Super Micro missed its revenue number this quarter. The stock went up 19% anyway — and almost nobody stopped to read why it missed.
The company reported $11.12 billion in Q4 net sales against roughly $11.55 billion expected. CEO Charles Liang explained the shortfall plainly: some customer projects were delayed by power, cooling and networking readiness. Not weak demand. Not cancelled orders. The servers were built. The customers wanted them. The buildings couldn’t run them.
That single sentence is the most important thing said in AI earnings this month, and it reframes the entire trade. The AI power bottleneck has quietly replaced chip supply as the binding constraint on the whole industry — and the market is still pricing the old constraint.
The market read the backlog and ignored the reason
Look at what actually happened on 12 August. CoreWeave reported Q2 revenue of $2.6 billion, up 112% year over year, with a revenue backlog of $104 billion. Super Micro booked more than $60 billion in new orders in a single quarter and guided fiscal 2027 to $65–72 billion. Nebius jumped 34%. Lumentum rose 14%. The S&P 500 hit a record.
The consensus read was straightforward: the AI bears were wrong, demand is real, buy the infrastructure layer.
But demand was never seriously in question. What those results actually revealed is a widening gap between what has been ordered and what can be energised. CoreWeave’s $104 billion backlog against 2026 revenue guidance of $12.4–13.2 billion is not primarily a sales achievement — it is an eight-year queue. Backlog that large stops being a demand signal and becomes a delivery-capacity signal. And the thing throttling delivery is not silicon.
The bottleneck moved down the stack, from chips to substations
For three years the scarce input in AI was the GPU. That story is over. The scarce inputs now are interconnection, transformers and firm generation — and each has a lead time measured in years, not quarters.
- The average time from interconnection request to commercial operation in the US has stretched to roughly five years, up from under two years in 2008.
- Around 2,300 GW of generation and storage sits stuck in US interconnection queues — more than the country’s entire installed capacity.
- Dominion alone reports roughly 70 GW of large-load interconnection requests against an all-time peak demand of 24 GW.
- High-voltage transformers carry lead times of about four years.
- AI racks draw 30–100+ kW versus 5–15 kW for conventional racks.
Now put the capital against that. Amazon, Alphabet, Microsoft and Meta plan roughly $725 billion of 2026 capex between them, up about 77% year over year. That money is being committed on quarterly earnings-call timescales into a physical system that responds on five-year timescales. Something has to give, and it will not be physics.
What a $329 megawatt-day does to AI returns
Here is where it hits the P&L. PJM’s capacity auction cleared at $28.92 per megawatt-day for the 2024/25 delivery year. For 2026/27 it cleared at $329.17 — an increase of more than 1,000% in two delivery years. Data centres were responsible for 63% of the increase in the 2025/26 auction.
The contrarian point is this: rising power costs do not kill the AI trade. They redistribute who earns from it.
Think about who sits where. A hyperscaler that has locked in long-term generation contracts, owns its substations, or has struck nuclear and renewable offtake deals converts scarce power into a moat. A neocloud that signed fixed-price, multi-year compute contracts and buys power at merchant rates has just written a short option on electricity prices. Same revenue line, opposite exposure.
This also complicates the fashionable bear case about depreciation. Much has been written — including by Michael Burry, who estimates hyperscaler depreciation will be understated by roughly $176 billion between 2026 and 2028 — about the industry stretching server useful life from three or four years to six. If power is genuinely scarce, older GPUs stay in service and stay utilised, which arguably supports the longer schedules. The real margin risk isn’t that the hardware becomes worthless. It’s that every incremental unit of compute costs more to run than the last. Depreciation is an accounting argument. Electricity is a cash one.
Regular readers will recognise the pattern from our look at how AI infrastructure is actually being financed — the debt was the first thing that migrated away from the companies whose logos are on the trade. Power economics is the second.
The bill is landing on people who never bought a GPU
This is where Wealtharian will be blunt. Roughly $9.3 billion of PJM capacity costs from the 2025/26 auction gets recovered from customers across the region. PJM expects bill increases of about 1.5% to 5% depending on state and utility. Pepco residential customers in Washington DC saw average bills rise by around $21 a month.
AI infrastructure creates real value — better diagnostics, cheaper drug discovery, genuine productivity gains. That is worth building. But the way the cost is currently allocated, a household in Maryland is subsidising the compute margin of a trillion-dollar company through its utility bill. That is not value creation. That is cost transfer, and it looks a lot like the K-shaped split we wrote about earlier this month: asset owners capture the upside, wage earners absorb the input costs.
Treat this as a live investment risk, not just an injustice. Cost transfers of this size and visibility invite rate cases, data-centre-specific tariffs, and state legislation. Several regulators are already moving. Any AI capex model that assumes today’s power cost structure holds through 2030 is assuming away the politics.
How to actually position
Three practical implications, none of which require you to guess which model wins.
1. The marginal AI dollar increasingly buys infrastructure, not intelligence. Grid equipment, high-voltage transformers, electrical engineering and construction, cooling, and independent power producers with interconnected capacity are all upstream of every hyperscaler’s plan. They are also, on the whole, less crowded than semis.
2. Ask a different question of every AI holding you own. Not “is demand strong” — it plainly is. Ask: does this company control its power cost, or does it buy at auction? That single distinction will separate winners from losers in this cohort more reliably than revenue growth over the next three years.
3. Check your concentration before you add. If you own an S&P 500 index fund, you already have substantial AI capex exposure, and the gap between index-level and mega-cap drawdowns shows how concentrated that has become. Adding a thematic AI position on top is often doubling down, not diversifying.
The honest counterargument
The strongest case against all of this: capital is extraordinarily good at solving supply problems when the returns justify it. Microsoft and Google are already funding dedicated generation through partnerships with Brookfield and others. Behind-the-meter gas, restarted nuclear, and on-site generation can bypass the interconnection queue entirely. If hyperscalers simply build their own power, the constraint loosens faster than the queue statistics suggest — and the utilities and IPPs never capture the scarcity rent at all.
That is a real possibility, and it is the main reason to own the picks-and-shovels layer (turbines, transformers, switchgear, engineering) rather than betting narrowly on regulated utilities. The equipment gets sold either way.
But note what even the bull case concedes: the money now goes to building power, not buying chips. That is the repricing, and it has barely started.
The number to watch
Ignore the backlog headlines next quarter. Watch the gap between orders booked and revenue recognised, and watch how many management teams blame “power, cooling and networking readiness.” Super Micro just told you where the constraint is. It missed a quarter to prove it, and the market gave it a 19% round of applause for the backlog instead.
Demand is not the story any more. Delivery is.
Want to track your own path to financial independence? The Wealtharian Wealth Tracker lets you monitor your net worth, FU money progress, and investment milestones in one place. Try it free →