity got repriced or delayed." Same underlying cause, completely different transmission mechanism, and the second one is systematically under-modeled. I have sat on the buy side of enough capacity negotiations to know that the customer-facing symptom of a physical constraint is almost never described as the physical constraint.
Contrarian: The Efficiency Escape Hatch Nobody Prices
Now the counter-intuitive part, and the reason I would not simply buy the "power is scarce" narrative wholesale.
The note treats electricity as an exogenous constraint. That is a sell-side habit โ linear extrapolation of a demand curve against a supply curve that the model holds fixed. It is almost always wrong at the margin, because the system being modeled adapts. And there is hard evidence the adaptation is fast.
Algorithmic efficiency is improving at a rate that embarrasses grid construction timelines. DeepSeek's published training economics put a frontier-class model at roughly 2.79 million H800 GPU-hours and about $5.58 million in compute, excluding prior research โ materially below contemporaneous estimates for comparable capability. That is not a rounding error. That is proof that the compute required to reach a fixed performance level can compress dramatically within a single cycle. Third-party inference price tracking shows LLM inference costs falling by multiples to hundreds of multiples per year depending on the task. Compare that to a five-to-ten-year grid build. The demand curve is elastic on the efficiency axis in ways the supply curve can never be.
Hardware reinforces it. Each generation โ Hopper to Blackwell and beyond โ claims roughly two to three times the per-watt inference performance in low-precision workloads. Add liquid cooling dropping PUE from the 1.3 range toward 1.1, and you get rising compute density per rack alongside falling energy per unit of compute. The rack gets hungrier; the workload gets cheaper. Both are true, and the second one is systematically ignored.
And demand itself is not rigid. Inference can be routed, distilled, quantized โ small models handling simple requests, large models reserved for hard ones. Training can be scheduled across geographies and time zones. None of this eliminates the constraint. All of it means the constraint bites unevenly and later than a static model predicts.
So here is my contrarian read, stated plainly: the "power delays" headline is directionally true and tactically misleading. It is true that a hyperscaler cannot energize 2026 capacity that the grid will not deliver until 2029. It is misleading to conclude that AI growth is therefore capped, because the first thing sacrificed under a hard power constraint is not inference โ it is large-scale frontier training concentrated in a handful of mega-clusters. Distributed inference is far more flexible. The note does not distinguish training from inference, and that distinction is the difference between a sector-wide bear case and a very narrow, very specific set of delayed projects.
There is also a structural blind spot in the note's framing that I find more interesting than its conclusion. It presents only the risk side of regulation. But power politics cuts both ways. On one side you have cost allocation fights โ the PJM auction repricing has already triggered hearings and legislative activity in Virginia, Georgia, Ohio, and Texas around large-load tariff classes and fairness of cost sharing. On the other you have federal permitting reform and fast-track lanes for AI data centers. Which way the policy vector points is genuinely uncertain, and the note presents it as a one-directional headwind. That is not analysis. That is a risk disclosure dressed as a forecast.
I have been on the wrong side of a one-directional model before. In 2022 I held $2 million in UST because I believed the peg was structurally sound and I had not modeled the failure of the mechanism itself. The lesson was not "stablecoins are risky." The lesson was that any model with a single point of failure and no scenario for that failure is not a model โ it is a hope with a spreadsheet attached. The "power is scarce" thesis has exactly that shape. It has one input โ demand grows โ and no scenario for the input that breaks it: efficiency arrives faster than the grid. I am not saying efficiency wins. I am saying the note does not price the possibility, and an unpriced possibility is where the drawdown lives. That is not a hypothesis I have measured yet โ nobody has, cleanly โ but the direction of the error is clear.
The Miner Trade, Priced Honestly
Let me connect the contrarian case back to the crypto angle, because this is where I actually have an edge and where most of the audience for this story is standing.
If the bottleneck is timing and the scarce asset is queue position, then miners with energized capacity are structurally long the exact thing hyperscalers are short. That is the trade. But it is not a free lunch, and the risks are specific.
First, transferability. Not every interconnect agreement can be assigned or repurposed at will. Utility contracts carry load-profile assumptions, and a mining load and an AI training load have different shapes. The deal structure matters more than the headline megawatts. A site that looks like 500 MW of capacity on a slide can be 300 MW of usable, assignable, energizable capacity after you read the actual agreement. I have seen enough token contracts that promised the world and delivered a governance exploit to be deeply suspicious of capacity figures that have not been measured yet against the actual contract terms.
Second, community and regulatory risk. Miners have spent years building local relationships around noise and land use. Converting a site to AI hosting can intensify power draw and cooling demand, which reopens exactly the fights the miner had already won. The queue position is valuable precisely because it is hard to replicate; that same difficulty means it is hard to defend once challenged.
Third, counterparty concentration. A miner that converts the majority of its capacity to a single hyperscaler has traded commodity exposure for single-name exposure. In a bear market, single-name exposure is the thing that kills you. I learned that lesson in the DeFi summer of 2020 โ 140% APY across Compound and Aave felt like genius until the bZx exploit turned six months of carry into a 60% drawdown. Yield is never free. It is compensation for a risk you have chosen, knowingly or not. The miner-to-AI hosting yield is compensation for counterparty and contract risk, and it should be underwritten exactly that way, not celebrated as a pivot.
So the honest framing of the miner trade is this: it is a queue-position arbitrage with contract, community, and counterparty risk layered on top. The spread is real. The spread is also narrower than the headline multiple suggests once you price the risk. That is true of every trade I have ever made, and it is why I stopped trusting decks and started trusting verified repositories and signed agreements.
Takeaway: What to Watch, and at What Level
I do not do price targets on narratives. I do levels on flows. Here is what I am watching, in order.
The first level is the PJM capacity auction clearing price. If the 2027/28 auction clears above the 2026/27 print of roughly $329 per MW-day, the firm-capacity scarcity is confirmed and the queue-position trade is validated. If it mean-reverts, the whole "power is the binding constraint" story loses its cleanest public print, and I would expect the miner-to-AI multiple to compress with it. Watch the auction, not the note.
The second level is transformer lead times. If the 120-to-210-week quotes start shortening, the physical ceiling is easing and the urgency premium on existing capacity decays. If they extend, the premium holds. This is a slower signal but a more honest one than any demand forecast.
The third level is the efficiency print. Every time a frontier model ships at materially lower compute per unit of capability โ the DeepSeek-style disclosure is the template โ the "power caps AI" thesis weakens by exactly that amount. I do not need to forecast it. I need to watch for it, because the note does not.
And the fourth, the one that matters most for anyone holding miner equity: the contract terms. Not the headline megawatts. The assignability, the load profile, the termination clauses, the counterparty credit. The gap between the announced capacity and the energizable, assignable capacity is where this trade is won or lost. That gap has not been measured yet by anyone publishing research, and until it is, every "MW locked" figure in a press release should be read as a marketing number, not an asset.
The broader point is the one I keep coming back to after five cycles. The market rewards the people who price the thing that is actually scarce and punishes the people who price the thing that is loudly discussed. Right now the loud discussion is about whether AI is too power-hungry. The quiet, tradeable fact is that a signed interconnect behind an existing substation is worth more than its replacement cost by a margin that reflects years of calendar time โ and the balance sheet that holds those agreements is, more often than not, a Bitcoin miner. The Goldman note did not say that. But the desk that routed it to a crypto audience understood it perfectly. So the real question is not whether power delays AI. The real question is whether you are pricing the delay, or pricing the headline about the delay. One of those is a trade. The other is a story you tell yourself while someone else buys the megawatt.