There is a line in the current AI capex story that nobody has priced, and it is not the one about demand.
Amazon is holding billions of dollars of Nvidia hardware it cannot switch on. Not because the silicon is late. Because the electricity is not there.
I read the source report four times looking for a publication date. There is none. No named executive, no earnings call anchor, no filing reference. Four sentences, two of them opinion. That is the entire evidentiary base, and any analyst who skips past it is building on sand.
My working rule: a claim about capital allocation with no date and no citation is a paraphrase until proven otherwise. This one tracks almost exactly to Amazon CEO Andy Jassy's public remarks from late 2024 โ the ones where he said the company had secured a large volume of Nvidia chips and that power, not silicon, was the binding constraint.
If that reading is right, "billions" has already been diluted by two rounds of transcription. The real figure is larger. It always is.
This is a brief about a constraint, not a company. The company is simply where the constraint surfaced first.
The 2023 narrative was chips. Allocation quotas. Who gets H100s, who waits, who pays over list on the secondary market.
That narrative is dead. The market has not fully repriced the corpse.
Chip supply loosened through 2024 as TSMC capacity scaled and Nvidia's allocation spread across a wider buyer base. Power did not loosen. Power is physical, permitted, and slow, and none of those three properties respond to a term sheet.
The report gives me zero quantitative data, which is itself worth naming as a finding. So here is the arithmetic I run, flagged as mine rather than the article's.
An H100 SXM node, fully loaded with CPU, networking and cooling at a PUE of 1.2 to 1.3, draws roughly 1.0 to 1.25 kW of facility load. At $20โ25k per card, $1 billion buys about 40,000 cards and about 40 to 60 MW. Scale that to the $3โ10 billion range the phrasing implies and you land at 100 to 500 MW.
That is not a rack. That is a campus. In some cases it is the largest single load in a regional grid operator's territory.
Liquidity doesn't wait for a transformer. Capital sitting in idle iron earns nothing while still depreciating on schedule.
Why should a crypto desk care? Because the marginal buyer of stranded power is no longer only a hyperscaler. It is a token issuer with a compute narrative, a restaking protocol selling security to operators, an HPC conversion deal priced as a yield product. The grid is now part of the collateral stack whether or not anyone wrote it into a whitepaper.
Now the supply side, which is where this story actually lives.
Large power transformers quote at 120 to 210 weeks. Some bids run three to four years. GE Vernova, Siemens Energy and Mitsubishi Heavy are effectively sold out on gas turbines into 2027 and 2028. The US interconnection queue holds roughly 2,600 GW of requests, and per LBNL the median time from request to commercial operation moved from under two years in the 2000โ2007 window to about five years for the 2018โ2023 cohort.
Five years. The accelerator refresh cycle is twelve to eighteen months.
Run that collision and the economics get ugly fast. Hopper to Blackwell to Rubin. If your power lands two generations late, you have financed a building full of parts that were flagship products at order date and mid-tier silicon at rack date. The capex is real. The revenue curve slips six to twenty-four months. That gap does not appear in a headline about AI demand. It appears in a depreciation schedule.
Which is the line item nobody is discussing.
Amazon extended the useful life of its servers from five years to six starting in 2024. That change contributed roughly $3 billion to 2024 operating income. It is a defensible choice under a stable technology cadence. It is a considerably more aggressive choice when a material share of your fleet is sitting in a warehouse waiting on a substation.
I don't care what a company says about demand. I care what assumption it books.
Extend useful life, defer depreciation, smooth earnings. Now stretch the idle period. The assumption and the reality diverge, and the divergence compounds quietly until someone restates it. That is the under-discussed financial spine of this entire buildout, and it does not stop at Amazon's fence line.
Demand is not in dispute. US data center electricity consumption was about 176 TWh in 2023, roughly 4.4% of national load. LBNL projects 325 to 580 TWh by 2028. Globally, data centers drew about 415 TWh in 2024, around 1.5% of world electricity, with the IEA projecting roughly 945 TWh by 2030.
Those are not forecasts from a pitch deck. They are grid planning inputs. Grids plan in decades.
The engineering layer makes it worse. Rack density moved from the traditional 10 to 15 kW to roughly 120 kW for a GB200 NVL72 rack. That is not a marginal shift. It forces liquid cooling, high-voltage DC distribution and a full mechanical retrofit. The last hundred meters, substation to rack, became the hardest part of the build.
So what is actually scarce?
Not GPUs. The scarce asset is the powered shell โ a building with energized, permitted, grid-connected capacity.
