AI Is Becoming a Power-Plant Trade - 8/7

Morning Watchlist: Friday Edition             

A quick note from Behind the Markets

Wall Street still wants to sell you AI as a software story.

The power grid disagrees.

The Pentagon is discovering that a missile order is only as good as the supplier who can actually deliver the motor, the nozzle, and the connector. Private credit is finding out what happens when "yield" meets a real refinancing wall. And the FDA just let a decision deadline come and go without saying a word — which tells you something about how much certainty a calendar date really buys you.

Here is the thread running through all four stories this morning: every popular market narrative eventually runs into a physical constraint. Electrons. Machine tools. Borrowers who can refinance. Regulators who accept the evidence.

The opportunity is in the bottlenecks, not the slogans. Let's find them.

1) AI Is Becoming a Power-Plant Trade

The numbers are stark. Cumulative U.S. grid power supply to data centers is forecast to fall more than 50 gigawatts short of demand by 2030. The mismatch that matters: a data center can be built in 18 to 24 months. Securing a grid connection takes three to seven years.

Hyperscalers cannot wait three to seven years. Their capital expenditure announcements jumped more than 50% in early 2026 alone, and the utility sector has not caught up.

That mismatch is turning "bring your own power" from a niche engineering choice into a competitive advantage. On-site generation, gas turbines, backup systems, transformers, switchgear, cooling, grid-modernization equipment — all of it becomes part of the AI stack. Small and mid-sized industrial suppliers benefit because they sell the physical components that hyperscalers cannot conjure with a software update.

As of the end of 2025, more than 2,060 gigawatts of generation and storage capacity were sitting in U.S. interconnection queues — roughly double the entire existing U.S. power plant fleet. Sounds like a boom. Now look at what happens to those requests: of the capacity that entered the queue between 2000 and 2019, only 13% had reached commercial operation by the end of 2024. Seventy-seven percent was withdrawn.

In Texas it is even starker. ERCOT's large-load queue ballooned to 226 gigawatts — about 2.6 times the state's all-time peak demand — and as of late 2025, only about 1.8% of it was actually operational and drawing power. The industry has a name for the rest: phantom loads.

So the hard questions are utilization, permitting, fuel access, customer concentration, and who pays for the capital. A data center with a power contract is not the same thing as a data center with dependable electrons. And a supplier with a big order is not necessarily a supplier with a good return on invested capital.

The Stock: Powell Industries (NASDAQ: POWL)

Powell is a Houston company that has been making custom-engineered electrical gear since 1947 — switchgear, control rooms, bus duct, the unglamorous steel-and-copper equipment that sits between a power source and a load. It is exactly the kind of physical bottleneck supplier this theme runs through.

On Monday it reported fiscal third-quarter results, and the order book was extraordinary: a record $934 million in new orders, up 158% year over year, producing a book-to-bill ratio of 3.0x. Backlog climbed to $2.4 billion, up 69% from a year ago and up 35% in a single quarter. That backlog now sits at roughly 7.7 times the most recent quarter's revenue. The quarter included three mega-orders: a data center project exceeding $400 million — the largest single contract in company history — plus a roughly $75 million petrochemical order and a $60 million LNG order.

The balance sheet is genuinely strong. Powell finished the quarter with $633.6 million in cash and short-term investments, $606.5 million of working capital, and no debt. Gross margin held at 30.6%.

Now the other side, and it is substantial.

Powell missed. Revenue of $311.7 million came in about 1.5% below consensus, and adjusted earnings of $1.42 a share missed the $1.47 estimate by 3.4%. That is now three consecutive quarters coming in below consensus. Revenue grew 9% while orders grew 158% — meaning almost the entire story is in a backlog that has not yet converted to revenue. Management itself cautions that cancellations, scope reductions, or delays could affect conversion.

Concentration is real: those three mega-orders alone were at least $535 million, or 57% of total bookings for the quarter. Petrochemical revenue fell 49% year over year and the Canadian market softened. The company is expanding its manufacturing footprint by more than 20%, which means capital going out the door ahead of revenue coming in. Copper, aluminum, and steel inflation are live headwinds.

And the stock has been violent. It fell on the earnings report, is down roughly 34% over the past 90 days, and sits well below its 52-week high of $336 — yet it is still up about 79% year to date with a one-year total return near 168%. A stock that can fall a third and still be up that much is a stock carrying enormous expectations. Analyst coverage is also thin — only three or four firms publish on it — so the "consensus" is a small sample. Insiders have been sellers this year.

