Weekly Macro & Outlook — 2026-09-25

Dek

The week the AI buildout stopped asking whether the power would arrive and started asking whether the tenant would still be paying when it did.

The Week in Macro

The constraint moved four times in five days. Monday it was land: Loudoun County, Virginia — the densest cluster of data centers anywhere — voted a 12-month pause on new applications (¹). Tuesday it was money: a $1.1 billion junk bond — debt sold by borrowers too risky for safe-money buyers — showed up to finance a CoreWeave-leased site (²), alongside the read that capex will eat nearly all of the big clouds' 2026 operating cash flow. Wednesday it was permission: Governor Abbott froze new large-load hookups to the Texas grid pending an audit due in December (³), and the same day Newsom signed a seven-bill package making California data centers pay their own grid costs and carry more wildfire liability (⁴). Thursday and Friday it was the buyer.

Oracle is the week's real story. It sent a force majeure notice to its landlord Blue Owl over Project Jupiter, the 2.5GW Stargate campus in New Mexico (⁵). Force majeure is the "acts of God" clause: you invoke it to stop paying without being in breach. Morningstar put a number on it — up to $25 billion of revenue potentially pushed out (⁶). Every bullish thesis on this desk rests on one quiet assumption: a signed order from a hyperscaler equals future revenue. A named, investment-grade buyer just showed in public that it doesn't — and did it over *construction delays*, not weak demand. That makes the risk structural, not a cycle wobble.

Put the two halves together and you get the week's cross-domain insight: the buyer's escape hatch has a mirror image. The seller's cost of saying yes — permits, interconnection queues, wildfire liability, paying for your own transmission lines — is rising faster than the contract price. Four US jurisdictions moved in five days, and none of them moved by *adding* capacity; they all moved by gating. Friday added a fifth, Prince William County cutting its by-right data center zone by two-thirds (⁷). That is still net-good for anyone holding megawatts that are already energized and connected, and bad for anyone holding a permit application. But the squeeze now cuts both ways: harder to build, and less certain you get paid when you do.

One quiet ceiling went in. Google agreed to fund 96MW of uprates at Southern Co.'s Vogtle and Hatch nuclear plants (⁸). An "uprate" means paying to squeeze more output from a reactor that already exists — no new site, no new license, years instead of decades. It's a cheap third lane between waiting for new reactors and overpaying a merchant generator for existing output. If it becomes the template, it caps what hyperscalers will pay for scarce existing nuclear megawatts — the exact premium the Constellation trade depends on.

Materials handed back their scarcity story. A worker death shut Escondida, the world's largest copper mine, on Thursday; it restarted Friday (⁹). The structural deficit is untouched, but the week's tightening impulse is gone. Meanwhile hopes of reopening the Strait of Hormuz pulled oil down and put a risk-on bid under stocks (¹⁰). Two of the week's three energy-security premiums deflated at once.

One number I can't reconcile. Monday's Goldman figure was $1.4 trillion of hyperscaler capex by 2027. Friday's Bloomberg headline from the same bank reads $1.2 trillion, up 50%. Different scope, or a quiet trim — I can't tell from the headline, and it matters, because that single number feeds VRT, ETN and GEV. Flagging it, not pricing it.

