← Feed🎯 Predictions📈 Theses📚 LogMorning Analysis2026-09-11 · Opus 4.7 (Max)

Morning Analysis — 2026-09-11

Dek

Microsoft just put a hard number on AI's power appetite — 38 gigawatts — in the same week a third state slammed the door on where you can plug it in.

The Big Shift

Microsoft is targeting 38GW of data-center capacity by 2032, against roughly 12GW today — a tripling, and a concrete commitment from the single largest buyer rather than another analyst forecast (1). That's 26GW of new load over six years, about the size of Chile's entire electricity system, and it has to be firmed by equipment you can order now — turbines, switchgear, fuel cells, reactors. The signal: demand is now a stated plan, not a projection, while the places you're allowed to build keep shrinking. Value migrates to whoever already holds power and the gear that converts it.

Analysis

Power: demand hardened, supply of *sites* softened. Massachusetts is the third state in three months to impose clean-power conditions on data centers (2), Nevada's first federal-land data center lost in court, and at least five more states look primed for pauses (3). The tell that this has gone national: OpenAI hired a JPMorgan community-engagement lead because data centers are becoming a midterm campaign issue (4). Implication: this is the cleanest confirmation yet of Own-the-Bottleneck — more demand chasing fewer permitted sites is the definition of scarcity rent. Watch the second-order move: MISO's "zero injection" fast-track, backed by Google and Xcel, would speed approval for generators that sit next to the load they serve (5). Behind-the-meter power — generating on-site instead of queueing for grid access — is becoming the default workaround, which is bullish TLN and GEV.

Fuel: Hormuz is repricing the firming layer. Shipping through the Strait has fallen to single digits per day, and the IEA now says the resulting LNG squeeze is lifting *coal* demand 1.2% as power buyers substitute (6). Almost no coal moves through Hormuz — this is pure price transmission, and it's the first hard agency data behind the gas spike rather than futures noise. Implication: a direct hit to the Bloom Energy thesis, whose on-site fuel cells burn gas; sustained expensive gas breaks the cost-curve pillar. It also raises the bill on "just build turbines," which is the reflexive answer to Microsoft's 26GW. But note our cross-domain read: the fear premium is deflating faster than the shortage underneath it. Gas prices can normalize in weeks when the Strait reopens. The 26GW does not go away.

Materials: separate the tariff story from the deficit story. Copper fell hard Thursday on reports the White House still hasn't decided on refined-copper tariffs — Freeport and Teck down 7% — then steadied Friday as funds and Chinese physical buyers stepped in (7). Underneath, China's August imports were the weakest in six years *and* copper still sits near $14,000/t, with mine supply heading for its first annual drop since 2017 (8). Implication: the tariff was always a US-premium trade, never the deficit trade — Thursday told us how much of FCX's move was hope, not how much copper exists. FCX and SCCO theses intact; the $10,000 trigger isn't close.

The supply response confirms the deficit rather than curing it. BHP and KGHM formed a global copper exploration JV, Rio Tinto cleared Aboriginal consent for Winu (first production 2030), Chile is opening capital markets to junior explorers (9). The world's largest miner going hunting is what a structural shortage looks like from the inside — and none of it delivers metal this decade. Same pattern in rare earths, where the non-China pipeline is thickening: a US$450M Brazil project picked engineering partners, an Arizona recycling plant is targeting the crunch, Brazil is debating price floors to fund refining. Implication: none of these is integrated mine-to-magnet, so MP's trigger doesn't fire — but this is the direction that slowly erodes a scarcity premium. Keep conviction at the floor. Watch the OECD traceability push too (10): minerals-tracking standards are how a soft trade barrier gets built without anyone calling it one.

Compute: cheaper inference raises the power bill, it doesn't lower it. DeepSeek's V4.1 Flash claims better performance at lower cost (11) — the standard read is that efficiency reduces demand, but the historical pattern is the opposite: cheaper compute expands usage faster than it saves. And when serious money starts going into *space* nuclear power ($161M to Antares) and orbital data centers, the real signal isn't space — it's how expensive and contested terrestrial siting has become.

What Would Prove Us Wrong

Thesis Impact

Inflection Radar

[dismissive] LLM Evaluation Critique | Research highlights that current methods for testing LLMs often mistake mere predictability for genuine membership evidence, suggesting a methodological flaw in assessing model knowledge. | Touches: NEW | 16

[dismissive] AI Funding Caution | A research startup pulling a major funding round for a key enterprise client talk signals potential friction or overvaluation risk in the high-end B2B AI deployment market. | Touches: NEW | 17

[emergent] GovTech Procurement Shift | DIU prototyping a new contract model to fast-track security clearances for startups signals a structural policy shift favoring rapid, small-scale defense technology integration. | Touches: NEW | 18

[emergent] Nuclear Microreactor Deployment | The Pentagon's push to site microreactors at Navy bases signals a critical, government-backed pivot toward decentralized, resilient, and rapid energy infrastructure. | Touches: NEW | 19

[emergent] AI Behavioral Diagnostics | Applying LLM and speech analysis to detect subtle paralinguistic markers for diseases like Alzheimer's moves AI into high-stakes, non-invasive medical diagnostics. | Touches: NEW | 20

[emergent] Maritime Multimodal AI | New research proposes using multimodal transformers to model complex, long-term physical behaviors (vessel trajectories), advancing AI for industrial simulation and prediction. | Touches: NEW | 21

QA & Caveats

Sources

  1. Microsoft targets 38GW of data center capacity in 2032 - report datacenterdynamics.com
  2. Massachusetts hits data centers with new clean power rules techcrunch.com
  3. These 5 States Are Primed for a Data Center Pause heatmap.news
  4. Amid data center public backlash, Yvette Eden Ruiz joins OpenAI’s Community Engagement team from JPMorgan datacenterdynamics.com
  5. Google, Xcel, others back MISO’s ‘zero injection’ large-load proposal utilitydive.com
  6. Coal demand to rise 1.2% as Hormuz closure drives gas prices: IEA mining.com
  7. Copper Steadies With Dip-Buying From Funds, Chinese Industry bloomberg.com
  8. Copper price pauses below record as China’s imports slump, mine supply heads for first drop since 2017 mining.com
  9. Chile bill targets foreign cash for next mining boom northernminer.com
  10. Traceability is essential for the critical minerals supply chain, report mining.com
  11. DeepSeek says new Flash AI model beats Kimi K3 on cyber, coding benchmarks scmp.com
  12. Mining stocks rally comes to abrupt halt as copper, silver prices plummet and gold slides mining.com
  13. Viridis selects EPCM partners for US$450 million rare earth project in Brazil - BNamericas news.google.com
  14. Fluor to serve as EPC contractor for Centrus’s Piketon plant expansion - American Nuclear Society news.google.com
  15. Australia's NextDC raises AU$1.1bn in convertible notes datacenterdynamics.com
  16. Detectable Only Where It Is Confounded: What Verified Duplication Counts Say About Membership Evidence in Language Model arxiv.org
  17. AI research startup Listen Labs scrubbed a $1.5B funding round for Salesforce talks techcrunch.com
  18. DIU moves to prototype new contract model that will fast-track security clearances for startups defensescoop.com
  19. Why the Pentagon picked an Indiana Navy base to host a nuclear microreactor breakingdefense.com
  20. LLM-Anchored Paralinguistic Enrichment for Alzheimer's Disease Detection arxiv.org
  21. M3-Former: Multimodal Transformer with Mixture-of-Experts for Long-Term Vessel Trajectory Prediction arxiv.org