Bottom line: AI is moving from a cloud-capex story to an electricity-and-equipment story. Powell is a real business with a record book and a fortress balance sheet — and a market that has already priced in a great deal of it, three straight misses, and a backlog it now has to actually execute. Follow the companies that remove the physical bottleneck, then check whether the cash flow survives the buildout.

2) The Pentagon's Missile Ramp Is Really a Supplier Audit

The headline is a multiyear missile-production push. The investable detail is the supplier map.

On Monday, Northrop Grumman signed two multiyear framework agreements worth a combined $3 billion — $2 billion covering PAC-3 MSE solid rocket motors and ignition safety devices, and $1 billion covering THAAD components over seven years. Northrop plans to triple tactical solid rocket motor capability at its Allegany Ballistics Laboratory in West Virginia by 2027, supporting an Army plan to lift annual PAC-3 MSE production from roughly 600 units today into the thousands.

Here is the detail most coverage buried: this deal makes Northrop a second source. L3Harris is currently the sole producer of solid rocket motors for the PAC-3 interceptor — and L3Harris signed its own similar framework agreement with the Pentagon the week before. When the Defense Department pays to stand up a second supplier for a critical component, it is telling you the single-source chokepoint was real.

The Pentagon has signed roughly a dozen framework agreements with defense companies since January, and officials describe the approach as working with suppliers at every level rather than primes alone — using multiyear commitments to entice firms to invest their own capital in capacity.

That matters because the defense industrial base is not a vending machine. A survey of companies across the supplier chain found the base is not prepared to execute a reported $900 billion backlog without work on resilience, financing, technology, and capacity. The problem is not only the prime contractor. It is the machine shop, the propellant supplier, the nozzle producer, the test house, the specialty-materials plant, and the quality-control system underneath it.

So resist the reflex to buy the biggest defense name after every contract announcement. Look for suppliers with qualification barriers, recurring sustainment work, scarce manufacturing know-how, and enough balance-sheet strength to fund tooling before reimbursement arrives.

What "Good" Looks Like — And What It Costs: Moog (NYSE: MOG.A)

Moog makes precision motion and actuation systems that get designed into missiles, military aircraft, and space platforms. Once its hardware is qualified onto a program, switching costs are punishing — which is the textbook version of the "qualification barrier", plus decades of recurring sustainment work behind it.

The business is performing. Moog reported record fiscal third-quarter results on July 31 — net sales up 15% year over year to $1.1 billion, adjusted operating margin expanding 280 basis points to 16.4%, free cash flow of $133 million, and a twelve-month backlog up 23% to $3.3 billion. It raised full-year 2026 guidance and declared a $0.30 quarterly dividend.

And here is why you should put it on your watchlist rather than buy right now. The stock is up roughly 127% over the past year and trades around $417 — at or slightly above the average analyst price target, which sits in the $335 to $415 range depending on whose panel you use. When a stock trades through the consensus target, you are no longer being paid to be early. You are paying up for a story the market has fully absorbed. Coverage is also thin at about five firms, and one screen flags it as technically overbought.

Moog is what a good sub-tier supplier looks like. It is not currently what a cheap one looks like. Note it, watch it, and let volatility do you a favor before you act.

3) Private Credit Just Printed a 6% Warning

Private credit is still marketed as a calm, floating-rate income product. The latest data is less calm.

Fitch's U.S. private-credit default rate reached 6.0% over the twelve months through June, up from 5.7% in the prior quarter. Some context the headline number hides: this is not a fresh spike. The rate first touched 6.0% back in April, a record since Fitch launched the series in August 2024, and it has been sitting at that level since. This is a plateau at the highs, not a blip.

The detail to watch is not the percentage. It is the mechanism. More than half of the quarter's default events involved maturity extensions rather than payment defaults or rate deferrals — a pattern that has now persisted for months, with extensions outpacing every other default scenario. In one recent monthly read, five of seven extensions pushed maturities out by one to two years.

That is a market buying time, not clearing bad credits.

And the sector data cut against the popular narrative. Health care default rates rose, and industrials and manufacturing posted the highest rates among the largest sectors. Technology software — the sector everyone assumes AI is about to disrupt into oblivion — showed the lowest stress of the major categories at 1.2%, down from 2.3%. Whatever is breaking in private credit, it is not the thing the headlines told you would break.

