The Golden Goose Doesn't Lay Eggs in the Dark: Trump's AI Sermon and the Power Queue Crypto Already Owns

The queue does not care about speeches. In the PJM interconnection queue — the waiting line where every new power plant, every data center, every hungry load queues up to plug into the largest grid in America — projects sit for four to seven years before a single engineering study is finished. Some wait longer. The number that should have anchored every headline on September 14 was not the gigawatts announced in a press release. It was the gigawatts that can actually be energized before this decade ends. On current physical constraints, that number is small, and it is shrinking against the demand curve. Then a sentence landed. Donald Trump, with the serene confidence of a man who has never once stood in a queue, told the country not to "stifle the golden goose" of artificial intelligence. The line moved through every wire service — and, strangely, through crypto feeds too. Ours included. We didn't ask why. We didn't explain the connection. We copied the bird and moved on. That silence is the story. Because the golden goose and the bitcoin miner are standing in the same line, and only one of them was invited to the podium. Let me place my cards on the table before I deal. I run a crypto newsroom in Riyadh. I have watched three full narrative cycles from the inside — the 2018 collapse, the 2020 liquidity-mining summer, the 2022 aftermath — and I have learned that when a blockchain outlet suddenly discovers an AI policy quote, something is being left unsaid. Usually, that something is power. Literal, metered, megawatt power. The kind measured in dollars per kilowatt-hour, not in tweets per minute. Here is the background the fast-news wires skipped. The statement — "don't stifle the golden goose" — is not a technology policy. It is a capital-markets policy wearing a technology costume. Its real function is to lower the political and regulatory uncertainty premium attached to AI infrastructure spending. Strip the rhetoric and the mechanics are simple: reduce permitting friction, reduce environmental review friction, reduce grid-interconnection friction, and the expected return on a multi-billion-dollar data center rises without a single line of model code improving. That is the entire trade. The scale it is trying to protect is enormous. The four largest hyperscalers — Microsoft, Google, Amazon and Meta — were guiding toward roughly 350 to 400 billion dollars in combined capital expenditure, with a meaningful share aimed at AI compute. That figure is an estimate, because each company defines its fiscal year and its "AI" bucket differently, but the order of magnitude is not in dispute. What is in dispute is whether any of it gets paid back. And here is where the crypto reader should sit up, because we have seen this movie before. In 2020, during DeFi Summer, I helped coin the phrase "liquidity mining as social contract" — the idea that yield farming was less a financial mechanism than a governance experiment dressed in a spreadsheet. That framing was useful. It was also, in hindsight, a little too kind. Liquidity mining was a subsidy structure that looked like a community and behaved like a treadmill. The AI capex boom has the same rhetorical architecture: a technology story wrapped around a subsidy, with the subsidy quietly billed to someone who wasn't in the room. That someone, in both cases, is the user of the underlying commodity. In DeFi, it was the late liquidity provider. In AI, it is the ratepayer. The hardest constraint in the AI economy is not silicon. It is copper, steel, and the signature at the bottom of an interconnection agreement. Start with the demand curve. Data centers consumed roughly 4% of total U.S. electricity in 2023. Depending on whom you believe — the EIA, EPRI, or the IEA, whose methodologies diverge sharply — that share is projected to reach 6% to 12% by the end of the decade. That range is wide because nobody actually knows, and the width of the range should itself be read as a warning: the industry is forecasting its own future electricity demand, and it has every incentive to forecast high, because a high forecast is what justifies the build. Take the midpoint, roughly 9%, and translate it into physical terms. The absolute increment is on the order of tens of gigawatts — the equivalent of dozens of large generating units, plus the transmission to move those electrons to where the campuses sit. You cannot summon tens of gigawatts with a deregulatory gesture. You have to build them, and building them runs into a supply chain that has been quietly starving for two years. The supply chain does not read press releases. Large power transformers — the iron-and-copper beasts that step voltage up and down between the grid and a campus — have seen lead times stretch to 80 to 120 weeks at the peak. That is not a shipping delay. That is a manufacturing-capacity delay: there are only a handful of qualified global suppliers, their order books are full, and adding a new production line for large transformers takes years, not months. Every week a transformer is late, a data center sits as a concrete shell with nothing to power it. It gets worse one layer up the stack. The gas turbines that would supply the fast, dispatchable power data centers love are booked out to the late 2020s by the major manufacturers. Combined-cycle and simple-cycle gas is the pragmatic near-term answer to intermittent renewables, and the pragmatic answer is sold out. High-voltage transmission lines — the ones that would carry power from where it is cheap to where the campuses are — routinely take five to ten years to permit and build. These are physical clocks. They do not accelerate because a politician is enthusiastic. Here is the insight most market commentary is missing: the binding constraint is not generation capacity. It is interconnect permission. This is where my years in this strange industry pay off, because crypto learned this lesson the hard way and no one outside it noticed. A bitcoin mine's most valuable asset was never the ASICs. Hardware depreciates. The moat was the power