Morning Analysis — 2026-06-28

Dek

The cheapest path to more AI just got cheaper — and the only question that matters is whether that shrinks the power bill or feeds it.

The Big Shift

Overnight, China's DeepSeek shipped a major V4 upgrade built around "DSpark," which it says sharply speeds up AI responses while easing the strain on chips — in plain terms, getting more answers out of the same silicon. That matters because a durable jump in inference efficiency (the cost of *running* a model, not training it) is the exact trigger we've flagged as the thing that could break the AI-power-demand story. The signal cuts against our Own-the-Bottleneck thesis at the margin — but history says cheaper compute usually *expands* total usage (the Jevons effect), so one model release is a flag to watch, not a thesis-killer. ¹

Analysis

Efficiency vs. demand is today's central tension. DeepSeek's DSpark and OpenAI's custom inference chip "Jalapeño" (built with Broadcom) both point the same direction: the industry is racing to make each AI query cheaper to serve ². The bear read is that this softens the explosive power-demand curve our theses lean on. The bull read — and the one the money is still backing — is that cheaper inference pulls forward *more* AI use, not less. Watch which one the capex confirms: hyperscalers are still guiding toward a $600B infrastructure spend in 2026 ³. Spending that scale while efficiency improves is the tell that demand, not thrift, is winning.

Firm power keeps getting locked in, not retired. The clearest cross-day thread is that nobody is allowed to walk away from reliable generation. Constellation filed to extend the life of its Ginna nuclear plant ⁴, Walmart signed on to nuclear via Constellation ⁵, and the Trump administration has laid out its buildout plan ⁶. Talen's (TLN) fresh 8-K — a new material agreement plus an acquisition and new debt — fits the same pattern of nuclear-and-gas operators signing long-term load deals. Implication: even if inference gets more efficient, the contracted demand for around-the-clock "firm" power is being written into multi-decade contracts now. That underpins the power thesis regardless of the efficiency debate.

But the local backlash is real and spreading. The data-center site stream is relentless — new campuses in Iowa, Texas, Arkansas, Ontario, Wales, and Australia — yet the pushback is hardening just as fast. East Fishkill, NY passed a three-year moratorium that kills a million-square-foot project ⁷, and a senior House Democrat is now calling for a *national* data-center moratorium ⁸. Implication: the binding constraint is shifting from "can you build the megawatts" to "will the community let you." That favors developers with sites that already have power and permits — the scarcity our bottleneck thesis prices in.

Equipment names look priced for perfection. "GE Vernova Is Overheating" is a bearish valuation take, and it's consistent with the risk we already flagged on GEV — a premium multiple that bakes in years of flawless execution ⁹. It's a lone, lower-tier opinion with no harder corroboration, so it doesn't move conviction. The point: the move from here needs *execution* proof (orders, deliveries), not just narrative.

Geopolitics has cooled from the front page but not off the board. A US–Iran peace memorandum has reopened the Strait of Hormuz, yet the US already carried out its first post-deal strike after an Iranian attack on a cargo ship ¹⁰, and the ECB's Schnabel is still warning on upside inflation risk despite the truce ¹¹. Implication: the energy-price tail risk is smaller than in March but not gone — a re-closure would slam back into power and materials costs across the whole stack.

What Would Prove Us Wrong

• Inference efficiency shows up as flat or falling power demand. If a hyperscaler trims or delays its 2026 capex guide, or PJM/ERCOT data-center interconnection requests stall, that's the Jevons assumption breaking — a direct hit to Own-the-Bottleneck. The DeepSeek DSpark release is the leading indicator; watch the capex numbers next.
• A hard GEV execution miss, not just a valuation gripe. A canceled or pushed-out turbine order, a guidance cut, or a margin warning would convert the "overheating" opinion into a real GEV problem. A lone Seeking Alpha note does not.
• Moratoria go national or cluster regionally. If the Pallone national moratorium gains co-sponsors, or two or three more states/large counties follow East Fishkill, the demand pipeline our power theses assume could choke on permitting rather than physics — bearish for the whole buildout chain.

Thesis Impact

Own-the-Bottleneck | Conviction: HOLD | Surprise: MED | DeepSeek's V4 upgrade (DSpark) sharply cuts inference cost and chip strain — this is the exact "inference-efficiency step-change" listed as a break trigger across the power-demand theses. CONTRADICTS at the margin, but one model release ≠ durable step-change, and cheaper inference historically *grows* total compute (Jevons). Not enough to move P=0.81; flagging as the most important disconfirming item to watch. T2. | ¹

GEV | Conviction: HOLD | Surprise: LOW | "GE Vernova Is Overheating" — a bearish valuation take consistent with the thesis's own flagged risk (premium multiple prices much in). T3 / leads-only, no T1-T2 corroboration, so no move on conviction. Watch for a credible execution or order-cancellation item to confirm. | ⁹

Everything else is recurring or already in the prior (CEG Ginna renewal, Walmart/Constellation nuclear, the DCD site-buildout stream, Hormuz energy items, hyperscaler $600B capex) — LOW surprise, no move.

Inflection Radar

[emergent] Data Center Power Regulation | FERC developing plans for data center power capacity suggests grid constraints are moving from engineering problems to regulatory bottlenecks, impacting all compute CapEx planning. | Touches: NEW | ¹²

[emergent] Institutional AI Adoption in Finance | The documented rise of firms like Jane Street using proprietary AI at scale signals a structural shift in market access and alpha generation, favoring deep tech capability over traditional market presence. | Touches: NEW | ¹³

QA & Caveats

• Own-the-Bottleneck: The claim of a durable step-change from DeepSeek's V4 upgrade is not clearly supported by the single source; the call correctly notes this contradicts historical trends (Jevons) and requires further evidence beyond the article.
• GEV: This is a T3, leads-only item with no T1 or T2 corroboration, confirming the assessment that it lacks conviction based on the provided data.

Sources

1. Faster AI, lower costs: DSpark eases inference bottlenecks and chip strain, says DeepSeek scmp.com
2. techcrunch.com techcrunch.com
3. AI Infrastructure Investment 2026: $600B Hyperscaler Boom - Intellectia AI news.google.com
4. Constellation Energy Generation, LLC; R.E. Ginna Nuclear Power Plant; Subsequent License Renewal Application federalregister.gov
5. Walmart Goes Nuclear heatmap.news
6. Trump’s Big Nuclear Play Is Here heatmap.news
7. Three-year data center moratorium passed in town of East Fishkill, New York State, blocking planned campus datacenterdynamics.com
8. Key House Democrat Calls for a National Data Center Moratorium heatmap.news
9. GE Vernova Is Overheating seekingalpha.com
10. U.S. Carries Out First Strike On Iran Since Peace Memorandum Signed (Updated) twz.com
11. ECB’s Schnabel Sees Upside Inflation Risks Despite Peace Deal bloomberg.com
12. FERC Has a New Plan for Data Centers heatmap.news
13. Jane Street's AI-driven rise in global finance - MSN news.google.com