Evening Analysis — 2026-10-06

Dek

Hyperscalers and the federal government spent the day buying the same scarce thing — existing nuclear megawatts — while a quiet FERC docket gave regulators the power to claw back the revenue that makes those megawatts worth owning.

The Big Shift

Google signed 20-year power purchase agreements with Constellation for 3,590 MW in the PJM grid — 890 MW of genuinely new capacity from nuclear "uprates" (squeezing more output from existing reactors) plus 2,700 MW of existing generation, backed by $4.3B of investment (¹). This is the second hyperscaler-nuclear deal in a week, after Amazon's Calvert Cliffs agreement (²), and it matters because it converts uncontracted power — sold at whatever the market pays, year to year — into two decades of locked-in revenue at scarcity prices. That is the central CEG thesis, and it is being proven faster than we assumed. What it points to: the window for buying unhedged nuclear cheaply is closing, because the buyers are now bidding against each other in public.

Analysis

Power: the state joined the trade. Hours after the Google deal, the DOE said it intends to lend Vistra $4.2B for nuclear fleet improvements — 433 MW of uprates and preserving roughly 4 GW of existing PJM capacity (³). Uprates were the weakest leg of the VST case, because they're slow and capital-hungry. Subsidized federal debt removes exactly that objection. Implication: the cheapest new firm megawatts in America for the next three years are uprates at plants that already exist, and both corporate and government balance sheets are now paying for them. Conviction up.

But the revenue side can now run backward. FERC opened a Section 206 proceeding against PJM with a refund effective date (⁴). Plainly: the regulator can revise PJM rates retroactively, not just going forward. Every thesis in the book that rests on capturing high PJM prices now carries a backward-looking risk it didn't price yesterday. It hits Talen hardest — one big asset, one big counterparty, thin interest coverage, and a pending co-location question already in front of FERC. The paradox of the day is that the same grid is simultaneously the most valuable place to own generation and the least certain place to collect for it.

Gas is winning the load that solar was supposed to win. Proposed grid-connected gas capacity jumped 44% in eight months and has nearly tripled since early 2023, explicitly because of data centers — while new wind plans shrank (⁵). AWS's new 36-building campus sits at a gas-anchored site, the Homer City Energy Campus (⁶). That forward pipeline is what turns into turbine slot reservations and backlog, so it strengthens GEV. It cuts against First Solar: developers are choosing power they can dispatch on demand over power that shows up when the weather allows, for precisely the 24/7 AI load the solar thesis was counting on. The article names wind, not solar — read it as directional, not settled.

Demand is splitting, not breaking. Data center cancellations topped $260B in 2026 against more than $800B of expected spend (⁷) — the day's most serious disconfirming number, and the first real brush against the Eaton trigger. Yet Q2 data center capex rose 92% year-on-year on AI demand and memory costs (⁸). Both are true: speculative and locally-opposed projects are dying while committed hyperscaler budgets accelerate. Eaton and Vertiv sell into the second group, so pillars hold — but the load forecasts underpinning the bottleneck case are now less trustworthy at the margin. Two smaller items point the same direction: Alabama will require utilities to name data centers and disclose their expected peak demand (⁹), and MISO is moving to speed up its interconnection queue with state and developer backing (¹⁰). Disclosure strips out phantom demand; a faster queue is the first visible crack in the scarcity that makes the bottleneck valuable.

Materials and geopolitics: supply is loosening slowly, risk premia are not. The World Bank is backing $1.4B for Codelco and pitching the structure as a template for mineral-rich states (¹¹), and Greenland approved the Tanbreez heavy rare earth mine (¹²) — both real, both years from shipping metal, so near-term copper and magnet tightness for transformers and turbines stands. Meanwhile Taiwan overtook Korea as the preferred AI equity market on deeper supply-chain exposure (¹³) in the same 24 hours US Marines moved anti-ship systems to Japan's southwesternmost island near Taiwan (¹⁴), and oil rose on Iranian tanker attacks in the Strait of Hormuz (¹⁵). Implication: the compute leg of the thesis concentrates ever more value in the single most militarily contested supply chain on earth, which is the strongest argument for owning the domestic power and electrical-equipment legs instead.

