Amazon bought the chips. The electrons didn't show up.
That's the entire story compressed into one sentence, and it should terrify anyone who still believes the AI capex boom is a supply-chain problem waiting to be solved by more fabs. According to a brief circulating through crypto and tech aggregators, Amazon is sitting on billions of dollars of Nvidia hardware that cannot be switched on because the power isn't there yet. No specific figure. No timeline. No quoted executive. No filing anchor. Just a claim that, if accurate, marks the exact moment the AI infrastructure trade stopped being about silicon and started being about watts.
The constraint migrated. Almost nobody repriced their models to reflect it.
I've spent the better part of a decade auditing systems where the failure mode was always hidden in the layer nobody photographed for the press release. In 2017, I pulled apart forty ICO whitepapers in the middle of the Ethereum mania and found reentrancy bugs sitting in contracts that had already raised eight figures. The pattern repeats across every cycle: the crowd stares at the visible asset โ the token, the GPU, the headline โ while the actual bottleneck quietly relocates somewhere upstream. Right now, that relocation is complete. Nvidia allocation is no longer the binding constraint. Grid interconnection is.
Here is what the aggregate numbers look like when you run the math nobody in the press release bothered to run. An H100 SXM node, fully loaded with CPU, networking, and cooling overhead, draws somewhere between 1.0 and 1.25 kilowatts of facility load once you account for a power usage effectiveness ratio of 1.2 to 1.3. At a procurement price between $20,000 and $25,000 per card, a single billion dollars buys roughly forty thousand accelerators โ which translates to somewhere between forty and sixty megawatts of continuous draw. Scale that to the "billions" the report gestures at, call it three to ten billion, and you land at 100 to 500 megawatts. That is not a server room. That is a mid-sized data center campus sitting dark, depreciating, waiting on a wire.
And the wire is not coming quickly.
Large power transformers are quoting at 120 to 210 weeks. Some configurations stretch to three or four years. Gas turbine capacity from GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries is effectively sold out through 2027 and into 2028. The US interconnection queue sits at roughly 2,600 gigawatts of proposed capacity, and the median time from request to commercial operation has stretched from under two years in the 2000โ2007 window to approximately five years across 2018โ2023, per Lawrence Berkeley National Laboratory data. This is not a cyclical shortage you arbitrage away with a phone call. It is a physical queue with a physical length, and Amazon just got in line behind everyone else.
The demand side is accelerating into the same wall. US data center electricity consumption ran around 176 terawatt-hours in 2023 โ about 4.4 percent of national demand. LBNL projects 325 to 580 terawatt-hours by 2028. Globally, data centers consumed roughly 415 terawatt-hours in 2024, about 1.5 percent of world electricity, and the IEA expects that to reach approximately 945 terawatt-hours by 2030. Meanwhile rack density has jumped from a comfortable 10 to 15 kilowatts into the 120-kilowatt range for GB200 NVL72 deployments, which forces liquid cooling retrofits and high-voltage DC power distribution. Every one of those engineering changes adds months to the last hundred meters between substation and silicon.
So why would Amazon buy hardware it cannot power? Because in a shortage, buying the scarce thing early is rational even when the complementary asset lags. Nvidia allocation windows are short and politically mediated. Waiting for electrons before ordering chips means accepting a generation behind. The trade is deliberate: lock the accelerators now, absorb the delay, hope the grid catches up.
The problem is what happens if it doesn't. Hopper gave way to Blackwell, and Blackwell is giving way to Rubin on roughly a twelve-to-eighteen-month cadence. If the interconnection lags by more than a generation, you do not get delayed compute. You get stranded compute โ capital that powers on already obsolete. The pool remembers what the ticker forgets, and right now the ticker is pricing these assets as if they were humming at full utilization.
There is a quieter accounting dimension here that almost nobody is discussing. Amazon extended its server depreciation schedule from five years to six starting in 2024, a change that contributed a low-single-digit-billion positive swing to operating income. That assumption only holds if hardware is actually productive for six years. Idle assets still accrue depreciation on schedule. Extend the idle period far enough and the depreciation curve stops being a modeling choice and becomes an audit finding. This is the buried financial line item, and it is the one I would want to see in the next ten-Q before believing any of the bullish capacity narratives.
The genuinely counterintuitive conclusion is not that power is tight. Everyone has heard that. The counterintuitive conclusion is that the scarce asset has been repriced without anyone announcing the repricing. It is no longer the GPU. It is the powered shell โ a building with an executed interconnection agreement, energized switchgear, and cooling capacity already commissioned. Interconnection rights themselves have become tradeable, and the bundle of land plus power plus permitting is now the thing hyperscalers are actually bidding for. Watch what happens to valuations of operators holding energized capacity in PJM or ERCOT territory. That is where the real alpha sits, and it has nothing to do with which accelerator generation ships next quarter.
There is one mechanism working in the opposite direction, and the source material ignored it entirely. Power scarcity is the strongest possible incentive for efficiency innovation. Mixture-of-experts sparsity, aggressive inference optimization, distillation into smaller models, and custom ASICs all get dramatically better economics when a megawatt costs more than a rack. If effective compute per watt improves two to three times annually, a meaningful share of the projected deficit gets absorbed by efficiency rather than construction. Whether Jevons paradox swallows that gain or efficiency dividends outrun it is genuinely undecided, and anyone claiming certainty in either direction is selling something.
Volatility is the tax on uncertainty. Right now the uncertainty is not about demand โ demand is screaming. It is about whether the physical layer can be built fast enough to convert signed purchase orders into revenue before the silicon underneath them ages out. Amazon's parked hardware is the first public crack in that assumption, and cracks in infrastructure narratives tend to widen.
The question worth sitting with is not how many GPUs Amazon bought. It is how many of them will still be competitive when the breaker finally closes โ and who is holding the depreciation schedule when the answer arrives.