Let me walk through the actual mechanism — the queue, the price signals, the equipment lead times, the legal levers — so you can tell the soundbite from the engineering.

Start with the wrong word

Every AI leader says "lack of power." Take that literally and it's false. The US has plenty of generation nameplate capacity on paper. What it lacks is deliverability — the ability to get electrons from where they're generated to where the load actually sits.

Aravind Srinivas, founder of Perplexity, framed it this way in a 20VC interview, and the framing stuck because it's diagnostic. If the problem were fuel, you'd build power plants. If the problem is deliverability, building plants does nothing — the electrons still can't reach the load.

So the real question isn't "how much power exists." It's "can this specific substation evacuate a 300MW load." Often the answer is no, and no amount of signed power contracts changes that.

The queue nobody talks about

Any large load — over 100MW — has to clear an interconnection queue run by the regional grid operator: PJM in the mid-Atlantic, ERCOT in Texas, MISO in the Midwest.

The sequence is fixed. Feasibility study, then System Impact Study, then Facilities Study, then a signed interconnection agreement. Each stage identifies network upgrades — new transformers, breakers, reconductored lines, sometimes entirely new 345kV or 500kV lines — and then everyone fights over who pays for them.

This is a three-to-five-year process at minimum. Cost-allocation disputes can send you back to the start. A signed power purchase agreement means nothing if the study says the substation can't evacuate the load.

Do this tonight: pull up your regional operator's public interconnection queue. PJM and ERCOT both publish them. Look at how many large-load projects are stuck in study versus in service. That ratio is the real bottleneck.

The price that proves it

If you want one number that settles the fuel-versus-deliverability question, it's PJM's capacity auction.

PJM's 2025/26 Base Residual Auction cleared at roughly $270 per MW-day. The prior 2024/25 auction cleared at roughly $29. That's close to a 10x jump in a single cycle. Treat these as illustrative reference figures: they're RTO-wide locational clearing prices, not numbers that apply to Northern Virginia alone. But the scarcity driving them is concentrated in that data center corridor.

Here's why that number is diagnostic. Capacity auctions don't price raw energy. They price guaranteed deliverable capacity at a location. So a large spike isn't the market screaming "we need more fuel." It's the market screaming "we can't guarantee moving what we have to where the load is."

Fuel-scarcity and deliverability-scarcity show up in different prices. This one is deliverability.

Why AI loads break the grid's assumptions

A steel mill draws a big, smooth, predictable load. An AI training cluster does not. This is the part most coverage misses.

A GB200 NVL72 rack draws on the order of 120 kW at the GPU-system level — that's the commonly cited figure for the compute hardware itself, not the full facility draw once you add cooling, distribution, and overhead. A 100MW hall is thousands of GPUs running synchronized collective operations — all-reduce steps where every chip waits for every other. When a training job checkpoints, crashes, or hits a communication stall, tens of MW can drop in milliseconds across the hall, all at once.

To the grid's protection systems, that reads like a fault — like a generator tripping offline. The grid was never designed for a load that behaves like a synchronized on/off switch at that scale.

The consequence is concrete. Utilities now write ramp-rate limits and low-voltage ride-through obligations directly into large-load interconnection agreements. And campuses install battery and flywheel buffering not for backup power, but for grid smoothing — to hide the millisecond swings from the utility. That's a genuinely new engineering problem, not a marketing one.

Why money doesn't solve it

The optimist's move is "capital dissolves the constraint — just build behind-the-meter." It doesn't, and here's the specific stack of why.

Even if you're handed a hot substation, energization is gated by hardware with structural backlogs.

Large power transformers: lead times commonly quoted in the 24-to-30-month range, though it varies by voltage class and vendor. Supply is largely import-dependent, so treat it as a moving figure, not a fixed law.

Medium-voltage switchgear: constrained.

Gas turbines: lead times for GE Vernova, Siemens Energy, and Mitsubishi H- and F-class frames reportedly stretch into the late 2020s, per industry reporting.

So "just put gas turbines behind the meter" hits two walls at once — the turbine backlog itself, and air permitting (New Source Review / PSD, Title V). Behind-the-meter generation doesn't route around the opposition. It hands them an emissions filing to fight over.