What Would Prove Us Wrong

• A named hyperscaler cuts guided capex by 15% or more, or Q3 data center capex growth falls below roughly 30% year-on-year. That would turn the $260B cancellation figure from bifurcation into a genuine demand break and hit Eaton and Vertiv directly — their orders track hyperscaler budgets, and Vertiv carries a premium multiple that assumes those budgets keep compounding.
• FERC's Section 206 order actually sets lower PJM rates with refunds owed for the effective period. The docket is open; the question is the outcome. A retroactive revenue haircut hits Talen hardest and would also force a rethink of how much of the Constellation and Vistra re-pricing story is regulator-proof. Watch the order itself, not commentary about it.
• A hyperscaler signs a large firm-power deal with solar-plus-storage at a credible round-the-clock price, or new wind and solar interconnection requests rebound alongside gas. Either would undercut tonight's read that AI load is structurally choosing dispatchable gas and nuclear, and would restore the First Solar case we just marked down.

Thesis Impact

• CEG | Conviction: UP | Surprise: MED | Google signed 20-year PPAs totaling 3,590 MW in PJM (890 MW of new nuclear uprates + 2,700 MW existing, $4.3B investment) — the second hyperscaler deal in a week after Amazon/Calvert Cliffs, taking Constellation's nuclear expansions to ~1.1 GW. Pillar 2 (uncontracted fleet converts to 20-year PPAs at scarcity prices) is now being proven at a faster clip than the prior assumed; the uprate component means new MW, not just re-pricing old ones. CONFIRMS. | ¹
• VST | Conviction: UP | Surprise: HIGH | DOE intends to lend Vistra $4.2B for nuclear fleet improvements — 433 MW of uprates plus preserving ~4 GW of existing PJM capacity. Federal capital at subsidized cost directly underwrites pillar 4 (uprates add low-cost capacity), which was the weakest pillar at P=0.55. Not in the prior's expected-news list: this is government balance sheet, not an operating milestone. CONFIRMS. | ³
• GEV | Conviction: UP | Surprise: MED | Proposed US grid-connected gas capacity jumped 44% in eight months and has nearly tripled since early 2023, explicitly driven by data centers — while wind plans shrink. Pair with AWS's 36-building campus at the Homer City Energy Campus (a gas-anchored site). Pillar 1 (gas turbines are the near-term firm-power answer) is strengthening in the forward pipeline, which is what feeds slot reservations and backlog. CONFIRMS. | ⁵
• Own-the-Bottleneck | Conviction: HOLD | Surprise: MED | Data center cancellations have topped $260B in 2026 against >$800B of expected spend — the day's most important disconfirming item. But Q2 data center capex still rose 92% year-on-year, so this is bifurcation, not a demand break: speculative and locally-opposed projects are dying while actual hyperscaler spend accelerates. Pillar 1 survives; the load forecasts underpinning it are now less reliable at the margin, and MISO's accelerated interconnection queue plan is the first visible crack in a bottleneck. MIXED — confirms the constraint, weakens the forecast. | ⁷
• ETN | Conviction: HOLD | Surprise: MED | The $260B cancellation figure is the first real brush against the named trigger (industry-wide data center capex cut of 15–30%), but the same day's data cuts the other way: +92% Q2 data center capex on AI demand and memory costs. Cancellations are concentrated in speculative/opposed projects, not in the hyperscaler budgets ETN's content sells into. Pillars intact at P=0.85 — but this is now the thing to track, not background noise. MIXED, leaning CONFIRMS. | ⁸
• VRT | Conviction: HOLD | Surprise: MED | +92% Q2 data center capex is direct confirmation of pillar 2 (orders track hyperscaler capex) and the strongest print we've seen on the demand source. Offset by the $260B cancellation number, which at a premium multiple is what the 0.28 prior is pricing. Net: the demand driver is accelerating, the valuation risk is unchanged. CONFIRMS pillar 2, does nothing for pillar 3. | ⁸
• TLN | Conviction: DOWN | Surprise: LOW | FERC instituted a Section 206 proceeding against PJM with a refund effective date — meaning PJM rates can be revised backward, not just forward. For the most fragile, single-asset, single-counterparty name in the book, with thin interest coverage and a pending co-location/FERC 203 question, retroactive revenue risk hits pillar 2 (clean front-of-meter transition post-FERC) hardest. Recurring, so already partly in the prior, but the refund date is the part that bites. CONTRADICTS. | ⁴
• FSLR | Conviction: DOWN | Surprise: MED | The gas-plan surge for data center load comes with shrinking new wind plans — utilities and developers are choosing dispatchable capacity for AI load, not intermittent. The article names wind, not solar, so read it as directional rather than conclusive, but it undercuts pillar 1 (cheap, fast solar fills the AI power gap): solar is being passed over for exactly the load the thesis was counting on. CONTRADICTS. | ⁵

