Morning Analysis — 2026-09-04

Dek

The scarcity trade is splitting in two overnight: copper's deficit hardened into arithmetic while America's one-mine lithium monopoly got word that eight or nine neighbors are moving in.

The Big Shift

Bloomberg reports copper now faces a rare outright *decline* in mine supply as project setbacks pile up — not slower growth, less metal (¹). That matters because a demand forecast can be wrong, but a mine that didn't get built is a fact you can't argue with; the deficit stops being a story about AI datacenters wanting copper and becomes a story about there not being enough. It points to price setting on the supply curve, which pays the low-cost incumbents first — we raised conviction on SCCO and, with more torque to price, FCX.

Analysis

Copper: the US answer is real, and it's too slow. The Southwest pipeline is genuinely activating — Ivanhoe Electric is chasing $1B of US-backed debt for Santa Cruz in Arizona (²), Gunnison is hunting government cash (³), and the Forest Service just cleared South32's Hermosa plan (⁴). But the binding constraint isn't permits, it's water — whether these projects can secure decades of it (⁵). None of this delivers tonnes inside the window that sets prices for the next three years, which is exactly why the incumbent-producer read holds. The live counter-argument: Reuters' Andy Home's line that there's no copper shortage, it's just *in the wrong place* (⁶) — a refining and logistics problem, not a geology one. If he's right, the money goes to smelters and traders, not miners. That's the fork to resolve.

Lithium: the first bottleneck visibly unclenching. Albemarle's Silver Peak in Nevada is the only producing lithium resource in the US, and as many as nine new mines could join it (⁷). This is disconfirming evidence for Own-the-Bottleneck — the thesis that owning a chokepoint beats owning the end product. One chokepoint is easing on a 2–3 year view. It doesn't trip the "several ease at once" trigger: transformers, HALEU (the enriched uranium new reactors need), and dysprosium (a magnet metal China controls) are all still tight, and rare-earth answers like the $100M Missouri magnet-recycling JV (⁸) are years from mattering. Note the cross-current: China's largest lithium mine just lost its licence, halving CATL's output forecast there (⁹), even as Chinese battery shipments head toward roughly double today's volume by 2030 (¹⁰). Supply is loosening in the US and tightening in China at the same time — that's a relocation of the bottleneck, not its disappearance. Hold, watch copper and rare earths for the same pattern.

Power: the constraint has moved inside the fence line. The question is no longer whether electricity exists, it's whether you can get it delivered to your own meter, and on what terms. PG&E's Flex Connect lets big loads energize in *months* instead of the several years a firm grid connection takes, and interest is rising (¹¹). Meanwhile California's legislature sent the governor bills creating separate electricity tariffs for large data centers (¹²). Read together: grid power for AI is getting faster to obtain but more expensive and more politically conditional. That's mildly good for BE and behind-the-meter generation, and it's the real risk to PWR — if flexible interconnection scales, new AI load routes *around* the transmission Quanta builds, which weakens the "wires are the binding step" pillar. Storage is doing the same work from the other side: Equinor's 100MW Texas battery came online (¹³) and batteries are why ERCOT's summer was uneventful (¹⁴).

Compute: capex confirmed, but the marginal information has moved to backlogs. Every hyperscaler raised capex again (¹⁵), which confirms the VRT/ETN pillars — and is already fully in the price at 0.89 and 0.69 conviction. So it's noise until an order book prints. Today's actual read: VRT filed two 8-Ks on Sept 2, one flagged as a material agreement plus a Reg FD disclosure. That's where new information would live, not in the capex aggregators. Two structural tells underneath: Taiwan's Liteon paid ~$176M for a quarter of Polish liquid-cooling firm DCX (¹⁶) — the Asian supply chain buying European thermal capacity, because cooling is the scarce input now and it's being locked up rather than bought. And the buildout keeps dispersing to second-tier geographies — Bangkok, Cancún, a 65MW site outside Frankfurt, Tasmania, rural Texas — which is what you do when the primary markets are power-constrained. Watch the froth signal too: GoPro is becoming a data-center company via merger with an optical photonics outfit (¹⁷). Late-cycle reverse mergers are not a bullish tell.

