Evening Analysis — 2026-06-25

Dek

A tanker burns in Hormuz, OpenAI tapes out its own chip, and four theses end the night exactly where they started — the loudest news was the least relevant, and the quiet news was the one to watch.

The Big Shift

The day's biggest *structural* move wasn't the oil drama — it was the chip market turning against Nvidia on two fronts at once. OpenAI unveiled "Jalapeño," its first custom inference chip built with Broadcom, while Qualcomm rolled out its Dragonfly data-center line with Meta already signed up as a buyer (¹). Why it matters: these chips are pitched on efficiency — more AI output per watt — and the industry is now openly arguing that "tokens per watt" should be the metric that counts (²). That points at the one thing that could undermine the whole AI-as-power-demand story: if inference gets radically cheaper to run, the electricity curve everyone is building toward bends downward.

Analysis

Power (CEG) — confirming, but on autopilot. Walmart signed a nuclear offtake deal with Constellation (³), a clean hit on the "turn uncontracted nuclear into long contracts" pillar, and the policy tailwind kept blowing — DOE's $17.5B nuclear loan package and a swelling small-reactor pipeline (Deep Fission now claims 18.5GW of interested customers, ⁴). But all of this is recurring and already priced into our view. The thesis is intact and getting *more* confirmed, just not in a way that earns a re-rate tonight.

Compute vs. power — the real tension. Here's the thread that cuts against CEG and the broader power trade: custom silicon (Jalapeño, Dragonfly) plus software tricks like speculative decoding (a method that drafts several tokens at once — one paper, JetSpec, claims up to ~9.6x speedups) are all aimed at squeezing more AI work out of each watt. If they deliver, the data-center electricity demand that justifies new nuclear and grid build-out grows slower than the consensus assumes. None of this is proven at scale yet — the efficiency leads are still lab-grade (T3) with no hard corroboration — but it's the first credible counter-current to the "AI eats the grid" thesis, and it's worth tracking as a cluster, not as one-off headlines.

Materials (SCCO, FCX) — supply drumbeat hit a crescendo. Generation Mining all but closed financing on its ~$969M Marathon copper-palladium project (⁵), capping a week of supply news (Anglo–Codelco unlocking +2.7Mt, NexMetals +63% resource, Faraday drilling). Layer on SCCO's own 8-K today — it took on material debt and entered a material agreement — and you have a steady accumulation of evidence against the copper-deficit pillar. But none of it is the trigger that would actually break the thesis: this is all long-dated, brownfield (expansions of existing mines), small relative to a ~22Mt/yr market. The right read is patience: log the accumulation, don't move.

On-site power (BE) — the watch-item showed up again. FuelCell Energy's up-to-380MW deal to power data centers (⁶) cuts both ways for Bloom: it confirms that on-site fuel cells for data centers are a real, growing market, but it's also exactly the competitive threat we flagged. Recurring, so already in the prior — but a second data point in the "credible rival winning DC contracts" column.

Geopolitics — loud, off-thesis. The Hormuz cluster dominated the wires: a cargo ship was hit, the UN paused its escort plan, and ships U-turned out of the strait (⁷) — all this just after the US Energy Secretary declared Iran's ability to close the waterway "ended." Yet oil round-tripped back to pre-war levels (⁸). The implication for us is mostly what it *doesn't* do: a live shooting incident in the world's most important oil chokepoint failed to produce a sustained price shock. We hold no oil; the signal is that the energy-security premium isn't transmitting to prices, which keeps the macro backdrop calm for everything else.

What Would Prove Us Wrong

• Efficiency bends the power curve (hits CEG + the whole power-demand trade). Watch for a hyperscaler cutting or flattening its capex/power guidance, or a credible "watts per token" figure dropping sharply as custom silicon ships in volume. If inference gets cheap to run, the electricity-demand story that underwrites nuclear offtake weakens. Concrete trigger: any top-four hyperscaler guiding data-center power growth *down* year-over-year.

• Real new low-cost copper, not brownfield (hits SCCO + FCX). Tonight's supply news was all expansions and long-dated projects. The thesis breaks if a large *greenfield* low-cost mine (>250kt/yr, near-term first production) gets sanctioned. That's the "new cheap supply online" trigger we haven't seen — the brownfield drumbeat alone doesn't qualify.

• A rival underprices Bloom on a data-center deal (hits BE). Watch for FuelCell or another competitor signing a repeat DC offtake at a cost clearly below Bloom's $/kW, or converting that 380MW framework into firm, priced orders. One signed PPA at a lower price point would turn the "watch-item" into a thesis problem.