purchase agreement and the interconnect queue position — the legal right to draw a specific number of megawatts at a specific location for a specific period. When the hashrate economics turned brutal after 2022, the survivors were not the operators with the best machines. They were the operators who already held the wire. Now watch what happens when the AI boom collides with that asset. The bitcoin miners who spent 2019 through 2021 hoarding interconnect rights suddenly discovered that the same substation, the same land, the same permits they already held were precisely what a hyperscaler needed for an AI campus — and the hyperscaler could not wait five years to build the queue position from scratch. So the miners did what any rational holder of a scarce, non-reproducible asset does. They became landlords. They converted power sites into AI hosting facilities and signed long-term leases with the very cloud giants that were busy telling their shareholders they had everything under control. This is the quietest transfer of value in the current cycle: crypto mining capacity is being re-rated as AI real estate, and the re-rating has almost nothing to do with crypto. Think about what that means for a bear market. Here is a sector — proof-of-work mining — that was left for dead twice, that every serious analyst wrote off as a stranded asset, and that now sits on the single most contested resource in the developed world: an energizable, permitted, interconnected load. In a market where survival matters more than gains, the survivors in mining did not survive because they were better. They survived because they happened to be holding the one thing the AI economy cannot manufacture on demand. I should be precise about the limits of my own certainty here. When I reverse-engineered the Raptor Protocol contracts back in 2018 — forty hours of unpaid obsession that ended in a $2 million reentrancy exploit and a thesis that went viral for all the wrong reasons — I learned exactly how dangerous it is to confuse a compelling narrative with a durable position. The miners-as-AI-landlords story is compelling. It is also at the mercy of one variable I cannot fully audit: whether the AI demand actually materializes at the scale the leases assume. If it does, the miners win again. If it does not, they are holding empty concrete and the hyperscaler walks. Sentiment is a shifting tide, not a solid ground. The wire beneath it, though, is nailed to the earth. Let me turn to the compute layer itself, because this is where the crypto-native reader will recognize an old friend wearing a new coat. There is a growing ecosystem of "decentralized compute" and "DePIN" projects that promise to aggregate idle GPUs into a permissionless market for machine-learning workloads, and to pay contributors in tokens. The pitch is beautiful. It is also, in most cases, a familiar architecture. The coordinator — the entity that schedules jobs, sets prices, and decides who gets paid — is typically a small set of nodes, sometimes literally one, operating under the word "decentralized" the way a sequencer operates under the word "rollup." I have written about this before and I will write it again: the "decentralized sequencer" has been a PowerPoint slide for two years. Decentralized compute suffers the identical disease. The marketplace is permissionless; the metering and settlement are not. There is a sharper problem underneath, and it is the one I keep circling back to in my own work. These compute markets depend on a price feed — a live reference for what a GPU-hour of a given spec should cost, adjusted for utilization, latency, and reliability — and that feed is only as good as its oracle. Oracle feeds lag. They get front-run. They get gamed at the edges where liquidity is thin, and a compute-price oracle is exactly the kind of thin, manipulable market where a six-second latency window is a free option for an adversary. When I mapped 10,000 AI-agent interactions on-chain for my "Silent Market" research, the single most striking finding was not that agents transact autonomously. It was that roughly 70% of the observed transactions were micro-payments for data verification — tiny, machine-speed settlements settling through oracles whose latency was never designed for entities that decide in milliseconds. Code is law, but humans write the bugs — and agents exploit them at a speed no human auditor can match. Now the chip layer, quickly, because it is the most understood and therefore the least interesting part of the stack. Advanced training chips remain under the near-total dominance of one supplier. The in-house ASICs — Google's TPUs, Amazon's Trainium and Inferentia, Meta's MTIA — are real and improving, but they principally serve internal workloads. They relieve pressure on the buyer's own margin; they do not meaningfully loosen the external supply picture. Export controls remain a live policy variable, and their long-run effect is disputed: they slow one competitor while financing, through necessity, a cheaper path for another. I will return to that in the contrarian section, because it is the hinge on which the whole "America is far ahead" claim swings. Cooling deserves more attention than it gets. Rack power density has moved from the ten-kilowatt range toward a hundred kilowatts and beyond, and at those densities air cooling stops being a design choice and becomes a physics problem. Liquid cooling is going from niche to mandatory, which pulls in a whole supply chain of cold plates, manifolds, pumps, and CDUs that most equity analysts do not model. It also pulls in water — and water is where the story turns political again. In drought-prone states, the same campuses that promise tax revenue and construction jobs are presenting a bill measured in acre-feet, and the bill is being read aloud at county commission meetings by people who vote. This is the sentence that should be bolded in every AI infrastructure model: the social license to build is more fragile than the regulatory license to build. Now to the money, where the crypto tape and the AI tape share DNA. The four hyperscalers' combined capex sits at roughly 350 to 400 