Energy geopolitics: the number everyone is trading is unreliable. Oil is set for a weekly surge on renewed US–Iran fighting and Hormuz disruption (¹⁸). But the flow estimate at the center of it — the EIA's 4.9 million barrels a day for Q2, against 21.6 million in Q4 2025 — comes from data the EIA itself calls "especially unreliable" (¹⁹), and Goldman's work suggests Gulf exports have recovered further than the official numbers show (²⁰). Two implications. First, positioning built on that figure is fragile in both directions — a revision is a market event on its own. Second, sustained high energy prices raise the operating cost of every mine (diesel, reagents, power), which widens the gap between low-cost and marginal producers and reinforces the case for cost-curve winners like SCCO over higher-cost leverage plays. It also keeps the political tailwind behind domestic materials and behind nuclear, including the Army's push for microreactors on bases (²¹).

What Would Prove Us Wrong

Rising copper treatment and refining charges (TC/RCs), or ICSG mine-supply prints coming in flat-to-up. TC/RCs are what smelters charge to process ore — they rise when concentrate is abundant. If they climb while prices stay high, the shortage is a refining and location problem, not a tonnage one, and the money belongs to processors. Hits SCCO and, harder, FCX, whose price leverage assumes a genuine supply deficit.
A second and third bottleneck easing alongside lithium — specifically transformer lead times dropping under ~24 months, or a credible non-China dysprosium or HALEU supply commitment with a date attached. Lithium alone is one crack. Three at once means the scarcity regime is over. Hits Own-the-Bottleneck directly and takes the wind out of the whole materials complex.
PG&E publishing measurable Flex Connect enrollment in gigawatts, or another large utility (Dominion, Oncor, AEP) launching a copycat flexible-interconnection program. That would show AI load can be served without new transmission, undercutting the "wires are the binding step" pillar. Hits PWR first, and secondarily the grid-buildout leg of ETN.

Thesis Impact

SCCO | Conviction: UP | Surprise: MED | Bloomberg (T1) says copper faces a rare outright *mine supply decline* as setbacks mount — the deficit is now a supply-side fact, not just a demand forecast; low-cost producers capture the price. CONFIRMS pillars 1 and 2. | ¹
FCX | Conviction: UP | Surprise: MED | Same Bloomberg supply-decline read, with more torque: FCX's leverage to price is higher than SCCO's. CONFIRMS pillar 1. Offset: the recurring "no shortage of copper, it's just in the wrong place" thread argues the problem is refining/location, not tonnes — which favors processors over miners. | ¹
Own-the-Bottleneck | Conviction: HOLD | Surprise: MED | Disconfirming: as many as nine new US lithium mines could join Albemarle's Silver Peak, ending the single-mine monopoly. One bottleneck visibly easing on a 2–3 year horizon. Doesn't trip the "multiple bottlenecks ease at once" trigger — transformers, HALEU, and Dy are all still tight — but it's the first crack. Watch whether copper/REE follow. | ⁷
PWR | Conviction: HOLD | Surprise: MED | Disconfirming: PG&E's Flex Connect lets large loads energize in *months* instead of "several years" for firm interconnection, and interest is rising. If flexible interconnection scales, it routes new AI load around the wires Quanta builds — softening pillar 1's "binding step" claim. Recurring item, so no move yet; it's the thread to watch. | ¹¹
VRT | Conviction: HOLD | Surprise: LOW | Every hyperscaler raised capex again — CONFIRMS pillar 2 (orders track capex), but it's exactly what the prior expects and already at P=0.69. T2 aggregator, no new order data. Noise until VRT's own backlog print. | ¹⁵
ETN | Conviction: HOLD | Surprise: LOW | Same capex-raise item; CONFIRMS pillars 1–2. Already the highest-conviction name at 0.89 and fully in the prior. No move. | ¹⁵
BE | Conviction: HOLD | Surprise: LOW | California's data-center ratepayer bills (separate tariffs for large loads) now sit with the governor — grid power for DCs gets more expensive and more political, which is the case for generating behind the meter. CONFIRMS pillar 1 directionally. Recurring, and the same day's PG&E flex-interconnect news cuts the other way. Not enough to move 0.10. | ¹²

Inflection Radar

[emergent] Tool-Using Agent Architecture | LLM agents are moving beyond pure inference; the critical bottleneck is managing serial action-observation turns and environment transitions, requiring new architectural models for stateful tool use. | Touches: NEW | ²²

