Evening Analysis — 2026-09-16

Dek

Two doors closed today — one on nuclear's cheap capital, one on cheap power for data centers — and both were opened by the same force the grid operator just put a number on.

The Big Shift

Holtec, a nuclear services firm and small-reactor developer, pulled its US IPO the day before it was set to price (¹). Pulling a deal at the eleventh hour means the book wasn't there at any price the company would accept — that's a buyers' strike, not a scheduling problem. It matters because the entire "new reactors create a fuel demand wave in the 2030s" story assumes builders can raise money now, and this is the first hard evidence they can't on the terms they expected. Watch whether a second nuclear issuer pulls or prices badly in the next two weeks; one withdrawal is a company story, two is a sector repricing and a real hit to the LEU thesis.

Analysis

Power scarcity stopped being a forecast and became an admission. PJM — the grid operator for 65 million people from Chicago to DC — modeled loss-of-load running more than 100x its own planning standard by 2030 if data center growth continues, which works out to 13+ days a year of not having enough electricity (²). Grid operators do not publish numbers like that casually. For CEG this is the thesis simply working: firm, always-on generation gets scarcer and dearer, and capacity prices follow. It's confirmation, not news — already priced into a 0.95 conviction — so it moves nothing tonight. The useful read is what it licenses politically.

And politics moved the same day. The House passed a bill to shield households from data center power costs (³). The backlash has climbed from county zoning meetings to federal law in about a year — and you don't legislate over *who pays* unless the thing is genuinely scarce. The second-order effect is the one that matters for Own-the-Bottleneck: if data centers can't socialize their power bills onto ratepayers, building your own generation on-site starts looking cheap by comparison. Cost-shift rules push hyperscalers toward behind-the-meter deals, which is the bottleneck thesis expressed through a different channel. Local resistance keeps supplying fuel — Philadelphia is now fighting a proposed campus in a neighborhood already scarred by a dead oil refinery (⁴), and Heatmap is already naming the midterm races this will decide (⁵).

The tape is bullish for the wrong reason. Lumentum and Coherent — optics companies that make the laser links between AI servers — were the S&P's biggest gainers, and the stated reason was investors hunting "further down the supply chain" because they fear a spending slowdown at the top (⁶). That's defensive rotation dressed as a rally. VRT's second pillar is that its order book tracks hyperscaler capex, and it trades at a premium that assumes that capex keeps compounding. Pair it with the recurring Semafor line that *slowing* AI development would help hyperscaler balance sheets, and you have two independent items pointing the same direction: the market is quietly rehearsing a capex pause. Conviction down. Note the honest tension — the same day's PJM report says physical demand is overwhelming the grid. Demand for power and willingness to fund the next capex cycle are not the same variable, and they can diverge for a year.

Capital is the common thread, and it's getting selective. Holtec can't get an IPO done. Nuclear paper is being repriced while SMR costs still haven't come down — Big Tech keeps signing anyway (⁷), which tells you the buyers are strategic, not financial. Uranium miners are spending and drilling, but development timelines mean supply arrives late regardless (⁸). Meanwhile money for the boring middle of the stack is abundant: Serverfarm added $895M to a now-$3.89B development fund, Goodman raised $455M for Hong Kong. Investors will fund concrete and racks on ten-year leases. They will not fund reactor developers at 2030s multiples. That split is the story of the day.

Materials: the moat is being surveyed, not yet crossed. Critical Metals claims it cracked eudialyte — a rare earth ore that's notoriously hard to process, which is why nobody bothered — and plans a $2.2B/yr refinery in Romania fed by its Greenland deposit (⁹). This is exactly MP's named thesis-breaker: a second Western mine-to-magnet chain. It's a preliminary study, so it's years out and announcements are cheap — but the direction is one-way, and it's why conviction sits where it does. Adjacent: Volta hit 214m of rare earths and gallium in Ontario, and the Pentagon put $5.2M into Bitterroot's nickel-copper drilling. Western critical-minerals supply is being built by many small hands at once. Each one is immaterial; the aggregate erodes exclusivity.