Where the Theses Moved

VRT — down. The pillar is "order growth tracks hyperscaler capex." The week's new information is that capex-in-a-contract and capex-actually-spent are now visibly different things. The $25B quantification is the proof. Small offset: VRT is buying fluid-management firm King Environmental, which deepens the liquid-cooling franchise but doesn't touch the demand question.
BE — down. Bloom's marquee relationship is the 2.8GW Oracle fuel-cell order, tied to the very site Oracle is writing itself out of. Orders-convert-to-revenue is the pillar under strain, and there was little room left below to cut.
WULF — down, reversing mid-week. Wednesday's "energized beats permitted" logic pushed it up: Lake Mariner is live and outside Texas. Friday pushed back — its pitch is that signed hosting contracts are bankable revenue, and Oracle just showed those contracts have give. The physical advantage is intact; the contract-quality assumption isn't.
CEG — up, but only on schedule. The NRC issued an environmental finding of no significant impact for the Crane Clean Energy Center restart, with DOE's financing arm cooperating (¹¹). Real progress, but expected news arriving on time. The Google uprate deal is the bigger medium-term drag on CEG's pricing power.
ISRG — up. Tesla is building several hundred Optimus humanoids a week, roughly 10x its Q2 rate, and they still can't handle general tasks (¹²). The same day, Agility — maker of the Digit humanoid — said it's exploring wheels. The field is scaling manufacturing while retreating on capability. ISRG stays the only robotics business on earth with proven, profitable unit economics.
Own-the-Bottleneck — down at the margin. Escondida reopened, Hormuz hopes deflated the oil premium, and Oracle added counterparty risk on the demand side. Every confirming item Friday was a repeat; every disconfirming one was new. The thesis isn't broken, but 0.88 conviction is priced for one-sided news flow and this week wasn't that.
FSLR — down. Arevia, a *solar* developer, is pitching a gas-fired data center on federal land in Idaho (¹³). When solar developers pick gas to serve AI load, the "cheap fast solar fills the gap" pillar is being contradicted by the people best placed to prove it.

Next Week — Prediction Calls

Each resolves on public reporting by Friday, Oct 2 close.

A second escape hatch surfaces (P = 0.55). A named AI tenant other than Oracle is reported renegotiating, delaying, or contractually hedging a data-center lease or large equipment order. This is the highest-value call of the set: one bad project is noise, two is a repricing of the entire landlord model. Bears on WULF, IREN, BE, and every neocloud with a "signed contract" slide.
Blue Owl responds publicly (P = 0.50). Blue Owl or a Jupiter partner disputes, clarifies, or discloses its answer to the force majeure notice. If the response shows the notice was narrow and site-specific, my VRT/BE/WULF downgrades were an overreaction and I'll say so.
No ERCOT thaw (P = 0.88). Texas does not resume approving new large-load interconnections; the December audit stays the gate. Keeps IREN's Childress timeline pushed out and keeps the premium on already-energized sites like WULF's Lake Mariner.
Another US jurisdiction gates data centers (P = 0.80). At least one new county or state announces a moratorium, size cap, or cost-allocation rule. Five did it in five days; the template is spreading, and each one raises the value of interconnected existing capacity while raising the cost of every greenfield plan.
Copper holds $10,000/t (P = 0.92), but drifts lower (P = 0.60). LME 3-month copper closes no day below $10,000/t, yet finishes Oct 2 below its Sep 25 close as Escondida ramps back. The deficit survives a mine reopening; the near-term squeeze doesn't. Bears on SCCO and FCX — hold, not buy.
Humanoids stay unproven (P = 0.90). No humanoid maker announces generalized task capability validated by a paying third-party customer. Direct support for the ISRG upgrade: production volume is not capability.
Suncatcher flies (P = 0.70). Google confirms the launch of its orbital TPU test (¹⁴). Not investable yet, but it's the clearest signal of how seriously hyperscalers take the terrestrial power ceiling.

What would force a rethink: a Blue Owl statement showing the force majeure was narrow and site-specific — calls 1 and 2 collapse and the equipment-maker downgrades were wrong. Or an ERCOT-style freeze landing in PJM territory, which sends Own-the-Bottleneck back up hard and takes CEG, VST and TLN with it.

Thesis Impact

The week in one line

The constraint moved four times in five days: from megawatts → permission → money → counterparty. Monday it was Loudoun County's 12-month pause. Tuesday it was a $1.1B junk bond funding a CoreWeave site and capex eating nearly all hyperscaler operating cash flow. Wednesday it was Abbott freezing ERCOT hookups and Newsom making California data centers pay their own way. Thursday and Friday it was Oracle sending a force majeure notice on a 2.5GW Stargate campus — and Morningstar putting a number on it: up to $25B of delayed revenue.

The through-line. Every bullish thesis on this desk rests on one assumption: a signed order from a hyperscaler equals future revenue. Oracle just demonstrated in public that it doesn't — and over *construction delays*, not demand weakness. That makes it structural, not cyclical. The cross-domain read holds: the buyer's escape hatch has a mirror image — the seller's cost of saying yes (permission, interconnect, wildfire liability, transmission cost allocation) is rising faster than the contract price.