[emergent] Causal Foundation Models | Moving beyond correlation to estimate the effect of interventions (causal inference), signaling a shift from descriptive ML models to prescriptive, real-world decision engines. | Touches: NEW | ²³

[emergent] On-Device LLM Scaling | Addressing the physical constraint of model weights exceeding available DRAM on edge devices by developing highly efficient, low-parameter streaming frameworks. | Touches: NEW | ²⁴

[emergent] Generalist Humanoid Embodiment | The development of open-source platforms that decouple morphology and control design is necessary to move beyond specialized tasks toward general, human-level physical AI. | Touches: NEW | ²⁵

[dismissive] Geopolitical Material Constraint | IAEA's inability to verify Iran’s enriched uranium stockpile signals a persistent, unquantifiable material risk and potential choke point for global energy supply chains. | Touches: NEW | ²⁶

[emergent] Advanced Counter-UAS System Integration | The maturation of counter-UAS technology is shifting from standalone prototypes to integrated, distributed kill chains linking sensors, C2, RF effects, and lasers. | Touches: NEW | ²⁷

[emergent] Energy Infrastructure Regulatory Filing | FERC accepting a preliminary permit application for DeepGreen's Cook Inlet SPV signals active, high-level regulatory maneuvering in frontier energy development. | Touches: NEW | ²⁸

[dismissive] AI Startup Valuation Bubble | Multiple reports confirming $30B valuations and large contracts (e.g., Crusoe/Jane Street) suggest a period of extreme, potentially unsustainable capital deployment in the AI infrastructure layer. | Touches: NEW | ²⁹

QA & Caveats

No issues found.

Sources

Copper at Risk of Rare Supply Decline as Mine Setbacks Mount bloomberg.com
SW US CEO chat: Ivanhoe Electric maps $1B debt path for Arizona copper project northernminer.com
SW US CEO chat: Gunnison Copper seeks government cash and partner northernminer.com
US Forest Service signs off on South32 Hermosa’s mine plan in Arizona mining.com
SW US mine rush collides with water limits northernminer.com
There’s no shortage of copper, it’s just in the wrong place mining.com
SW US: New US lithium mines are coming northernminer.com
Ionic Rare Earths, US Strategic Metals form JV to build $100M magnet recycling plant in Missouri mining.com
China’s biggest lithium mine loses licence northernminer.com
China’s lithium-ion battery shipments could top 3,600 GWh by 2030, roughly double the 2025 figure. reddit.com
PG&E sees rising interest in ‘customer-driven’ flexible interconnection pilot utilitydive.com
California lawmakers pass data center ratepayer protection bills, send to governor for approval datacenterdynamics.com
Equinor just brought its biggest US battery online in Texas electrek.co
The Texas Grid Has Had a Boring Summer Thanks to Batteries heatmap.news
Every Hyperscaler Raised Capex Again. These 3 Boring Industrials Get Paid No Matter Who Wins the AI Race. - The Motley F news.google.com
Taiwan-based Liteon purchases 25 percent stake in Polish liquid cooling firm, DCX datacenterdynamics.com
GoPro to enter AI data center market via merger with optical photonics company datacenterdynamics.com
Latest Oil Market News and Analysis for Sept. 4 bloomberg.com
The most important oil number right now is based on data the EIA calls “especially unreliable” reddit.com
Crossing Hormuz in the dark: The secret oil operation testing Trump’s claims reddit.com
Why the Army wants to deploy nuclear microreactors on US bases breakingdefense.com
Speculative Macro Commit for Faster Tool-Using Agents arxiv.org
Causal Foundation Models arxiv.org
LeanStream: A Speculate-and-Refine Streaming Framework for Efficient on-Device LLM Inference arxiv.org
BRIDGE: An Open-Source Humanoid Platform via Morphology-Control Co-Design for Physical AI arxiv.org
IAEA Says It Cannot Verify Iran’s Enriched Uranium Stockpile - kurdistan24.net news.google.com
Three exercises that pushed counter–UAS from prototype to proven breakingdefense.com
DeepGreen Cook Inlet SPV, LLC; Notice of Preliminary Permit Application Accepted for Filing and Soliciting Comments, Mot federalregister.gov
Accel reportedly in talks to lead $1B round for Thinking Machines at $40B valuation techcrunch.com