Morning Analysis — 2026-09-22

Dek

Google spent the week buying clean power at two very different prices — and the cheap one quietly undercuts the whole scarcity trade.

The Big Shift

Google signed a deal to fund 96MW of uprates at Southern Co.'s Vogtle and Hatch nuclear plants in Georgia (¹). An "uprate" means paying for equipment upgrades that squeeze more output from a reactor that already exists — no new site, no new license, years instead of decades. That matters because it opens a third lane between waiting on new reactors and bidding up the scarce uncontracted output of merchant plants. If uprates become the template, they put a ceiling on what hyperscalers will pay for existing nuclear megawatts — which is the exact premium the Constellation thesis rests on.

Analysis

There is now a visible price ladder for firm power, and buyers are climbing it from the bottom. Cheapest and fastest: uprates at regulated utilities, where the ratepayer-backed balance sheet absorbs the risk. Next: behind-the-meter gas ("behind-the-meter" = the plant feeds the data center directly, skipping the public grid and its multi-year interconnection queue). Most expensive and slowest: new nuclear. Google is buying from the bottom and the top simultaneously. For CEG, the middle rung — paying a scarcity premium for existing merchant nuclear — is the one getting squeezed from both sides. Nothing has broken yet, but watch whether a second hyperscaler copies the uprate structure. That would be the confirmation, not this deal alone.

The new-nuclear top rung got materially more real overnight. Kairos Power named Samsung C&T as its EPC partner — the firm that actually engineers and builds the thing — for a 50MW small modular reactor for Google in Tennessee, targeting 2030 (²). A named builder and a named offtaker turn a slide-deck reactor into a dated, funded project. Kairos runs on HALEU-based TRISO fuel — high-assay enriched uranium, the specialty product LEU exists to make — so this converts projected fuel demand into scheduled fuel demand. The regulatory plumbing is moving in parallel: NRC is taking comment on ³, the rulebook these designs need. And the market is already marking the sector up — Cameco's $2.1B stake in Westinghouse could be worth ~$24B at listing (⁴).

Meanwhile West Texas keeps answering the power question with gas engines. New Era signed a 20-year PPA with a Vistra subsidiary to run a 250MW data center off an adjacent Odessa plant (⁵), and Kodiak is deploying 40 reciprocating gas units for 76MW at another West Texas site (⁶). Two implications. For VST, twenty-year contract terms on the *gas* fleet — previously a nuclear-only structure — widen the contracted book and deserve a re-rate in how we value those assets. For BE, the grid-bypass problem is real but the market keeps choosing gas engines over fuel cells; that's a slow bleed against the thesis, already priced into our low conviction. Siting politics reinforce the Texas pull: two years of bruising data center fights in Pennsylvania (⁷) show what developers are paying to avoid.

The binding constraint may be quietly shifting from electrons to equity. JPMorgan's CEO put hyperscaler capex at $1 trillion next year and still climbing, with investors shrugging (⁸); Goldman sees $1.4T by 2027. The uncomfortable companion fact is that capex will consume ⁹. Past that line, the buildout gets funded with debt and equity rather than profits — which is why today's Nscale IPO is the real event. Nscale books most of its revenue from two customers, Microsoft and Anthropic (¹⁰). That is the same customer-concentration shape as IREN's AI Cloud and the same open capital window KEEL needs to fund 2.2GW with no anchor lease. A weak print marks down the whole comp set and raises everyone's cost of capital in one afternoon.

On the physical inputs, sovereign money is moving faster than private money. US copper warehouses are nearly full with another 100,000 tonnes inbound, Shanghai premiums hit a four-year high, and London flipped back into backwardation — near-term metal pricing above future metal, the classic tightness signal (¹¹). A UAE-backed buyer paid a 73% premium for Global Lithium to build an integrated supply chain (¹²), and Washington is taking equity in a Brazilian rare earth mine and a Mozambican graphite mine with governance rules still unwritten (¹³). The pattern: states are buying the inputs while markets buy the compute. That's a tax on the buildout that shows up in transformer, cable, and switchgear lead times long before it shows up in a headline. One macro risk did ease — Saudi pipeline flows and Hormuz diplomacy pushed oil lower with the S&P near record highs (¹⁴) — which removes an excuse for a risk-off tape today and makes the Nscale print a cleaner read on AI-infra appetite specifically.

