Morning Analysis — 2026-09-21

Dek

Goldman just raised the price tag on the AI buildout to $1.4 trillion — and the hard part isn't the chips, it's finding the megawatts, the metal, and the permission to plug it all in.

The Big Shift

Goldman Sachs now models US hyperscaler capital spending — what Amazon, Microsoft, Google and Meta spend on data centers and gear — at $1.4 trillion by 2027, above prior estimates (¹). That matters because it is the single input that feeds almost every thesis on this desk: Vertiv's order book is explicitly a function of hyperscaler capex, and the power, cooling and copper that money buys is the bottleneck everyone else is trying to own. It signals the demand-source risk — the worry that the spending was a one-off spike — is receding, and shifts the question from *will they spend* to *can the physical world absorb it*.

Analysis

The capex number turns into megawatts, everywhere at once. Overnight we got a 100MW groundbreaking in Madrid (²), a 200MW-by-2028 target in Oman (³), and Rio Tinto supplying low-carbon aluminum for the cabling inside an Amazon site in Ohio (⁴). None of these are big on their own. Together they confirm the Own-the-Bottleneck premise: the constraint is grid interconnection and physical inputs, not demand. No new information — but a drumbeat this steady is itself the signal.

Buyers are now paying for power that doesn't exist yet. Data-center M&A is increasingly priced with *earnouts* — deal structures where part of the purchase price is only paid if the site actually gets energized (⁵). Read it two ways. The bullish read is that unenergized power rights have become a tradeable asset, which is exactly the IREN/WULF pipeline bet. The cautious read is that sellers are being asked to carry the risk — capital is getting pickier about paying upfront for a promise. Watch which way that tilts as deals print.

On the supply side, distributed storage just proved it can show up for a peak. Sunrun and Tesla pulled a record 580MW from 140,000-plus home batteries onto the California grid during a September 9 heat wave (⁶) — roughly a mid-sized gas plant, assembled from garages. That confirms the broader "cheap, fast power fills the gap" thesis behind FSLR's pillar 1. But be precise: this is residential solar-plus-batteries, not utility-scale panels. It validates the category without touching First Solar's order book or pricing. Category win, not a stock win.

The materials thread is quietly getting more crowded — which is a problem for MP. Bridge Green and Hartree signed a $500M+, eight-year recycled-lithium supply deal with an equity investment attached (⁷). Separately, Washington has been taking equity stakes in *foreign* mines — rare earths in Brazil, graphite in Mozambique (⁸). The article's own argument is about weak governance on those deals, not a policy retreat. But the fact pattern mildly cuts against the "the US has no alternative but to fund MP" framing. A thread to watch, not a downgrade.

Fuel and permission, the other two bills. Oil fell back toward $100 — its longest losing streak since June — as Hormuz flows held up and diplomacy on the US-Iran war advanced, with satellite data showing Saudi Gulf loadings jumping over the weekend (⁹). That eases one of Bloom Energy's thesis-breaking triggers, though $100 is still not cheap and Bloom's real problem is converting pipeline to contracts, not fuel cost. On permission: the NRC is out for public comment on materials guidance for advanced (non-light-water) reactors (¹⁰) — dull, routine, and exactly the regulatory scaffolding LEU's demand curve depends on. Meanwhile five separate LocalLLaMA threads circled the same point — consumer VRAM is capped near 12–16GB, and one developer spent three weeks of agent-run work building custom CUDA kernels that still lost to llama.cpp (¹¹). That argues *against* a near-term inference-efficiency step-change, which is the one thing that would deflate the whole power-demand story. Low-tier chatter, no corroboration, nothing moves — but it's the right thing to be watching.