There is also a measurement problem you should understand. Moody's estimates that distressed restructurings — debt exchanges and maturity extensions agreed under duress — accounted for roughly 65% of all 2025 private credit defaults. Strip those out and headline rates fall to somewhere between 1.6% and 4.7%. Include them and the picture is materially worse. Two honest analysts can hand you very different numbers for the same market depending on where they draw that line. Know which number you are being shown.

Meanwhile, retail-focused funds face elevated redemption requests — a liquidity problem in a market built around loans that are hard to sell quickly.

This does not mean every lender is broken. It means the manager matters. Underwrite non-accruals, payment-in-kind income, amendment activity, leverage, borrower concentration, and the gap between stated net asset value and a real clearing price. A high distribution can be compensation for risk — or it can be a dividend funded from a balance sheet that is quietly deteriorating.

No stock idea in this section, and that is deliberate. We're not going to hand you a lender to buy in the same breath that we tell you the disclosure quality across the sector is the entire problem. When the honest answer to "which one?" is "I cannot verify that from the outside," the professional move is to say so rather than fill the slot. We will revisit this lane when the marks are clearer.

Bottom line: Private credit is no longer a clip-the-coupon trade. The first question is whether the borrower can refinance. The second is whether the lender is marking reality.

4) The FDA's Clock Ran Out — And Nothing Happened

Here is the setup. Replimune's RP1 — an oncolytic immunotherapy given with Bristol Myers Squibb's Opdivo for advanced melanoma that has progressed after anti-PD-1 therapy — has been through a brutal regulatory gauntlet. Breakthrough therapy designation in November 2024. A complete response letter in July 2025. A second complete response letter in April 2026, both citing the single-arm design of the IGNYTE trial. A third resubmission.

On July 30 the FDA's Cellular, Tissue, and Gene Therapies Advisory Committee voted 10 to 3 that IGNYTE's efficacy results were evaluable and clinically meaningful. The trial enrolled 140 patients and produced a 34% response rate with a median duration of response of 24.8 months. The stock nearly doubled in a day.

The target action date was Sunday, August 2.

That date came and went. The FDA did not act. As of this week the decision remains pending, with no explanation offered. Retail sentiment shifted from extremely bullish to neutral, and the theories range from a third rejection to a negotiation over labeling. Leerink took the optimistic read, upgrading to Outperform and lifting its target to $17 from $11, arguing the agency is now "boxed into approval." Other firms upgraded after the panel too — Piper Sandler to Buy at $14, Wedbush to Buy at $12, BMO reiterating at $16. Shares closed Wednesday around $11.77, roughly flat, on volume well below average.

Sit with that. An advisory panel voted overwhelmingly in the company's favor. Analysts upgraded. And four days past the deadline, holders know exactly as much as they did before — which is nothing.

That is the real lesson. An advisory vote is not an approval. A deadline is not a decision. A favorable outcome would still leave commercialization costs, label limitations, competition, and financing needs untouched. And a negative one can gut the stock even where the underlying science stays interesting.

Replimune posted a net loss of roughly $73 million last quarter against about $209 million in cash and $269 million in total liquidity. Its auditor has flagged going-concern doubt. The company has proposed doubling its authorized share count to 300 million — which speaks directly to dilution. A securities class action filed after the first rejection letter is still being litigated. The 52-week range runs from $1.50 to $13.86.

So this is a case study, not a stock idea. If you already own it, know that you are holding an unhedged binary in a company that needs capital regardless of the outcome. If you do not, understand that "the deadline is Sunday" turned out to mean nothing at all.

The retail mistake is to treat a catalyst as a thesis. The better framework: size for the binary, read the FDA documents instead of the social-media headlines, and model what happens if approval is delayed or restricted — because delay is exactly what happened.

Bottom line: The FDA calendar can create opportunity. It does not create certainty. Treat binary biotech as risk capital, not as a lottery ticket with a lab coat.

Before You Go

Every popular market story eventually runs into a physical constraint.

AI needs power — and 77% of the projects that queue up for a grid connection never get built. Missiles need qualified suppliers — and the Pentagon is now paying to stand up second sources because the first ones were chokepoints. Private credit needs borrowers who can refinance — and more than half of last quarter's defaults were extensions, not resolutions. Biotech needs regulators who accept the evidence — and this week a regulator simply declined to answer on time.

That is the work: find the constraint, identify who gets paid to solve it, and then check whether the balance sheet can carry the wait.