The honest counterpoint

Now the part that keeps this from being a soundbite. "Power is THE bottleneck" is an oversimplification, and it's worth saying why.

Go to Abu Dhabi or a Chinese hyperscaler and power is cheap and abundant. What binds there is silicon — specifically CoWoS-L advanced packaging capacity at TSMC and HBM3E memory allocation from SK Hynix, Micron, and Samsung, locked 12+ months out. A GPU without HBM is a paperweight.

And the binding chip constraint is a moving target. One quarter it's packaging, the next it's substrates, then it's 800G optical transceivers and switch ASICs for the scale-out fabric. So today, in most places, chips and packaging bind first. Power binds hardest.

Why power still wins the long game

Both things are true, so which one is the ceiling? It depends on the timescale, and this is the whole argument in one line.

Packaging is a 12-to-18-month capital problem. Transmission is a 10-to-15-year right-of-way, eminent-domain, FERC-jurisdictional problem.

You can double CoWoS capacity by spending money — TSMC is doing exactly that. You cannot buy your way past a NIMBY fight over a transmission line through six counties. There is no check you can write to skip a county commission's vote.

Call it capital-elasticity: how fast money converts into relief. Chips are high-elasticity. Transmission is near zero. That's why chips bind first and power binds last — and last is what caps the scaling curve.

"Community opposition" sounds vague, like a mood that PR can manage. It isn't. It's a set of concrete, dilatory legal levers, each one able to add 12 to 18 months.

Zoning moratoria at the county level — freeze siting outright.

Water-withdrawal permits — evaporative cooling towers consume real volume; in Arizona and Texas this is now a harder gate than power. It's why the industry's sudden pivot to "closed-loop, direct-to-chip cooling" is partly a permitting strategy.

Noise ordinances — generator load-bank testing and chiller noise.

Ratepayer-advocate interventions at the PUC — the big one.

None of these is emotional. Each is a filing with a comment period and a hearing. That's what makes backlash a machine you can predict rather than a wave you hope passes.

The battleground that decides everything: tariff design

If you want to know whether backlash gets worse, watch one thing at your state utility commission: large-load tariff design.

The question is simple to state. When a data center triggers network upgrades, who pays? If the PUC forces the hyperscaler to fully fund its own upgrades and post collateral, the project economics tighten and some deals die. If it doesn't, those upgrade costs get socialized onto residential ratepayers' bills — and that's the thing that turns diffuse annoyance into an organized political revolt.

Look at Virginia's PW Digital Gateway litigation, Georgia Power's rate cases, and Ireland's EirGrid connection freeze around Dublin. The fight isn't really about the building. It's about whose electricity bill goes up.

FAQ

Is power really the bottleneck to AI scaling?

Not first — chips and advanced packaging usually bind first today. But power and transmission have the longest lead time and the least ability to be fixed with money, so they cap the multi-year scaling curve.

Can't nuclear or SMRs fix this?

Not soon. SMRs are roughly a decade out and none are deployed at commercial scale in the US yet. And the binding delay is usually a permit or a PUC ruling, not the reactor.

Why can't hyperscalers just build behind-the-meter gas?

Turbine lead times reportedly stretching into the late 2020s, plus air permits (NSR/PSD, Title V) that become public-comment battlegrounds. It moves the fight; it doesn't end it.

What single indicator should I watch?

Capacity auction clearing prices and your state PUC's large-load tariff dockets. Those two tell you deliverability scarcity and cost-shift risk directly.

Do these three things now

Open your regional grid operator's interconnection queue (PJM or ERCOT) and count large-load projects stuck in study versus in service. That ratio is the real bottleneck, not chip supply.

Look up your state PUC's large-load tariff docket and read who's arguing about cost allocation. That tells you where backlash goes next.

Next time you hear "power is the bottleneck," ask the follow-up: fuel, or deliverability? The answer determines whether the fix is a power plant or a transmission line — and those are a decade apart.

One open question worth arguing about: over the next ten years, which one actually stops a given AI campus first — the packaging queue or the transmission line? It probably depends entirely on where you build. Where do you think it binds hardest?