What Would Prove Us Wrong

• A hyperscaler trims or defers capex guidance at Q3 earnings (late October). Concretely: any of the big four guiding 2027 capex below consensus, or language shifting from "accelerating" to "optimizing." That hits VRT pillar 2 directly — orders track capex — and undercuts Own-the-Bottleneck's premise that demand is the sure part.
• A credible, corroborated drop in serving cost per token at scale. Not a Reddit benchmark: a published inference-cost figure from a major lab or cloud showing a step-change (call it >3x) in tokens per watt. That would break the shared power-demand assumption under CEG, VST, ETN and Own-the-Bottleneck at once.
• Another US government equity stake in a non-MP rare earth supplier, or a DoD offtake awarded outside MP. Each one erodes the exclusivity argument in MP's pillar 1. Two more and the "only domestic option" framing is no longer true.

Thesis Impact

• VRT | Conviction: UP | Surprise: MED | Goldman now models US hyperscaler capex at $1.4T by 2027 — a bigger number than the prior assumed, and Vertiv's pillar 2 is literally "orders track hyperscaler capex." CONFIRMS; at P=0.37 the demand-source risk is the main thing holding conviction down, and this lifts it. | ¹
• Own-the-Bottleneck | Conviction: HOLD | Surprise: LOW | $1.4T capex plus a steady drumbeat of new builds (Madrid 100MW, Oman 200MW by 2028, Amazon Ohio cabling) is exactly the "rising DC load + hyperscaler power deals" the prior already expects. CONFIRMS pillar 1, no new information. | ⁴
• FSLR | Conviction: HOLD | Surprise: MED | Sunrun/Tesla dispatched a record 580 MW of home batteries to the California grid in a heat wave — solar+storage proving it can answer a peak. CONFIRMS pillar 1 (cheap, fast power fills the gap), but it's distributed residential, not utility-scale modules, so it validates the category without touching FSLR bookings or ASPs. | ⁶
• MP | Conviction: HOLD | Surprise: MED | Washington is putting equity into *foreign* critical-mineral supply — a rare earth mine in Brazil, graphite in Mozambique — which cuts against the "US must fund MP because there's no alternative" framing. Mildly CONTRADICTS pillar 1's exclusivity; the piece's actual argument is about weak governance on those deals, not about pulling back, so it's a thread to watch rather than a downgrade at P=0.20. | ⁸
• IREN | Conviction: HOLD | Surprise: LOW | Data-center M&A is increasingly priced with earnouts on future megawatts — the market is paying for power rights before they're energized, which is the whole IREN/KEEL/WULF pipeline bet. CONFIRMS pillar 1, but it's an opinion piece with no deal terms, so nothing moves. | ⁵
• BE | Conviction: HOLD | Surprise: LOW | Saudi Gulf loadings jumped and oil fell back toward $100 on robust Hormuz flows — easing the energy-price spike that is one of Bloom's thesis-breaking triggers. Weakly CONFIRMS pillar 3, but it's oil not gas, and recurring; at P=0.06 the problem is contract conversion, not fuel cost. | ⁹
• LEU | Conviction: HOLD | Surprise: LOW | NRC out for comment on materials-compatibility guidance for non-light-water reactors — routine regulatory groundwork for the advanced-reactor pipeline that is LEU's demand curve. CONFIRMS pillar 4; recurring, already in the prior. | ¹⁰

Watch, no move: five T3 LocalLLaMA threads on VRAM limits and inference engines point at the shared "inference-efficiency step-change" trigger under CEG/VST/ETN/Own-the-Bottleneck. One of them — three weeks of agent-built CUDA kernels that still lost to llama.cpp — argues *against* a step-change. T3 only, no corroboration, so nothing moves. (¹¹)

Inflection Radar

[emergent] AI Interpretability & Reasoning | New arXiv work moves beyond simple inference testing to focus on whether models truly internalize the *intent* of a task, signaling a shift toward AGI benchmarks that test understanding rather than pattern matching. | Touches: NEW | ¹²

[emergent] AI Compute Efficiency | Advances in Radius-Bounded Sparse Prefill address the core computational bottleneck (prefill phase) limiting context window size for LLMs, suggesting near-term performance gains for massive models. | Touches: NEW | ¹³