Power. Gating replaced building as the policy tool of choice. Four jurisdictions in five days (Loudoun, California, Texas, Prince William) moved by restricting, not by adding capacity. That's still net-bullish for anyone holding energized, interconnected megawatts and bearish for anyone holding a permit application. But Google's 96MW Vogtle/Hatch uprate deal quietly caps the ceiling — it's a cheap third lane between "wait for new reactors" and "overpay a merchant."

Materials. Copper gave back its scarcity headline. A worker death shut Escondida Thursday; it restarted Friday. The structural deficit is untouched, but the week's tightening impulse is gone, and Hormuz reopening hopes pulled oil down and put a risk-on bid under equities.

The one unresolved tension. Monday's Goldman figure was $1.4T by 2027. Friday's Bloomberg headline reads $1.2T, up 50%. Different scope or a quiet trim — I can't tell from the headline, and it matters, because that single number feeds VRT, ETN and GEV. Flagging it, not pricing it.

---

VRT | Conviction: DOWN | Surprise: MED | Oracle's force majeure got quantified — up to $25B of revenue potentially pushed out. VRT's pillar 2 is "order growth tracks hyperscaler capex"; the new information is that capex-in-a-contract and capex-spent are now visibly different things. CONTRADICTS. | ⁶
Own-the-Bottleneck | Conviction: DOWN | Surprise: MED | Three NEW items all cut the same way: Escondida restarted (a bottleneck easing), Hormuz reopening hopes pulled oil down (energy-security premium easing), and the Oracle precedent adds counterparty risk on the demand side. Every confirming item today was tagged recurring. The thesis isn't broken, but 0.88 is priced for one-sided news flow and this week wasn't. CONTRADICTS at the margin. | ¹⁰
ISRG | Conviction: UP | Surprise: MED | Tesla is building several hundred Optimus units a week — 10x Q2 — and they still can't generalize. Same day, Agility (Digit) said it's exploring *wheels*. Both NEW, both T2. The humanoid field is scaling manufacturing while retreating on capability, which is exactly the yardstick argument: ISRG remains the only scaled, profitable robotics economics on earth. CONFIRMS pillar 1. | ¹²
BE | Conviction: DOWN | Surprise: MED | The $25B quantification lands directly on the 2.8GW Oracle order — BE's marquee relationship, tied to the site Oracle is now writing itself out of. Little room left below 0.05, but the direction is unambiguous. CONTRADICTS pillar 2 (orders convert to deployed revenue). | ¹⁵
WULF | Conviction: DOWN | Surprise: MED | Wednesday's "energized beats permitted" call pushed WULF up; Friday's quantified escape hatch pushes back. Pillar 2 is "signed HPC hosting contracts ramp on schedule" — the Oracle precedent says a signed contract with an investment-grade AI buyer has give in it, and WULF is a far weaker counterparty than Blue Owl. Lake Mariner's physical advantage is intact; the contract-quality assumption is not. CONTRADICTS pillar 2. | ¹⁵
FSLR | Conviction: DOWN | Surprise: MED | Arevia — a *solar* developer — is pitching a gas-powered data center on federal land in Idaho. NEW. When solar developers choose gas to serve AI load, pillar 1 ("cheap, fast solar helps fill the AI power gap") is being contradicted by the people best positioned to prove it. Small but real. CONTRADICTS. | ¹³
CEG | Conviction: UP | Surprise: LOW | NRC issued an Environmental Assessment and Finding of No Significant Impact for the Crane Clean Energy Center — with DOE's Office of Energy Dominance Financing as cooperating agency. T1, and a discrete regulatory clearance on pillar 3 (restart by Q3–Q4 2027). But Crane progress is explicitly in the prior, so this is expected news arriving on time, not a surprise. CONFIRMS. | ¹¹
SCCO | Conviction: HOLD | Surprise: MED | Escondida is back after the fatality halt, easing the squeeze that low Chinese inventories and the approaching holiday pause had created. NEW and genuinely disconfirming for *near-term* price — but it touches none of SCCO's three pillars (structural deficit, cost position, Peru/Mexico output). The deficit thesis survives a mine reopening. Neither confirms nor contradicts; it removes a tailwind. | ⁹
FCX | Conviction: HOLD | Surprise: LOW | Same Escondida item, same read — marginal price pressure, no Grasberg or Indonesia news. FCX's idiosyncratic risk (pillar 3) went untouched all week. Noise. | ¹⁶
LEU | Conviction: HOLD | Surprise: LOW | Pezeshkian offered to dilute Iran's uranium and return to the June MOU. NEW, T2, and directionally it softens the energy-security narrative — but Iranian HEU is not a source of Western SWU, and none of LEU's four pillars (HALEU monopoly, Russian ban, DOE funding, SMR pipeline) are touched. Watching in case de-escalation broadens. Neither. | ¹⁷
IREN | Conviction: HOLD | Surprise: LOW | Nscale raised $3.36bn ahead of its IPO. Cuts both ways — neocloud capital is still abundant (good for the sector's funding story, pillar 3), but it's also a better-capitalized competitor bidding for the same AI tenants. Nothing NEW on Childress or the ERCOT freeze that drove Wednesday's downgrade. Neither. | ¹⁸
ETN | Conviction: HOLD | Surprise: LOW | Goldman's $1.2T capex headline is confirming for pillar 1, but it's the same story we've run all week — and I can't reconcile it against Monday's $1.4T-by-2027 figure from the same bank. ETN sells against short-cycle POs, which insulates it from the Oracle lease problem. Nominally CONFIRMS; I'm not moving on an unreconciled number. | ¹⁹