That repricing is already visible in three places, and this is where the crypto market should be paying attention rather than letting AI headlines scroll past.
First, existing data centers with live interconnection rights. They are no longer real estate. They are options on a queue position, and the queue is five years deep.
Second, interconnection rights themselves. In several ISOs these are transferable. A transferable queue position is a tradeable instrument with no ticker and no exchange listing. Some of the HPC conversion deals miners signed in 2024 and 2025 are effectively sales of that instrument dressed as hosting contracts. The market read them as yield deals. They were queue arbitrage.
Third, the bundle: land plus power plus permits. That package now prices above the silicon it will eventually host.
Bitcoin miners understood this first because they had no choice. They spent a decade negotiating interruptible load with utilities nobody else wanted to talk to. When the HPC buyers arrived with hyperscaler balance sheets, the miners discovered their most valuable asset had never been the hashrate. It was the energized fence line.
Watch the curtailment data. In several US markets, large loads are being signed under interruptible terms, which means the operator absorbs the downside when the grid tightens. That is hidden volatility exposure sitting inside a supposedly passive infrastructure asset. Negative pricing events are the tell. When power goes negative, someone is being paid to not consume, and that someone is increasingly a compute operator holding a contract it cannot reprice.
And treat the disclosed figure as a floor. Against Amazon's hundred-billion-dollar annual capex, the portion a communications team can safely describe as "waiting on power" is almost certainly the smallest defensible number. The real idle inventory is a multiple of it.
Which is why I am unimpressed by the decentralized compute narrative currently trading at a premium.
DePIN GPU marketplaces aggregate idle capacity. Most of that capacity sits behind residential or small commercial interconnects with no path to sustained high-density load. Aggregating ten thousand scattered cards does not produce a 120 kW rack. It produces a coordination problem, and coordination problems in permissionless systems have a long history of resolving badly.
I watched that failure mode up close in 2024, digging through EigenLayer's slashing conditions ahead of a restaking allocation. The attack surface was not in the cryptography. It was in operator coordination โ the assumption that a set of independent actors would not converge on a jointly profitable misbehavior. Grid interconnection has the same structure. A queue is a coordination mechanism. Everyone benefits from holding position, and everyone loses when the queue becomes the trade.
I spent weeks in early 2026 auditing autonomous AI trading agents, building a small open-source tool to trace their on-chain transaction patterns. The finding that stuck was not about model quality. It was that most agents had no disciplined key management, no spend caps, no circuit breakers โ full-throttle execution against infrastructure with no brake pedal. Inference demand has the same shape at a larger scale. Enormous capability, deployed against a base layer never designed for this load, with nobody accountable for the gap.
Here is where the consensus read is incomplete.
The standard conclusion from power scarcity is that compute stays expensive and incumbents with grid access win. Reasonable. Also lazy.
The counter-force is efficiency. Power scarcity is the strongest possible incentive to make each megawatt produce more tokens. Mixture-of-experts sparsity, aggressive inference optimization, distillation into smaller task-specific models and custom silicon all get economically amplified when the marginal megawatt is the binding constraint.
If effective compute per watt improves two to three times a year โ and the last two years suggest that pace is not fantasy โ a meaningful slice of the projected shortfall gets absorbed before it is ever billed to a customer.
Jevons says cheaper inference expands demand faster than efficiency saves supply. The efficiency dividend says the opposite. Nobody has settled it, and anyone who tells you the answer with confidence is selling something.
Second blind spot: generational risk. Buying chips before power is rational inside a shortage, because Nvidia allocation was the scarcer resource and waiting for a substation means accepting last-generation silicon. That logic inverts the moment the wait exceeds one refresh cycle. Then you have optimized for allocation and lost on obsolescence.
Most operators are running the first playbook while living in the second regime.
What would falsify this? A transformer lead-time collapse below 60 weeks, or interconnection reform that compresses the queue under two years. Neither sits in any credible pipeline. Plan accordingly.
Three numbers to watch, none of which appear in this report.
The generation mix of the idle hardware. H100, H200 or GB200. Six months of waiting and twenty-four months of waiting are different businesses with different residual values.
The ISO. PJM, ERCOT, Oregon, Ohio. Interconnection timelines and capacity prices differ by multiples across those markets, and the regional answer determines whether the write-down is a rounding error or a restatement.
And the depreciation footnote in the next 10-K. If server useful life moves back toward five years while idle inventory climbs, the smoothing is finished.