[emergent] Quantum Scaling Laws | Direct comparison of RydbergGPT's scaling behavior against established transformer laws provides a quantifiable, early assessment of quantum computing's viability as a general-purpose accelerator. | Touches: NEW | ¹⁴

[emergent] Geopolitical Capital Flight | Institutional investors are rapidly stockpiling cash and retreating from US/UK equities due to persistent inflation and mounting geopolitical risk, signaling a potential systemic liquidity shift away from developed markets. | Touches: NEW | ¹⁵

[emergent] Data Center Capitalization | Data center spending has surpassed housing as a primary driver of US capital expenditure, signaling that AI compute infrastructure is now fundamentally reshaping real estate and industrial investment cycles. | Touches: NEW | ¹⁶

[emergent] AI Energy Demand Coordination | Major players (Nvidia, Google, Emerald AI) forming an alliance to manage and promote demand flexibility signals that energy grid stability is becoming a core operational constraint for next-gen compute deployment. | Touches: NEW | ¹⁷

[emergent] Defense Contract Friction | Oasis challenging the US Army's C-UAS contract award decision highlights persistent legal and policy friction points in critical defense procurement areas, suggesting potential delays or restructuring of awarded contracts. | Touches: NEW | ¹⁸

[emergent] Robotics Embodiment Constraints | New research models robotic learning not just on information, but on the physical limits and dynamics of the body (penalty kicks), advancing robustness in complex, real-world interaction. | Touches: NEW | ¹⁹

[emergent] Advanced Demonstration Synthesis | Using Gaussian Splatting to synthesize high-fidelity training demonstrations from single scans solves a major bottleneck in creating data for complex visuomotor policies, accelerating robotics deployment. | Touches: NEW | ²⁰

QA & Caveats

No issues found.

Sources

1. Goldman Sachs expects U.S. hyperscaler capex to reach $1.4T by 2027 (GS:NYSE) - Seeking Alpha news.google.com
2. Nabiax breaks ground on 100MW data center in Madrid, Spain datacenterdynamics.com
3. Oman’s O-Green targets 200MW of data center capacity by 2028 datacenterdynamics.com
4. Prysmian, Rio Tinto to supply Amazon data center in Ohio with low-carbon aluminum datacenterdynamics.com
5. Pricing future megawatts: the growing role of earnouts in data center M&A datacenterdynamics.com
6. Tesla and Sunrun’s virtual power plant dispatch record 580 MW to California grid electrek.co
7. Bridge Green and Hartree Announce Transformational $500 Million+ Agreement for Recycled Lithium to Strengthen Domestic C prnewswire.com
8. Washington’s Foreign Equity Bets Need Better Governance warontherocks.com
9. Satellite Snapshot Points to Increase in Saudis’ Gulf Oil Flows bloomberg.com
10. Draft Interim Staff Guidance: Material Compatibility for Non-Light Water-Reactors federalregister.gov
11. The bear can dance: Qwen 3.8 27B on one 3090 for 3 weeks reddit.com
12. Implicit Rule Induction with Test-Time Task Embeddings in ARC-like Tasks arxiv.org
13. RBS-Attention: Radius-Bounded Sparse Prefill for Long-Context Large Language Models arxiv.org
14. Do Quantum Models Scale Like LLMs? arxiv.org
15. Geopolitical risks see institutional investors rush to stockpile cash, survey finds scmp.com
16. news.google.com news.google.com
17. AI Heavyweights Team Up to Promote Demand Flexibility heatmap.news
18. Oasis challenges US Army C-UAS contract award decision - Unmanned airspace news.google.com
19. Dynamics-Induced Commitment in Learning-Based Robotic Penalty Kicks arxiv.org
20. Demonstration Synthesis from a Single Scan via Gaussian Splatting for Visuomotor Policy Learning arxiv.org