---

Prediction calls — week of Mon Sep 28 to Fri Oct 2

Each resolves on public reporting by Friday Oct 2 close.

1. Another US jurisdiction gates data centers. At least one new county or state announces a moratorium, size cap, or cost-allocation rule. — P = 0.80

2. No ERCOT thaw. Texas does not resume approving new large-load interconnections; the December audit stays the gate. — P = 0.88

3. A second escape hatch surfaces. A named AI tenant other than Oracle-Jupiter is reported renegotiating, delaying, or contractually hedging a data-center lease or equipment order. This is the highest-value call of the set — it's the difference between one bad project and a repricing of the whole landlord model. — P = 0.55

4. Blue Owl pushes back publicly. Blue Owl or its Jupiter partners dispute, clarify, or disclose a response to the force majeure notice. — P = 0.50

5. Copper holds the floor. LME 3-month copper does not close below $10,000/t on any day. — P = 0.92. Secondary: copper closes Oct 2 *below* its Sep 25 close as Escondida ramps back — P = 0.60

6. Oil keeps falling on Hormuz. Brent closes Oct 2 below its Sep 25 close. — P = 0.58 (reverses fast if Iran talks stall)

7. Suncatcher flies. Google confirms launch of the orbital TPU test. — P = 0.70

8. Humanoids stay unproven. No humanoid maker announces generalized task capability validated by a paying third-party customer. Direct support for the ISRG upgrade above. — P = 0.90

What would force a rethink: a Blue Owl statement showing the FM was narrow and site-specific (call 3 and 4 collapse, VRT/BE/WULF downgrades were an overreaction), or a second ERCOT-style freeze in PJM territory (Own-the-Bottleneck goes back up hard, and CEG/VST/TLN with it).

Inflection Radar

Synthesis: Macro Through-Line

The week confirms three dominant, interconnected themes: Grid Constraint, Autonomous Systems Maturation, and Specialized Compute.

1. Power/Energy (Constraint Focus): The narrative has decisively shifted from *deployment* to *bottleneck management*. Massive federal funding ($1.9B DOE) is aimed at grid modernization (75, 85), but the underlying friction remains acute. Regulatory bodies (FERC) are dealing with interconnection queue backlogs and complex project approvals (69, 84). This suggests that utility-scale deployment speed will be dictated by regulatory/infrastructure capacity, not technology readiness.

2. Compute/AI (Specialization Focus): The academic frontier is moving away from generalized LLM performance toward highly specialized, resource-aware models. Key areas include spatial data inference (76), optimizing complex reasoning paths (78), and enabling real-time physical interaction via asynchronous robotics (79, 80). This signals a compute market bifurcation: massive infrastructure buildout vs. hyper-efficient edge/application AI.

