Morning Analysis — 2026-07-05

Dek

The oil-shock scare that shadowed every power thesis just deflated overnight — while the real chokepoint stays exactly where it was: the grid interconnect.

The Big Shift

Tankers are moving through the Strait of Hormuz again along the US-protected Oman-side corridor, a day after a batch of unexplained U-turns spooked energy markets, and OPEC+ is set to approve another output increase into the recovery (¹). That matters because a sustained Hormuz disruption plus a supply squeeze was the one macro tail risk that could have re-priced diesel, gas, and the whole cost stack under AI power buildouts. With shipping normalizing and more barrels coming, that risk is receding — which points back to the real story: the bottleneck for AI isn't the price of energy, it's getting power to the meter.

Analysis

Power — the grid has stopped being the road in. The sharpest signal isn't news at all, it's a pattern hardening. Data-center sites keep dying at the grid connection — the "20 percent problem," where interconnection kills projects before a shovel moves (²). The response is now structural: financiers are building whole power plants for a single tenant, like National Grid Ventures' $1.75B gas plant dedicated to one 2GW Microsoft campus (³). Implication: the grid-bypass and own-the-bottleneck theses keep getting confirmed — whoever controls firm, behind-the-meter power captures the scarcity rent. SK Telecom's reaffirmed 15GW buildout (⁴) just restates how bottomless the demand side is.

Materials — deficit intact, but the market is punishing froth. Two cross-currents. BHP is spending $1.5B to restart Chile's Cerro Colorado copper mine (⁵) — new supply, a mild dent to the copper-deficit thesis, but the tonnage is modest and won't close a structural gap. Meanwhile Standard Lithium cratered 48% on an overvaluation warning (⁶). Implication: the long-materials thesis (copper, lithium as AI/electrification inputs) is intact on fundamentals, but the crash is a reminder that the market is separating real supply from story stocks — own the producers, not the promoters.

Compute — the one thing that could break every power thesis at once. The efficiency chatter is building: multi-resolution flow matching for cheaper image generation (⁷) and speculation that dSpark/MTP-style tricks could make model spillover-to-disk tolerable (⁸). Implication: these are hobbyist-tier signals today, but they map to the single trigger that would hit power, grid-bypass, and materials theses simultaneously — a durable inference efficiency leap that flattens AI's power-demand curve. Nothing here clears that bar yet. It's a watch, not a move.

Geopolitics & policy — the slow squeeze is local, not global. With the Hormuz macro risk fading, the binding constraint is politics at home. New Jersey nixed a data-center tax credit and is establishing a large-load tariff (⁹); even Texas's Abbott, a longtime booster, now wants rural data centers banned (¹⁰); Amazon's emissions rose 16% on record capacity (¹¹). Implication: the direction is one-way — rising cost and restriction for grid-dependent siting. That's not bearish for AI power; it's the wind at the back of the behind-the-meter, own-your-own-plant model.

What Would Prove Us Wrong

• Inference efficiency leap goes mainstream. A T1/T2 outlet (not Reddit) corroborates a real, durable efficiency step-change — measured as a >30% drop in tokens-per-watt or energy-per-query at frontier scale. That would flatten AI power demand and hit every power thesis at once (Own-the-Bottleneck, grid-bypass/BE, gas-turbine/GEV).
• Copper supply reflates. More restarts like Cerro Colorado stacking up — two or three additional >100kt/yr projects greenlit within a quarter — would start closing the deficit and undercut the long-copper (SCCO/FCX) thesis.
• Energy shock returns. Hormuz transit reverses again or OPEC+ output growth stalls, pushing Brent through a sustained spike. That re-introduces the macro tail risk this morning just removed and would re-price the cost stack under every buildout.

Thesis Impact

No thesis-moving signal.

Nothing overnight clears the bar to change a prior. The material touches are all confirmation already in the priors or too thin to act on — logged below as HOLDs for the day:

• Own-the-Bottleneck | Conviction: HOLD | Surprise: LOW | SK Telecom's 15GW AI data-center buildout reappears — huge demand-side confirm, but it's yesterday's news (recurring 1d), already in the prior. CONFIRMS. | ⁴
• BE | Conviction: HOLD | Surprise: LOW | "20 percent problem" piece argues DC sites keep dying at the grid interconnect — squarely confirms the grid-bypass thesis, but it's an opinion piece with no new data. CONFIRMS. | ²
• GEV | Conviction: HOLD | Surprise: LOW | National Grid Ventures' $1.75B gas plant dedicated to a 2GW Microsoft DC — clean "financier builds a plant for one tenant" datapoint (matches the cross-domain thread), confirms gas-turbine demand, but recurring 2d. CONFIRMS. | ³
• SCCO / FCX | Conviction: HOLD | Surprise: LOW | BHP's $1.5B Cerro Colorado restart is new copper supply — the mild *disconfirm* to the deficit thesis — but modest tonnage and recurring 2d, not enough to touch the deficit. Slight CONTRADICT, immaterial. | ⁵

Watch today: the two inference-efficiency items ⁷⁸ are T3-only chatter now, but they map to the one trigger that would hit *every* power thesis at once — a durable inference step-change flattening AI power demand. If a T1/T2 outlet corroborates a real efficiency leap, that's the disconfirming signal to act on. Nothing there yet.

Inflection Radar

[emergent] Autonomous Military Integration | The Royal Navy's demonstration of combat-proven drone capabilities signals a rapid, deep integration of autonomous systems into core fleet operations, shifting defense spending toward networked, non-traditional platforms. | Touches: NEW | ¹²

[dismissive] Regulatory Erosion in Critical Infrastructure | Supreme Court action undermining the independence of agencies like FERC signals a systemic weakening of regulatory guardrails, increasing political risk and potential instability for large-scale energy and utility investments. | Touches: NEW | ¹³

QA & Caveats

No issues found.

Sources

1. Oil, Gas Tankers Cross Hormuz Via Oman-Side Route After U-Turns bloomberg.com
2. The 20 percent problem: why data center sites fail before they’re built datacenterdynamics.com
3. National Grid Ventures invests $1.75bn in Joulent for construction of gas plant powering 2GW Microsoft data center in Te datacenterdynamics.com
4. SK Telecom Pursues 15GW AI Data Center Buildout, Aiming to Become Asia's AI Infrastructure Hub prnewswire.com
5. BHP leverages $1.5B to restart Chile copper mine northernminer.com
6. Standard Lithium falls 48% after InvestingPro overvaluation warning investing.com
7. [Paper] Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling reddit.com
8. Is dSpark, dflash, MTP, QAT, and similar tech going to increase inference speed enough to where model spillover to disk reddit.com
9. New Jersey lawmakers pass bill to establish large load data center tariff datacenterdynamics.com
10. Texas Governor Abbott calls for data centers to be banned in rural areas datacenterdynamics.com
11. Amazon's carbon emissions grow by 16 percent in 2025, on the back of record data center capacity additions datacenterdynamics.com
12. Royal Navy’s Sea Launch Of Combat-Proven Nyan Kamikaze Drone Points To Fleet’s ‘Hybrid’ Future twz.com
13. FERC Was Already Losing Its Independence. Now It’s Gone. heatmap.news