Morning Analysis — 2026-10-11

Dek

Getting electricity is no longer the hardest part of building AI data centers. Getting permission is, so power that is already approved and already connected is becoming the most valuable kind.

The Big Shift

Local opposition to data centers has turned into a national political fight, and it is now heading for the permitting reform bill in Washington before the midterms (¹). Permitting reform is the push to speed up approvals for energy projects and power lines. The cost is already visible: Heatmap counts more than $260 billion of data center projects cancelled in 2026, against roughly $800 billion of planned spending (²). The trend points one way. New greenfield sites (empty land with no existing power hookup) get slower and riskier. Power at sites that are already permitted and already connected to the grid gains value.

Analysis

Power: the market is paying for capacity that already has approval. Look at which deals are getting done. Google is paying to "uprate" 11 existing Constellation nuclear reactors (³). An uprate squeezes more output from a plant that already has its licenses, its land, and its grid connection. PJM, the grid operator for 13 eastern states, fast-tracked 2.1 GW of new connections (⁴). Almost all of it is batteries plus a 455 MW boost at an existing gas plant, and none of it comes online before mid-2029. Two things follow. First, this supports the CEG and TLN theses: owners of existing plants hold a scarce asset that can't easily be copied. Second, the 2029 date shows that new power for anything outside existing sites is three-plus years away.

Siting: secrecy has stopped working. Amazon will stop using NDAs (non-disclosure agreements) when it negotiates with local governments, following Microsoft (⁵). Meanwhile, a council in northern Italy rejected an 80 MW project (⁶). In Poland, a local councilor exposed a project that he says was hidden from public scrutiny (⁷). And the Wilderness Society, a mainstream conservation group, now backs a moratorium (⁸). Going public is the right move, but it also gives opponents a clear target and a timeline to organize around. Expect approval timelines to get longer, not shorter.

Former bitcoin miners are winning because they already have approvals. Cipher extended its Barber Lake leases with Fluidstack and an unnamed AI lab to 20 years (⁹). A 20-year lease is the kind of contract lenders accept as collateral for debt. It shows AI tenants will commit for decades to sites that already have power and permits, which supports the WULF hosting thesis. Not every miner is making that move. Hyperscale Data is putting its Montana site into a new bitcoin mining subsidiary rather than leasing it to AI tenants (¹⁰). Today, check the WULF 8-K filed Oct. 5 (marked "other material event") for lease or financing news.

Compute: chips are being pulled away from consumers, and the efficiency risk is still small. Nvidia has reportedly stopped making the RTX 5090, its top gaming card, so it can shift supply to data center chips (¹¹). That suggests demand for AI chips still exceeds supply. The other side of the ledger is a hobbyist claim of a 27B-class model in "ternary" weights, where every weight is -1, 0 or +1. The post says it runs about 9x smaller and keeps 98% of full-precision quality (¹²). If that held up at frontier scale, AI would need less power per answer, which works against Own-the-Bottleneck. For now it is unverified, small scale, and saves memory more than power. Note it, but don't change the view.

Geopolitics and materials: Europe and the US are working on the slow problems. The EU granted strategic status to KGHM's $3.46 billion Polish copper projects, which speeds up their approvals (¹³). In the US, rare earth recycling is being pitched as a bridge while new mines take years to open. But the scale is tiny: Paladin is going from 40 to 150 tonnes a year (¹⁴). Taiwan's President Lai is urging democracies to stand together after the Trump-Xi talks (¹⁵). Any sign that Washington will trade Taiwan's interests for a China deal is a risk to the supply of advanced chips. China's holiday has slowed rare earth trading, so watch for Q4 price moves this week as Chinese buyers return (¹⁶). The common thread with data centers: approval and processing are the bottlenecks, not the rock in the ground.

What Would Prove Us Wrong

• Hyperscalers keep spending even as projects get cancelled. If Q3 earnings later this month keep 2026–27 capex at $800B or more and say cancelled sites were simply replaced elsewhere, then the backlash is moving projects around, not shrinking them. That would weaken the idea that already-permitted power deserves a big premium, which hits CEG, TLN and WULF valuations.
• Ternary weights work at frontier scale. If a major lab or a credible independent benchmark (T1/T2 source) shows a model of 70B parameters or more in ternary weights, within about 2% of full-precision quality and with a measured cut in power per token, this would trigger the "efficiency flattens demand" case. That is a direct hit on Own-the-Bottleneck.
• Permitting reform passes with fast lanes for data centers, or a FERC rule makes it easy to connect new sites. Either would erode the edge that existing approved sites have today. It would hurt TLN's Cornerstone premium most. FERC's 203 ruling on Cornerstone, which must approve the deal, is the near-term marker. VST's Oct. 5 8-K shows it signed a material agreement and took on debt. If that turns out to be a new-build deal rather than one at an existing plant, that is an early test of whether new projects can still get built at scale.