3. Geopolitics/Defense (Autonomy Focus): The defense sector is rapidly integrating autonomous capabilities across the stack—from C2 software (Anduril/Origin MoU, 73) to dedicated counter-UAS systems (74), and platform integration (RASWDC establishment, 63). This points toward a militarization of AI capability that requires resilient, multi-sensor data fusion.

Thesis Shifts:

Compute: Shifted from general "AI boom" hype to specific architectural challenges (e.g., memory stacking, asynchronous processing) and domain specificity (spatial biology).
Energy: Shifted from pure renewable capacity targets to the *physical ability* of the grid to handle new loads, making transmission and interconnection a primary risk factor.
Defense: Solidified focus on layered, networked autonomy rather than single-system breakthroughs.

***

Prediction Calls for Next Week (Falsifiable)

1. The DOE/FERC will issue guidance or proposals explicitly linking critical mineral sourcing (92) to grid interconnection timelines (85), creating a new regulatory choke point between materials and power infrastructure.

2. A major academic paper (arXiv, 76-81 range) will bridge the gap between physical robotics control (e.g., tactile sensing, 80) and large foundation models, moving beyond simulation into real-world deployment constraints.

3. The market will show increased focus on "de-risking" technology supply chains, evidenced by procurement signals (91) or government investment in domestic material stockpiles (92).

***

Emerging Signals

[emergent] Cross-Domain AI/Robotics: Combining advanced tactile learning for locomotion with real-time environmental data processing. | Touches: NEW | ²⁰

[emergent] Materials/Supply Chain Friction: Persistent, localized procurement bottlenecks (copper scrap invoices) signal systemic friction that will slow industrial adoption regardless of macro funding announcements. | Touches: NEW | [²¹)

[dismissive] AI/Compute Over-Optimization: The focus on highly specific, complex inference methods (e.g., NoThink leakage, 78) suggests a potential over-indexing on theoretical model efficiency at the expense of practical, general deployment utility. | Touches: N/A | ²²

QA & Caveats

VRT contradicts its pillar 2 regarding order conversion versus revenue based on the force majeure quantification.
BE and WULF contradict their pillar 2 assumptions regarding the conversion of orders into deployed revenue following the Oracle precedent.
FSLR directly contradicts its pillar 1 by noting solar developers are choosing gas for AI load, undermining the "cheap solar" thesis.

Sources

datacenterdynamics.com datacenterdynamics.com
bloomberg.com bloomberg.com
datacenterdynamics.com datacenterdynamics.com
datacenterdynamics.com datacenterdynamics.com
Oracle sends force majeure notice on its New Mexico Stargate data center techcrunch.com
Oracle: Data Center Force Majeure Implies Possibility of $25 Bln Revenue Delay - Morningstar news.google.com
Northern Virginia's Prince William County puts controls on new data center projects datacenterdynamics.com
datacenterdynamics.com datacenterdynamics.com
BHP’s Escondida restart eases tight copper market mining.com
US Stocks Gain as Oil Drops on Hormuz Reopening Hopes bloomberg.com
Constellation Energy Generation, LLC; Christopher M. Crane Clean Energy Center; Environmental Assessment and Finding of federalregister.gov
Tesla ramps Optimus to hundreds a week, but the robots can’t generalize electrek.co
Another Solar Company Trying Gas-Powered Data Center on Federal Lands heatmap.news
Project Suncatcher: Google to launch first space data center test in orbit next week datacenterdynamics.com
Oracle force majeure on US data center showcases AI infrastructure risks - Semafor news.google.com
BHP’s Escondida restart eases tight copper supply woes northernminer.com
Pezeshkian Offers to Dilute Uranium, Calls for Return to June MOU - Radio Free Europe/Radio Liberty news.google.com
AI cloud and data center firm Nscale raises $3.36bn ahead of its IPO datacenterdynamics.com
Goldman Sees Hyperscaler AI Capex Rising 50% to $1.2 Trillion - Bloomberg news.google.com
TactileStep: Sole Tactile Learning for Regulating Foot-Terrain Interaction in Humanoid Locomotion arxiv.org
Insufficient Invoices Hinder Scrap Utilization Enterprises' Procurement, Pre-Holiday Stockpiling Volumes Severely Inadeq news.google.com
Thinking Leakage: A Causal Audit of NoThink Post-Training in Hybrid Reasoning Models arxiv.org