What Would Prove Us Wrong

A second nuclear issuer pulls or prices below range within two weeks — or conversely, Holtec relaunches at or above its original range inside 60 days. The first confirms LEU's weakest pillar is cracking; the second says today was deal-specific and conviction should come back up.
Q4 hyperscaler capex guidance comes in flat or down versus Q3 (Microsoft, Amazon, Google, Meta). That would convert today's defensive rotation into a fact and break VRT's order-growth pillar. The inverse — raised guidance plus VRT backlog growth above 20% — says the optics rally was just investors finding a cheaper way in, and VRT's premium is safe.
Critical Metals signs a binding offtake or a financed FID on the Romania refinery (not another study) within 12 months. That turns a press release into a real second Western REE chain and takes MP's exclusivity pillar below its 0.05 prior. No FID by mid-2027 and it stays a slideshow.

Thesis Impact

LEU | Conviction: DOWN | Surprise: MED | Holtec — an SMR developer and nuclear services supplier — pulled its US IPO the day before pricing. Capital markets are backing away from nuclear paper right now, which is the first real crack in the "SMR pipeline = the HALEU demand curve" pillar (P=0.50, the weakest one). CONTRADICTS, at the margin: if SMR builders can't raise, the 2030s demand wave slips. Not a trigger yet — watch whether other nuclear issuers follow. | ¹

VRT | Conviction: DOWN | Surprise: MED | Lumentum and Coherent led the S&P today because investors are rotating *further down* the AI supply chain "amid concerns of a spending slowdown." The move is bullish tape, but the stated reason is defensive — money hunting for exposure that survives a hyperscaler capex pause. VRT's whole pillar 2 is "order growth tracks hyperscaler capex," and its premium is the risk. CONTRADICTS mildly. Pairs with the recurring Semafor piece on slowing AI development helping hyperscaler balance sheets — two items now pointing the same way. | ⁶

MP | Conviction: DOWN | Surprise: MED | Critical Metals says it solved the eudialyte processing problem and plans a $2.2B/yr refinery in Romania fed by Tanbreez (Greenland). That is a credible second Western mine-to-refined REE chain being drawn up — MP's explicit thesis-breaker. It's a preliminary study, so years out and cheap to announce. CONTRADICTS the exclusivity pillar; already reflected in a 0.05 prior, so this mostly confirms why conviction is there. | ⁹

Own-the-Bottleneck | Conviction: HOLD | Surprise: MED | The House passed a bill shielding households from data center power costs — the backlash has gone from local zoning fights to federal legislation. Read it as CONFIRMS: you don't legislate over who pays unless the resource is genuinely scarce. Second-order, it pushes cost onto data centers and makes behind-the-meter and co-located generation relatively more attractive versus the grid. No pillar moves; the political fight is the scarcity showing up in another form. | ³

CEG | Conviction: HOLD | Surprise: LOW | PJM study says loss-of-load could run 100x its planning criterion by 2030 — 13+ outage days a year — if data center load keeps growing. Strongly CONFIRMS the firm-baseload scarcity pillar and supports capacity prices. But this is recurring (0d) and already fully in a 0.95 prior. No move; it's the thesis working, not new information. | ²

Inflection Radar

[emergent] AI Compute Bottleneck | The industry focus is shifting from raw model size to optimizing inference latency and cost via distillation and specialized architectures for hot-path decision engines. This signals a critical hardware/software bottleneck for real-time AI deployment. | Touches: NEW | ¹⁰

[emergent] AI-Grid Convergence | Major tech players (Nvidia, Google) are forming alliances to manage the massive power draw of AI data centers, signaling that energy demand and grid flexibility will become primary strategic constraints for compute scaling. | Touches: NEW | ¹¹

[emergent] Advanced Autonomy Control | Robotics research is moving beyond end-to-end vision models toward neuro-symbolic, physics-constrained frameworks for whole-body control and compositional manipulation (e.g., humanoids). This requires integrating complex physical modeling into AI pipelines. | Touches: NEW | ¹²

[emergent] Localized Defense Power | Both the Navy (LCU 1700 prototypes) and Air Force (microreactor selection) are signaling a strategic pivot toward advanced, localized power sources for critical assets, reducing reliance on traditional grid infrastructure. | Touches: NEW | ¹³

[dismissive] Defense Procurement Friction | Persistent concerns regarding the industrial base and timelines for next-generation fighters (F/A-XX, F-47) highlight structural bottlenecks in major defense acquisition programs that are not purely technical. | Touches: NEW | ¹⁴

QA & Caveats

No issues found.