The Ledger Letter — AI Is Consuming Infrastructure Faster Than Capital Markets Are Pricing It
PJM capacity costs rose eleven-fold. The grid operator missed its own reliability target. Chip stocks set records.
The Ledger Letter
Finance Studio Advisors · Wednesday, June 4, 2026

AI Is Consuming Infrastructure Faster Than Capital Markets Are Pricing It

PJM Interconnection serves 67 million people across 13 states. It is the largest grid operator in the United States. In its most recent capacity auction, the cost of securing electricity supply cleared at $329 per megawatt-day, up from $29 two years ago. That is an eleven-fold increase driven almost entirely by data center expansion. And the operator still fell 6,623 megawatts short of its own reliability target. On Monday, chip stocks pushed the S&P 500 past 7,600 for the first time. On the same day, the Department of Energy launched a new platform to simulate AI campus power demands because the existing grid models cannot handle the load. The equity market sees AI demand. The grid sees the constraint. Both are correct, and they have not been reconciled.
The Breakdown
Today’s disagreement: equities are pricing an AI demand curve with no ceiling; the physical infrastructure underneath is pricing a supply constraint it cannot resolve in time.
01
The Auction That Priced the Bottleneck
PJM’s capacity auction for the 2026–2027 delivery year cleared at $329.17 per megawatt-day. Two years earlier, the same auction cleared at $28.92. The 2027–2028 auction hit $333.44, the maximum allowed under a temporary price cap. Without that cap, PJM estimates the price would have reached $530. Data center load accounted for 40 percent of the $16.4 billion in total auction costs.
02
The Shortage the Market Chose to Ignore
PJM fell 6,623 megawatts short of its 20 percent installed reserve margin target in the 2027–2028 auction. That is more generation capacity than a mid-size state uses at peak. The grid operator’s demand forecast jumped 5,400 megawatts year over year, driven overwhelmingly by data center interconnection requests. The queue is not shrinking. It is accelerating.
03
The Platform the DOE Had to Build
On Monday, the Department of Energy launched Agora, a test platform designed to simulate the electrical behavior of hyperscale AI campuses. The DOE built it because existing grid models cannot replicate the volatile, high-density power demands reshaping utility planning. When the federal government builds a simulation because the real system is too stressed to test safely, the bottleneck is no longer theoretical.
Partner Perspective

Dear Reader,

To avoid 67 million Americans losing power... President Trump is taking drastic action.

Using emergency executive powers, he has paved the way for a new technology 326 times more powerful than the most advanced power generators used by hospitals.

This tech could power the entire country without ever touching oil – and save our crumbling grid from collapse.

It’s now also backed by Nvidia CEO Jensen Huang, Elon Musk, and Sam Altman.

But the rights to this tech are cornered by companies most people have never heard of...

And virtually overnight, 3 little-known stocks could soar higher as the public learns how they’re involved.

Click here to see the little-known stocks at the heart of Trump’s $10 trillion plan for the American power grid.

Regards,

Joel Litman
Chief Investment Officer, Altimetry

The Cross-Asset Snapshot
PJM capacity price (2026/27)$329/MW-day — up from $29
PJM reliability shortfall (2027/28)6,623 MW — below target
Data center share of PJM costs40% — $6.5B of $16.4B
S&P 500 (Wednesday close)~7,553 — pulled back from record
Broadcom AI revenue (Q2)$10.8B — reported after close
10-yr Treasury yield~4.50% — rising
PJM auction data from FERC filings and Monitoring Analytics. Market data as of Wednesday close. Sources: PJM, DOE, CNBC, LSEG.

Eleven-Fold and the Question Nobody Is Modeling

The Cost Curve the Chip Rally Left Out

The wires this week led with chip stocks. Marvell surged. Hewlett Packard Enterprise beat. The S&P 500 crossed 7,600 for the first time. Behind that tape, the physical infrastructure required to run the AI buildout is repricing at a pace the equity market has not absorbed. PJM’s capacity auction results tell the story in a single number: $329 per megawatt-day, up from $29 two years ago. That is an eleven-fold increase in the cost of securing future power supply, driven almost entirely by data centers filing interconnection requests the grid was never designed to process at this volume.

The chip market is pricing AI demand. The power market is pricing AI constraint. They have not met in the middle.

The Shortfall That Is Not a Model. It Is a Deficit.

In this tape, PJM’s 2027–2028 capacity auction fell 6,623 megawatts short of its reliability target. That target exists for a reason: it models the amount of generation capacity the grid needs to prevent more than one unexpected outage every ten years. The shortfall is not a planning abstraction. It is a physical deficit in generation that has not been built, permitted, or contracted. And the demand side is not waiting. PJM’s load forecast jumped 5,400 megawatts in a single year, almost entirely from data centers.

Our view: when the largest grid operator in the country tells you it cannot secure enough generation to meet its own standard, and the equity market prices the demand driving that shortfall as a growth story, the tape is split. One side will reprice.

The Federal Signal Hiding in Plain Sight

Worth watching: the DOE’s Agora platform, launched this week, is the federal government’s admission that the grid cannot model AI loads with existing tools. The platform simulates volatile, high-density power demands from hyperscale campuses because real-world interconnection data is too unstable to plan around. This is not a research project. It is an emergency diagnostic.

The $700 billion in industry-wide AI capex announced this year assumes the megawatts exist. PJM’s auction says they do not. The gap between capex ambition and delivered power is the trade the equity market has not priced.

The Two Numbers That Settle This by Friday

Broadcom’s earnings, reported after Wednesday’s close, showed $10.8 billion in AI semiconductor revenue. That number confirms the demand side of the equation. Watch the stock’s reaction Thursday: if the market buys the demand story without pricing the infrastructure constraint, the AI trade is still running on the assumption that power will follow capital. Watch utility names tied to PJM’s footprint: if they begin to outperform chip stocks, the market is starting to price the bottleneck.

And watch Friday’s jobs data: if the 10-year pushes above 4.55 percent, the cost of financing every megawatt of new generation rises in real time, widening the gap between what AI needs and what the grid can deliver.

The equity market is pricing what AI can do. The grid is pricing what AI can’t plug in.
The Ledger Letter
When markets disagree, the signal is in the disagreement.
This newsletter is for informational purposes only and does not constitute investment advice. Past performance is not indicative of future results.

Recommended for you