Goldman Sachs has a warning for the market: power delays may hinder AI data center growth. Three sentences. No gigawatts. No terawatt-hours. No commissioning dates, no named author, no report title. And it reached me not through an energy desk but through a crypto outlet, republished for an audience that mostly does not own a utility. That is the first thing worth noticing, and the last thing anyone quoted. A sell-side warning stripped of every number is not a forecast โ it is a routing instruction, a way of telling institutional clients where to stand before the crowd arrives. Mining the liquidity where value truly pools has never meant chasing the headline. It means reading why the headline exists at this hour, in this feed, for these readers.
I have watched this pattern three times now. In 2017 the narrative was utility tokens, and the constraint everyone ignored was token distribution logic. In 2020 it was liquidity, and the constraint was impermanent loss nobody modeled. In 2022 it was algorithmic stability, and the constraint was trust itself. Each cycle, the crowd prices the visible asset and misprices the invisible bottleneck. Archaeology of the blockchain, layer by layer, always terminates in the same place: the scarce thing was never the thing being traded.
This cycle the bottleneck has a physical address. The International Energy Agency's April 2025 work puts global data center consumption near 415 TWh in 2024 โ roughly 1.5% of world electricity โ rising toward 945 TWh by 2030, with AI-optimized facilities carrying most of the increment. Lawrence Berkeley National Laboratory puts US data centers at 176 TWh in 2023, or 4.4% of national load, projecting 325 to 580 TWh by 2028. Those are not opinions. They are the reason a bank with no energy trading desk still feels compelled to say something out loud.
But the constraint is not generation. That is where the three-sentence version fails. The binding constraint sits in three segments, each with its own clock. First, interconnection: Berkeley's Queued Up series shows roughly 2,600 GW of capacity idling in US interconnection queues โ a figure that dwarfs national peak load โ with median time from request to commercial operation stretching from under two years in 2000โ2007 to about five years in 2018โ2023. Second, equipment: large power transformers now run 120 to 210 weeks in some categories on Wood Mackenzie's numbers, heavy-duty gas turbine slots are booked into 2028, and new units take five-plus years from order to operation. Third, siting and permitting, which has no standard clock at all.
Stack those and you get the only number that matters: a project approved today contributes zero megawatts before 2028. GPU generations turn over every 18 to 24 months. The grid turns over every five to ten. That mismatch, not scarcity of fuel or sunlight, is the actual story.
Now the part that reframes everything. Run the unit economics. A 100 MW IT load at PUE 1.2, $80/MWh and 80% utilization burns roughly $67 million a year in electricity. The GPU capital behind that same footprint โ call it 800 GB200 NVL72 racks, some 57,000 accelerators โ runs into the billions and depreciates over three years. Electricity is a rounding error against silicon. Which means operators are not optimizing for price. They are optimizing for availability, and they will pay almost anything for a year of head start. Price elasticity here is close to zero, and that is precisely the condition under which asset owners extract rents.
The shortage is also local, not planetary. Northern Virginia, Dublin โ where data centers once drew about 21% of Ireland's national electricity โ Singapore, and specific Texas nodes. Nowhere else is quite as tight. And the price signal has already turned political: PJM's capacity auction cleared near $29 per MW-day for 2024/25, roughly $270 for 2025/26, and about $329 for 2026/27. A nine-fold move in two cycles crosses the threshold where legislatures notice. Virginia, Georgia, Ohio and Texas are all mid-flight on large-load tariff classes and cost-allocation fairness.
Which brings us to the reason a crypto outlet carried this story at all. Bitcoin miners are the forgotten counterparty in the entire debate. A mining site is a pre-approved interconnection agreement with an energized substation attached โ the single scarcest object in American infrastructure right now. Following the code's whisper through the noise, the Core Scientific, IREN, Cipher and TeraWulf conversions to AI hosting are not pivots. They are arbitrage on queue position, the sale of a regulatory waiting period to the highest bidder. Anyone who modeled those deals as hash rate diversification missed the asset entirely.
The hyperscalers have already voted with contracts rather than commentary. 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 from Talen's Susquehanna site. Those are not sustainability gestures; they are hedges against exactly the delay the bank is describing. The genuinely scarce asset โ approved capacity, substation access rights, brownfield sites with existing thermal permits โ has been quietly repriced by private infrastructure funds for two years. Buying a company that already holds a queue position is two to three years faster than building one.
One distinction the warning elides, and it matters more than the warning itself: training versus inference. Frontier training concentrates in a handful of hyperscale campuses and is the load most exposed to a power delay. Inference is distributed, deferrable, and increasingly cheap to run on distilled models. If electricity becomes genuinely binding, what gets sacrificed first is the next frontier run, not the chatbot serving your queries. Sell-side notes rarely make this cut, because the linear extrapolation is the product. There is a second-order effect that will never print either: when power is constrained, cloud providers manage capacity by priority โ internal models and anchor tenants first, spot availability for everyone else. The startup assuming elastic compute from a hyperscaler is implicitly assuming it ranks above that hyperscaler's own roadmap. It does not.
Which returns us to the source. Goldman's institutional clients in this trade are utilities, transformer manufacturers, data center REITs and the hyperscalers themselves. A warning that power is scarce is, for that book, a buy recommendation wearing a risk label. I spent six months in 2024 interviewing portfolio managers at German banks about how digital gold was being repackaged as institutional-grade liquidity, and the lesson held: sell-side language is precise about positioning and deliberately vague about physics. The vagueness is not sloppiness. It is the message.
The counter-narrative is efficiency, and it carries harder numbers than the warning does. DeepSeek-V3's technical report put a headline training run at roughly 2.79 million H800 GPU-hours and about $5.58 million in compute, excluding prior research โ against public estimates several times higher for comparable capability. That is not a curiosity; it is proof that the compute required for fixed performance can compress within months. Per-watt inference improved roughly two to three times from Hopper to Blackwell at low precision, and liquid cooling pulls PUE from about 1.3 toward 1.1. Inference prices have fallen by multiples to hundreds of times a year depending on task, a rate no grid expansion can match. Demand itself is elastic: distillation, quantization and routing simple requests to small models all bend the load curve.
Where narrative fractures, the data speaks โ and here the fracture is the assumption that compute demand is rigid while supply is merely slow. Both are wrong. Supply is slower than advertised and demand is softer than advertised.
The deeper contrarian point is structural, though, and it is where most crypto-native optimism quietly dies. The decentralized-compute narrative claims this bottleneck will be solved by protocols. It will not. Power contracts are signed by legal entities with credit histories and indemnity obligations, not by smart contracts. The right to a substation sits with a handful of counterparties โ the same architecture I keep finding in DAO governance, where code is law dissolves the instant you locate the multi-sig that can upgrade the contract. Grid access is the physical analogue of admin keys. Nobody has proposed a credible way to decentralize it, because nobody can decentralize a credit rating.
What to watch is not the warning but its price. The miner-to-AI conversion spread โ dollars per megawatt of contracted capacity โ is the cleanest real-time quote on a queue position, and it surfaces in public filings long before it surfaces in any research note. PJM clearing prices are the second. Both move before the narrative does.
The question worth sitting with: when the marginal unit of AI capacity becomes a substation rather than a GPU, what exactly is the crypto industry holding โ a hedge, a toll road, or a story that outlived its collateral? Spotting the arbitrage in human psychology is the easy part. Pricing it is not.