Thesis Impact

SCCO | Conviction: HOLD | Surprise: LOW | Generation Mining locks in financing (~$969M) for its Marathon copper-palladium project, the latest brick in a week-long copper-supply drumbeat (Anglo–Codelco +2.7Mt, NexMetals +63% resource, Faraday). Mildly CONTRADICTS the structural-deficit pillar, but all long-dated/brownfield and mostly recurring — not the "large new low-cost supply online" trigger. Watch the accumulation, don't move yet. | ⁵

FCX | Conviction: HOLD | Surprise: LOW | Same copper-supply pipeline news; marginal CONTRADICT to the deficit thesis. 135kt/yr-class additions over decades don't dent a ~22Mt/yr market — prior intact at its already-cautious P=0.48. | ⁹

BE | Conviction: HOLD | Surprise: LOW | Competitor FuelCell Energy signs up-to-380MW supply deal to power data centers (first 30MW this year). Cuts both ways: CONFIRMS the on-site-fuel-cell-for-DC category is real demand, but is exactly the "cost-competitive on-site alternative" watch-item. Recurring, so already in the prior — no move. | ⁶

CEG | Conviction: HOLD | Surprise: LOW | Walmart goes nuclear via Constellation — a new corporate offtake counterparty, directly CONFIRMS the "convert uncontracted nuclear into long PPAs" pillar. Recurring (shown 1d), so already reflected; confirming but not new enough to re-rate. | ³

Everything else tonight was off-thesis (the Hormuz ship-strike cluster is an oil story, and we hold no oil), recurring (DOE $17.5B nuclear loans, FERC interconnection reform, $700B hyperscaler capex, AOC/Sanders DC-moratorium bill), or T3-only inference-efficiency leads (JetSpec, audio.cpp) with no T1/T2 corroboration. No prior moved.

Inflection Radar

[emergent] AI Compute Economics & Limits | Academic research is defining the hidden costs and physical constraints of LLMs, suggesting that optimization efforts (quantization, pruning) may inflate reasoning complexity or hit verification bottlenecks before achieving linear cost reduction. | Touches: NEW | ¹⁰, ¹¹, ¹²

[emergent] Embodied AI Safety & Airspace | The convergence of advanced robotics research (whole-body control, tactile sensing) with shared airspace prediction models indicates that the next bottleneck for autonomous systems is not computation, but reliable, low-latency physical interaction and collision avoidance. | Touches: NEW | ¹³, ¹⁴

[emergent] Critical Infrastructure Capitalization | State-level funding (DOE, FERC) is rapidly deploying massive capital commitments ($17.5B+) across nuclear power and data center buildouts, signaling a policy shift favoring high-CAPEX, regulated energy sources over decentralized alternatives. | Touches: NEW | ¹⁵, [heatmap.news/energy/ferc-data-center-plan]

[emergent] AI Infrastructure Viability | Multiple signals confirm the inflection point where AI revenue is expected to outpace compute build costs, shifting focus from raw model training (hyperscaler CAPEX) to specialized, localized neocloud deployment and resource optimization. | Touches: NEW | ¹⁶, ¹⁷

[emergent] Industrial Robotics Disruption | The emergence of specialized, domestic-supply industrial drones (Japan) alongside funding for disruptive enterprise software challengers signals a shift in automation from generalized IT services to physical, localized, and highly regulated operational environments. | Touches: NEW | ¹⁸, [dronelife.com/2026/06/24/prodrone-debuts-all-japanese-industrial-drone/]

QA & Caveats

No issues found.

Sources

1. Qualcomm unveils three new data center solutions including Dragonfly C1000 CPU, set to be deployed by Meta datacenterdynamics.com
2. Sponsored: Why ‘tokens per watt’ is crucial for measuring AI efficiency datacenterdynamics.com
3. Walmart Goes Nuclear heatmap.news
4. US SMR firm Deep Fission sees customer pipeline swell to 18.5GW, on the back of growing data center interest datacenterdynamics.com
5. Generation Mining nears C$1 billion copper project funding mining.com
6. FuelCell Energy signs up to 380MW fuel cell supply deal with Fit Energy to power data center operations datacenterdynamics.com
7. Plan To Move Ships Through Strait Of Hormuz Paused After Iran Strikes Cargo Vessel twz.com
8. UN halts escort of ships through Hormuz after vessel comes under attack investing.com
9. Generation Mining nears $1B copper project funding northernminer.com
10. Quantization Inflates Reasoning: Token Inflation as a Hidden Cost of Low-Bit Reasoning Models arxiv.org
11. Don't Go Breaking My LLM: The Impact of Pruning Attention Layers on Explanation Faithfulness and Confidence Calibration arxiv.org
12. Dustin: Draft-Augmented Sparse Verification for Efficient Long-Context Generation with Speculative Decoding arxiv.org
13. AeroCast: Probabilistic 3D Trajectory Prediction for Non-Cooperative Aerial Obstacles via Transformer-MDN Architecture arxiv.org
14. Self Capacitive Tactile Sensor System designed for Companion Robots arxiv.org
15. DOE offers $17.5B in loans to help build 10 large nuclear reactors utilitydive.com
16. Netris raises $15M Series A from a16z to help AI neoclouds go live faster techcrunch.com
17. AI revenue has finally outpaced the cost of building the infrastructure behind it - Startup Fortune news.google.com
18. Former Infosys chief has a new startup that wants to challenge the IT services world techcrunch.com