billion dollars, and the growth rate is still elevated. The bull case is straightforward: build compute, sell inference, collect rent. The bear case is a duration mismatch that any DeFi lender would recognize instantly. AI infrastructure depreciates over three to six years — that is the accountants' clock on a GPU cluster. Enterprise AI revenue, meanwhile, ramps along a curve that historically takes longer than the depreciation schedule of the asset that produces it. That gap is the whole story. When you finance a three-to-six-year asset against a revenue curve that has not yet arrived, you have built a duration bet, and duration bets get repriced when rates move or when the revenue just doesn't show. Yield is the bait, liquidity is the trap — and the trap in this cycle is called depreciation. There is a further wrinkle that anyone who survived 2022 will find unnerving. A meaningful share of AI demand is being financed through relationships that look circular. Chip suppliers invest in model developers; model developers buy compute from cloud providers; cloud providers buy chips from the chip suppliers; and the ventures that raise money from the first party spend it on the services of the second. In accounting terms, this inflates revenue. In economic terms, it dilutes signal — it makes it harder to tell whether the money represents genuine external demand or a closed financial loop breathing into its own lungs. I am not claiming fraud. I am claiming opacity, and opacity is the precursor to every repricing I have lived through since 2018. For the reader of a crypto publication, the practical mapping is this. The assets that benefit directly from the physical build are the boring ones: transformers, switchgear, gas turbines, grid equipment, fiber, optical modules, liquid cooling. The "AI plus energy" concept tokens that trade on narrative alone deserve the skepticism I reserve for NFTs that called themselves art when they were status auctions — because I conducted twenty interviews with Bored Ape collectors in 2021 and confirmed, on the record, that floor price tracked social signaling, not utility. Art without utility is just noise with a price tag, and the AI concepts with no compute revenue are the same phenomenon in a different wrapper. And the employment picture — the part of the story the golden-goose framing will not touch. Short term, the build is a construction boom: electricians, linemen, welders, HVAC technicians, at wage premiums. That is genuinely good work. The medium term is crueler. Customer service, junior coding, content production, paralegal document review — the roles that AI substitutes fastest — face displacement estimates that range from 20% to 60% depending on the task. The statement said nothing about worker transition. The absence of a word is itself a position. Now the counter-intuitive turn, and I want to aim it at the most comfortable sentence in the entire AI discourse. "America is far ahead of every other country." Read it twice, because the second reading is where it falls apart. In frontier closed models, in capital supply, in the installed base of advanced accelerators, in the sheer physical room to build — yes, the United States leads, and not narrowly. But "far ahead" is a political sentence wearing a scoreboard's clothes. Its function is to manufacture domestic consensus: to quiet the regulators, to bless the capex, and to bind voters to a national victory story. That is not an economic measurement. That is a campaign line. The place the claim corrodes fastest is open-source. Chinese open-weight models — I will not list them, they change monthly — have, on a number of public benchmarks, closed to within a hair of the American frontier, and they did it on dramatically less compute. That last clause is the one that should keep an American AI strategist awake, because it contradicts the foundational assumption of the entire capex boom — that leadership flows from owning the most iron. If a rival reaches near-parity at a fraction of the training cost, then the moat is not the cluster. The moat is the algorithm, and algorithms travel. This is the part where every bull run proves to be a myth waiting to be debunked. The 2018 cycle said the moat was the token. The 2021 cycle said the moat was the community. The AI cycle says the moat is the compute. Each time, a cheaper path dissolved the moat and the narrative repriced a quarter late. In the ledger's silence, the true story whispers — and the ledger here is silent about the one competitor the golden-goose speech never named. I want to be fair to the uncertainty, because I have been badly wrong before and I will be again. Export controls might genuinely slow a rival in ways I cannot see; the efficiency gap might be narrower than open-source benchmarks suggest once you account for the workloads that never get published; and a frontier gap of six months might be decisive in commercial terms even if it looks trivial on a leaderboard. Model rankings are quarterly objects. Anyone who tells you the gap with certainty is selling something. But the structural point stands. A speech that declares victory while omitting the rival is not analysis. It is a mood. And moods are the one input that has never once survived contact with a lower-cost competitor. So what do we actually hold, in a bear market, after the speech fades? Not the announcement. The queue. The interconnect position. The transformer on order. The liquid-cooling contract. The power purchase agreement signed with a nuclear operator. These are the assets whose value does not depend on whether the golden goose is, in fact, golden — only on whether the electricity to feed it can be delivered on time. And on the current physical timeline, delivered-on-time is the scarcest thing in the entire ecosystem. The miners already know this. They learned it while the rest of us were writing eulogies for their industry. They bought the wire. Now the wire is the story. The geese will keep quacking for the cameras. The question that decides the next four years is quieter, and it is asked at the substation: when the speeches are over, who actually holds the copper?