Thesis Impact

• Own-the-Bottleneck | Conviction: HOLD | Surprise: MED | A Reddit post claims a 27B-class reasoning model squeezed into ternary weights (each weight is just -1, 0 or +1). It says the model is about 9x smaller than full precision, keeps about 98% of the quality, and runs on ordinary laptops. If that holds up at scale, it points toward the "inference efficiency flattens AI power demand" trigger, so it CONTRADICTS the thesis. But it is T3, unverified and hobbyist-scale. A smaller model mainly saves memory; it doesn't clearly cut data-center power demand. Watch for T1/T2 benchmarks or a major lab adopting ternary weights before moving. | ¹²
• CEG | Conviction: HOLD | Surprise: LOW | Google signed a deal to uprate (raise the output of) 11 Constellation nuclear reactors. That CONFIRMS pillar 2, where hyperscalers pay up for existing nuclear. It has been in the feed for 4 days and the 0.85 prior already reflects it. | ³
• WULF | Conviction: HOLD | Surprise: LOW | Cipher stretched its Barber Lake leases with Fluidstack and an unnamed AI lab to 20 years. That shows AI tenants will sign long leases with former bitcoin miners, which CONFIRMS the hosting-pivot pillars. It is a peer deal, not TeraWulf's own, and it's a repeat item. | ⁹
• TLN | Conviction: HOLD | Surprise: LOW | A Seeking Alpha piece argues the Cornerstone gas deal and buybacks strengthen the buy case. That CONFIRMS pillar 3, but it is an opinion piece (T3) with no new facts, so it can't move conviction. We still need the FERC 203 ruling on Cornerstone and an interest-coverage update. | ¹⁷

Inflection Radar

[emergent] Local AI Compute Scaling | High-density, multi-GPU setups are proving viable for localized LLM inference outside major cloud providers, signaling a shift in deployment architecture and demand for specialized compute hardware. | NEW | ¹⁸

[emergent] AI in Personalized Media | Apple's interest in licensing tech from a personalized podcast startup suggests AI adoption is moving beyond general content generation into highly specific, individualized media experiences. | NEW | ¹⁹

[emergent] Defense Tech Adoption Bottleneck | The difficulty in transitioning promising military prototypes into large-scale, funded acquisitions highlights a systemic structural risk and bottleneck within the defense industrial base. | NEW | ²⁰

[emergent] Federal Counter-UAS Mandate | Bipartisan calls for federal legislation on counter-drone training indicate that the regulatory and operational necessity for anti-UAS capabilities is rapidly outpacing current state/local guidelines. | NEW | ²¹

QA & Caveats

• Own-the-Bottleneck: The claim about flattening AI power demand is speculative and based on a T3, hobbyist source. It does not clearly demonstrate a cut in data-center power demand.
• TLN: This call relies on an opinion piece (T3) with no new facts. Conviction should not be moved by this source alone.

Sources

1. The Data Center Backlash Is Coming for Permitting Reform heatmap.news
2. Exclusive: Data Center Cancellations Top $260 Billion in 2026 heatmap.news
3. Google Inks a Deal to Uprate 11 Nuclear Reactors heatmap.news
4. Engie, LS Power land 2.1 GW in fast-track interconnection approval from PJM utilitydive.com
5. Amazon and others are done keeping data center deals secret. Is it enough to build trust? techcrunch.com
6. Council rejects 80MW data center in northern Italy datacenterdynamics.com
7. Uncertainty over data center development in Trzebnica, Poland datacenterdynamics.com
8. Why the Wilderness Society Backs a Data Center Moratorium heatmap.news
9. Cipher extends Barber Lake data center lease commitments to 20 years datacenterdynamics.com
10. Hyperscale Data plans Patriot BTC subsidiary for Bitcoin mining business datacenterdynamics.com
11. No more RTX 5090 reddit.com
12. Now you can grow Bonsai on your potato reddit.com
13. EU fast-tracks KGHM’s $3.46B Polish copper projects northernminer.com
14. Rare earth recycling can bridge US supply gap as mines face long timelines, Paladin says mining.com
15. Taiwan’s Lai Urges Democracies to Unite After Trump-Xi Talks bloomberg.com
16. China Holiday Slows Rare Earth Trade; Market Awaits Q4 Demand【SMM Analysis】 - Shanghai Metals Market news.google.com
17. Talen: Cornerstone And Buybacks Strengthen The Buy Case seekingalpha.com
18. Building a 4x R9700 setup for a 10 person startup reddit.com
19. Apple discloses deal to hire team and license tech from personalized podcast startup Huxe techcrunch.com
20. One year in, Army’s FUZE leader eyes ‘bigger’ acquisition challenges breakingdefense.com
21. 29 Governors Call for Expanded Counter-Drone Training Under New Federal Law dronelife.com