What Would Prove Us Wrong

• Nscale prices below range or breaks issue in the first session. That's a direct markdown of concentrated AI-infra paper and hits IREN (Pillar 3 assumes it doesn't need expensive outside capital) and KEEL (Pillars 2 and 3 both need the window open). Measurable today. A strong print does the reverse and would argue for raising KEEL.
• A second hyperscaler signs an uprate deal instead of a merchant nuclear PPA — say Microsoft or Amazon buying incremental MW from a regulated utility's existing fleet in the next 60 days. That would confirm uprates as the template and force a downgrade of CEG's Pillar 2 scarcity-pricing premise from HOLD to a real cut.
• The Kairos–Google project sources fuel from somewhere other than the domestic HALEU chain, or the 2030 date slips publicly. Either would pull LEU's Pillar 4 back from "dated and funded" to "projected." Watch Centrus 8-Ks for a Kairos-linked supply agreement — there have been four LEU filings in the last two weeks, including a material agreement on Sept 11, and the contents matter more than the count.

Thesis Impact

• LEU | Conviction: UP | Surprise: MED | Kairos picks Samsung C&T to build a 50MW SMR for Google in Tennessee by 2030 — a named EPC partner and a hyperscaler offtaker turn Pillar 4 (the HALEU demand curve) from projection into a dated, funded project; Kairos runs on TRISO/HALEU fuel, so this is direct demand for the one thing Centrus makes. CONFIRMS. | ²
• VST | Conviction: UP | Surprise: MED | New Era signed a 20-year gas PPA with a Vistra subsidiary to power a 250MW data center in Ector County, TX, off an adjacent Odessa gas plant — small in MW but it proves Vistra can monetize the *gas* fleet on nuclear-style 20-year terms, widening the contracted book beyond the Meta/Amazon nuclear deals. CONFIRMS Pillars 1–2. | ⁵
• CEG | Conviction: HOLD | Surprise: MED | Google is funding 96MW of uprates at Southern's Vogtle and Hatch — the residual here is that hyperscalers now have a cheaper, faster nuclear option than bidding for Constellation's uncontracted MWh: pay a regulated utility to squeeze more out of existing reactors. Mildly CONTRADICTS Pillar 2's scarcity-pricing premise; doesn't touch Pillars 1 or 3. Watch whether uprate deals become the template. | ¹
• VRT | Conviction: UP | Surprise: LOW | JPMorgan's CEO puts hyperscaler capex at $1T next year and still growing — the demand source for Pillar 2 keeps getting marked up. The tell is the subhead: investors are shrugging it off, which is the premium-compression risk (Pillar 3) rather than a demand problem. CONFIRMS. | ⁸
• IREN | Conviction: HOLD | Surprise: MED | Nscale's IPO prices this week as the market's read on concentrated AI-infra bets — it books most revenue from two customers (Microsoft, Anthropic), the same shape as IREN's AI Cloud. A weak print marks down the comp set and raises the cost of the capital Pillar 3 assumes IREN won't need. No result yet, so no move — but this is today's cleanest disconfirming risk for the crypto-to-AI names. | ¹⁰
• KEEL | Conviction: HOLD | Surprise: MED | Same Nscale read, sharper: Keel has no anchor lease yet and is burning cash, so it needs the IPO window open to fund the 2.2GW pipeline. If Wall Street turns on concentrated AI-infra paper, Pillars 2 and 3 both get harder. CONTRADICTS on financing risk; nothing new on the pipeline itself. | ¹⁰
• Own-the-Bottleneck | Conviction: HOLD | Surprise: LOW | The $1T capex number lands alongside the recurring finding that capex will consume nearly all hyperscaler operating cash flow in 2026 — Pillar 1 (power is binding) is intact and reinforced, but the funding ceiling is now visible. Already in the prior at 0.88; this is the thread to watch, not a move. CONFIRMS. | ⁹
• BE | Conviction: HOLD | Surprise: LOW | Two more West Texas data centers went behind-the-meter with plain gas — New Era off a Vistra plant, Kodiak with 76MW of reciprocating engines. The grid-bypass problem is real, but the cheap answer keeps being gas engines, not solid-oxide fuel cells. CONTRADICTS Pillar 3; already reflected in a 0.05 prior, so no move. | ⁶

Inflection Radar

[emergent] AI for Climate Modeling | High-resolution solar radiation downscaling using flow matching shows a path to physically consistent energy forecasting beyond current coarse reanalysis products. | Touches: NEW | ¹⁵