Inflection Radar

[emergent] Open-Source Bimanual Robotics | An open, CE-certifiable bimanual service robot design is shared on public forums, bypassing commercial gatekeepers and accelerating hardware/safety development outside traditional industrial channels. | NEW | ¹⁶

[emergent] Advanced Humanoid Locomotion | Multiple arXiv papers detail advanced, coupled-load locomotion (e.g., pulling wheeled loads) and continual learning for humanoids, moving the focus from simple tracking to complex, real-world utility tasks. | NEW | ¹⁷

[emergent] AI Simulation & Safety | Research is rapidly advancing deep learning surrogates (e.g., Transformers) for high-fidelity simulations like crashworthiness, suggesting AI will soon move from prediction to core engineering design validation. | NEW | ¹⁸

[emergent] Energy Infrastructure Friction | Regulatory oversight is flagging significant conflicts (e.g., methane turbine deployment without air permits), signaling that rapid energy tech scale-up will increasingly hit local permitting and environmental compliance bottlenecks. | NEW | ¹⁹

[emergent] AI Agents for Code Reliability | The focus of autonomous coding agents is shifting from mere code generation to reliable, multi-step execution (reading, running, editing files), suggesting a move toward true software engineering automation. | NEW | ²⁰

[hype] IP Licensing in AI Compute | Major tech players are formalizing cross-domain patent agreements (e.g., Huawei/Qualcomm for 5G/AI), indicating that the next phase of compute growth will be defined by IP consolidation rather than just raw silicon capacity. | NEW | ²¹

QA & Caveats

• CEG: Confirms.
• VST: Confirms.
• GEV: No issues found.
• Own-the-Bottleneck: Mixed.
• ETN: Confirms.
• VRT: Confirms.
• TLN: Contradicts.
• FSLR: Contradicts.

Sources

1. Constellation-Google deal will bring $4.3B investment, 890 MW new nuclear to PJM utilitydive.com
2. Amazon, Constellation sign 20-year PPA to expand Maryland nuclear plant utilitydive.com
3. DOE intends to loan Vistra $4.2B for nuclear fleet improvements utilitydive.com
4. PJM Interconnection, L.L.C.; Notice of Institution of Section 206 Proceeding and Refund Effective Date federalregister.gov
5. US gas plant plans jump 44% in just 8 months – here’s why electrek.co
6. AWS plans 36-building data center campus in Indiana County, Pennsylvania datacenterdynamics.com
7. Exclusive: Data Center Cancellations Top $260 Billion in 2026 heatmap.news
8. AI demand and memory costs drove a 92% second-quarter data center capex increase - Lightwave Online news.google.com
9. rtoinsider.com rtoinsider.com
10. States, renewable developers back MISO’s accelerated interconnection queue plan utilitydive.com
11. World Bank backs $1.4B for Chile’s copper giant northernminer.com
12. Greenland government approves Tanbreez rare earth mine plan mining.com
13. Taiwan Dethrones Korea Atop Global Markets as AI Trade Widens bloomberg.com
14. US Marine Corps anti-ship weaponry to see exercise action on remote Japan island near Taiwan defensenews.com
15. Latest Oil Market News and Analysis for Oct. 7 bloomberg.com
16. I designed an open, CE-certifiable bimanual service robot in Italy: CAD, safety electrics and RL training are all on Git reddit.com
17. Humanoid Rickshaw Pulling: Whole-Body Locomotion under Coupled Wheeled Loads arxiv.org
18. BAT-NO: A Boundary-Condition-Aware Transformer Neural Operator for Crashworthiness Prediction of Vehicle Components arxiv.org
19. Elon Musk under fire after new report exposes controversial project: 'It is appalling'. xAI has brought in 35 or more po reddit.com
20. Teaching Agents to Code Reliably arxiv.org
21. Huawei, Qualcomm strike multi-year patent agreement across 5G, AI scmp.com