Morning Analysis — 2026-07-18

Dek

The grid can't add power fast enough for AI — so the hyperscalers stopped waiting and started building their own.

The Big Shift

Bank of America now says AI data-center demand will outrun everything U.S. utilities have *planned* to build by more than 100 gigawatts through 2030 — a gap bigger than the entire grid of most countries ¹. That number matters because it turns "power is the bottleneck" from a talking point into a measured shortfall, and it tells you where the money goes next: on-site gas, batteries, and behind-the-meter deals, because waiting in the utility interconnection queue is no longer an option. The signal for the day — the buildout is quietly routing *around* the public grid, and that reshapes who captures the spend.

Analysis

The clearest proof is in the deal flow. Google just committed to a 2.5 GW solar-plus-storage project in Arkansas ², and smaller players like Duos are locking in on-site capacity for hyperscalers on multi-year contracts ³. Meanwhile the public grid keeps flashing scarcity: PJM's capacity auction hit its price cap *again* ⁴, the market's way of saying supply can't meet demand at any reasonable price. This is the core "own-the-bottleneck" thesis playing out — the constraint is real, it's priced, and it's driving self-supply. It also explains why nuclear license renewals (Ginna ⁵) and firm generators keep bid up: anything that produces steady electrons now has a buyer.

The materials layer is tightening underneath all of this. The IEA says the copper supply outlook has "worsened considerably," with China's export ban and Middle East conflict choking the sulphuric acid needed to leach the metal ⁶, and BMI just lifted its 2026 copper forecast 6.7% to $12,700/tonne on structural deficits ⁷. Copper is the wiring of every data center and every new generator, so a deficit is a tax on the entire buildout. The newer wrinkle: rare earths now have AI as a *demand* driver — data-center cooling, power, and the robots being built to run these facilities all need them, on top of defense and clean energy ⁸. What this implies: the bottleneck isn't just megawatts, it's the physical inputs to build the megawatts, and China controls the choke points.

Geopolitics keeps its thumb on the same scale. Iran is stepping up attacks on shipping in the Strait of Hormuz, and the striking part is that the U.S. Navy hasn't been able to shut it down ⁹ — enough that Iraq, Chevron, and the UAE are all planning physical routes *around* the strait ¹⁰. Same pattern as the grid: when the shared chokepoint fails, everyone builds their own bypass. For the energy thesis, sustained Hormuz risk keeps a floor under oil and gas prices, which raises the cost of the on-site gas turbines the data centers are now leaning on.

On compute financing, the first GPU-backed lenders are pivoting to inference chips in a $400M chip-collateralized loan ¹¹. This is a tell that the market believes inference — running models, not training them — is where sustained demand lives, which supports the durable-compute-demand thesis. But watch the other edge: Cerebras is reportedly "winning the inference race" on efficiency ¹². If inference gets radically cheaper per query, that's the one thing that could flatten the power-demand curve the whole thesis rests on. Needs corroboration before it moves anything.

The one thesis actually moving today is ISRG: the stock plunged on its 7/16 earnings ¹³. The headline doesn't name the driver, but a sharp post-print drop on the sector's quality benchmark is a disconfirming reaction that pressures Pillar 2 (procedure growth of 13.5–15.5%). Conviction nudges down until we see *why* — the 10% procedure-growth line is the trigger to watch.

What Would Prove Us Wrong

• Inference efficiency flattens power demand. If Cerebras-style efficiency gains ¹⁴ get T1/T2 confirmation — a hyperscaler publicly cutting its power-per-query or trimming a capex forecast — the 100 GW gap narrows and the own-the-bottleneck thesis weakens. Watch for a hyperscaler guiding *down* on data-center energy intensity.
• ISRG procedure growth breaks 10%. If the earnings driver turns out to be procedure-volume deceleration (not one-off guidance or China), and growth prints below ~10%, that hits Pillar 2 directly and takes conviction from DOWN toward exit.
• Copper deficit resolves. If China lifts its export controls or the copper price rolls back below ~$11,000/tonne on easing supply, the "materials tax on the buildout" thread loses force — a sign the physical-input constraint was overstated.

Thesis Impact

ISRG | Conviction: DOWN | Surprise: MED | Stock plunged today, the market's verdict on the 7/16 earnings (8-K). The headline doesn't state the driver, and I couldn't open the article to confirm whether it's procedure-growth deceleration, guidance, or China — but a sharp post-print drop on the sector's benchmark quality name is a disconfirming reaction that pressures Pillar 2 (procedure growth 13.5–15.5%). Nudges conviction down pending the "why"; watch for the 10% procedure-growth trigger. | ¹³

_Note: everything else in today's feed is tagged recurring (copper, Hormuz, data-center buildout, hyperscaler-capex, nuclear-decommission) — already in the prior, LOW surprise, HOLD. The one T3 lead worth flagging but not acting on: "Cerebras winning the inference race" ¹⁴ touches the AI inference-efficiency/demand-flattening trigger shared across CEG/VST/ETN/VRT/Own-the-Bottleneck — needs T1/T2 corroboration before it moves anything._

Inflection Radar

[dismissive] AGI Hype Cycle Correction | The refusal by key players to label their work "AGI" signals a necessary technical distinction from current hype cycles, suggesting focus must remain on verifiable capability rather than conceptual milestones. | Touches: NEW | ¹⁵

[emergent] Compute Infrastructure Bottlenecks | FERC's increasing focus on interconnection "seams" and capacity auction stress highlights that the primary constraint for AI buildout is not compute power, but grid access and regulatory bandwidth. | Touches: NEW | ¹⁶

[emergent] Global Compute Capital Deployment | Large foreign capital commitments (e.g., Indonesia) for AI data centers and network infrastructure indicate a pattern of compute buildout bypassing traditional Western markets, creating new regional power nodes. | Touches: NEW | ¹⁷

QA & Caveats

• ISRG call is an interpretation of market reaction; no direct support in candidates.
• The "Cerebras winning" point is a T3 signal needing T1/T2 corroboration before action.

Sources

1. AI data center growth could force US utilities to rethink generation plans, BofA says utilitydive.com
2. Google inks deal for massive Arkansas solar and storage project utilitydive.com
3. Duos Edge AI to provide 10MW of capacity for five years to “investment-grade hyperscaler” datacenterdynamics.com
4. PJM Once Again Hits Its Price Cap at Latest Auction heatmap.news
5. Constellation Energy Generation, LLC; R.E. Ginna Nuclear Power Plant; Subsequent License Renewal Application federalregister.gov
6. northernminer.com northernminer.com
7. northernminer.com northernminer.com
8. AI emerges as new driver of rare earth demand, Sprott says mining.com
9. A naval failure in the Strait of Hormuz breakingdefense.com
10. Iraq and Chevron Are Planning a Route Through Syria to Evade Strait of Hormuz heatmap.news
11. Why the first GPU financiers are turning to inference chips in a $400 million deal techcrunch.com
12. seekingalpha.com seekingalpha.com
13. Why Intuitive Surgical Stock Plunged Today - Investopedia news.google.com
14. Cerebras Is Winning The Inference Race, Yet I'm Still Not Buying seekingalpha.com
15. Why AMI Labs’ Alexandre LeBrun won’t call his AI ‘AGI’ or ‘superintelligence’ techcrunch.com
16. PJM capacity auction results compound ‘alarm bells’: FERC Chairman Swett utilitydive.com
17. Agencia Comercial Spirits Ltd Announces Agreements for Indonesia AI Data Center Construction, Network Infrastructure Pro news.google.com