[emergent] LLM Planning in Physical World | FRAMES addresses the critical gap where LLM-selected robot skills fail during physical execution, moving planning from theory to reliable action. | Touches: NEW | ¹⁶

[emergent] AI Energy Management Alliance | Major tech players (Nvidia, Google) are forming an alliance to manage AI's massive energy draw, signaling the commoditization of 'demand flexibility.' | Touches: NEW | ¹⁷

[emergent] Geopolitical Resource Shift | Potential agreement for Iran to surrender enriched uranium creates a sudden, high-impact signal regarding global nuclear fuel supply dynamics. | Touches: NEW | ¹⁸

[emergent] Data Center Regulatory De-risking | Texas PUC softening interconnection rules (eliminating non-refundable fees, easing requirements) signals a regulatory shift favoring rapid hyperscale CAPEX deployment. | Touches: NEW | ¹⁹

[emergent] Prospective Memory Retrieval | Moving personal memory retrieval from merely *resembling* a query to anticipating what the user needs next fundamentally changes human-computer interaction models. | Touches: NEW | ²⁰

[emergent] Humanoid Skill Synthesis | CHOREO aims to unify disparate humanoid skills (RL, imitation, generative modeling) into a single framework, addressing the current siloed nature of advanced robotics capabilities. | Touches: NEW | ²¹

[emergent] Cross-Domain AI Inference Optimization | Multiple arXiv papers (e.g., 47, 51, 53) signal a collective industry focus on making LLM inference cheaper and faster via sparse attention, quantization, and speculative decoding. | Touches: NEW | ²²

QA & Caveats

• LEU and VST are directly supported by the cited deals.
• CEG’s contradiction regarding scarcity pricing is an implication, not a direct statement from the source.
• VRT and IREN rely on market sentiment/IPO pricing; they are observations rather than hard confirmations of the thesis mechanics.
• KEEL and Own-the-Bottleneck restate known trends about capex and pipeline funding; no new risk is introduced here.
• BE contradicts Pillar 3, but the source only notes the reality of gas usage versus SOFCs, not a direct refutation of the pillar premise.

Sources

1. Google inks deal to support nuclear uprates at Georgia Power's Vogtle and Hatch plants datacenterdynamics.com
2. Kairos Power selects Samsung C&T to help build 50MW SMR for Google in Tennessee datacenterdynamics.com
3. Draft Interim Staff Guidance: Material Compatibility for Non-Light Water-Reactors federalregister.gov
4. Cameco may turn $2.1B into $24B with Westinghouse listing this fall northernminer.com
5. New Era signs 20-year gas PPA with Vistra subsidiary to power 250MW data center in Ector County, Texas datacenterdynamics.com
6. Kodiak Gas Services to supply 76MW of behind-the-meter gas power to West Texas data center datacenterdynamics.com
7. Everyone can find a reason to dislike data center construction techcrunch.com
8. JPMorgan CEO Sees Hyperscaler Spending Hitting $1 Trillion Next Year And Still Growing — Investors Are Shrugging It Off news.google.com
9. Capex to Consume Nearly All Hyperscaler Operating Cash Flow in 2026 - Broadband Breakfast news.google.com
10. Nscale’s IPO will test Wall Street’s appetite for concentrated AI bets once again techcrunch.com
11. Copper prices bounce as US runs out of warehouse space mining.com
12. Global Lithium soars 50% on A$333M Titan takeover mining.com
13. Washington’s Foreign Equity Bets Need Better Governance warontherocks.com
14. Latest Oil Market News and Analysis for Sept. 22 bloomberg.com
15. SolarFlowRefiner: Refinement-Aware Flow Matching for Surface Solar Radiation Downscaling arxiv.org
16. FRAMES: Failure Recovery And Monitoring of Embodied Skills for Humanoid Loco-Manipulation arxiv.org
17. AI Heavyweights Team Up to Promote Demand Flexibility heatmap.news
18. Iran agrees to surrender enriched uranium stockpile by… Odds & Prediction Market Analysis - CryptoSlate news.google.com
19. Texas PUC adopts softened rules on data center interconnection utilitydive.com
20. Memory That Looks Forward: A Zero-Inference Prospective Term for Personal Memory Retrieval arxiv.org
21. CHOREO: Every Humanoid Skill as a Trajectory arxiv.org
22. RBS-Attention: Radius-Bounded Sparse Prefill for Long-Context Large Language Models arxiv.org