# The AI trade nobody priced

*September 2026*

Leopold Aschenbrenner's fund, Situational Awareness LP, filed its first 13F in February 2025. It reported [$255 million across six positions](https://13f.info/manager/0002045724-situational-awareness-lp), and the top holdings were Marvell, Vistra, Vertiv and Talen Energy. None of them trains a model. They design networking chips, generate electricity, and power and cool server halls. Over 2025 the book filled with bitcoin miners. By March 2026 the fund held [IREN at $401 million, Core Scientific at $389 million, Applied Digital at $320 million, Riot at $142 million and CleanSpark at $104 million](https://finance.yahoo.com/markets/options/articles/situational-awareness-lp-discloses-5-155553432.html), and it had grown its CleanSpark stake from 1.64 million shares to 12.28 million in one quarter. By June the reported book was [$20.2 billion](https://filingsflow.com/manager/0002045724/situational-awareness-lp). In July the fund was [forced to sell much of it to Citadel](https://www.cnbc.com/2026/08/14/situational-awareness-filing-shows-ai-bets-before-forced-portfolio-sale-to-citadel.html).

The miners mattered because they owned what AI needed and couldn't get quickly: power contracts, substations and land zoned for heavy load. In June 2024 Core Scientific signed [12-year contracts to provide about 200 MW](https://www.sec.gov/Archives/edgar/data/1839341/000162828024027044/ex991-corzunsolicitedpropo.htm) of hosting to CoreWeave, and rejected CoreWeave's $5.75-a-share offer as too low. A year later CoreWeave came back with [about $9.0 billion](https://investors.coreweave.com/news/news-details/2025/CoreWeave-to-Acquire-Core-Scientific/default.aspx). Core Scientific's shareholders [voted that down too](https://www.reuters.com/business/core-scientific-terminates-9-billion-merger-deal-with-coreweave-2025-10-30/). The market was repricing a company that looked like a crypto miner and was actually a grid connection.

On August 1 I told a friend what I thought the next version of that trade was (my translation): "A lot of factories have had grid infrastructure ready for years, and a bunch of neoclouds and labs are going to buy them, but they're priced as if that won't happen. Like when Leopold bought bitcoin mining datacenters knowing they would become AI datacenters, before the market knew or had priced it."

The timeline is messier than my version of it. By the time the fund held miners in size, in the third quarter of 2025, the CoreWeave deal was public. But the first filing says the same thing louder. Before he bought a single miner, Aschenbrenner bought power.

Here is the thesis. The part of the AI trade the market still misprices is physical and illiquid. Most attention goes to models and to the companies that train them, and the physical layer gets treated as procurement. I think the binding constraints are interconnection time, memory, turbines and reactors. The liquid tickers that own them got priced, and then some. The assets that own them without being labelled AI are still priced like the old economy. On the short side, I'd look at neocloud debt, because GPUs lose value faster than the books assume. Above all of it sits a political layer: labs drifting toward the state, a state that doesn't understand what it depends on, and a middle class that might break the buildout before it finishes.

**This is not investment advice.** I'm an engineer writing down bets I discussed with friends. The numbers come from public filings and reports that can be wrong or stale, and several of these bets are expensive to hold even if the thesis is right.

## The binding constraint is a queue

To plug a new power plant into the US grid, you file an interconnection request and wait for studies. Lawrence Berkeley National Laboratory tracks that wait every year. In the [2026 edition of Queued Up](https://eta-publications.lbl.gov/sites/default/files/2026-06/queued_up_2026_edition.pdf), the median project built in 2025 took 61 months from request to commercial operation, "compared to 36 months in 2015 and 22 months in 2008." Size makes it worse. The median project under 5 MW took 20 months. The median 200+ MW project, the scale a serious AI campus needs, took 59.

![Median months from interconnection request to commercial operation, by year and by project size](https://future-seems-so-good.com/blog/assets/the-ai-trade-nobody-priced/charts/interconnection-wait.svg)

The queue itself is a graveyard. At the end of 2025, 2,061 GW was waiting for a connection, against 1,374 GW of total installed US capacity. Of the 2,290 GW in the queue a year earlier, "only 53 GW reached operational status in 2025, yielding an annual throughput rate of just 2%."

One caveat. LBNL measures generators, and says so on its first page: "there are separate queues for large loads and those are not included in this report." A data center is a load, so the chart measures how long new supply takes to reach the grid that data centers draw from, not the data centers' own wait.

Two lines in the same report tell you where developers think the power will come from. Active natural gas capacity in the queues rose 86% in 2025, to 253 GW, while solar, wind and storage all shrank. Active nuclear capacity roughly doubled, from 5.3 GW to 10.4 GW.

## Old sites already won the queue

If new supply takes five years, the scarce asset is a site that already has an interconnection and a substation. Old industrial land often has both. It was built for heavy load decades ago, and the utility relationship is already there.

The conversions are happening.

**Factories into data centers.** The EPA's guidance on brownfield reuse cites Meta, which ["converted two adjacent brownfield sites into a data center"](https://www.epa.gov/brownfields/reuse-considerations-data-centers-brownfield-sites) on land "previously used for textile and manufacturing operations."

**Retired plants.** A listing for a retired gas plant in Bayonne, New Jersey advertises its [192 MW legacy interconnection](https://gridcensus.com/brownfield-sites/new-jersey/bayonne-plant-holding-llc-6f9b4828) as "collapsing the speed-to-power timeline versus a greenfield build."

**Speed.** An industrial adaptive-reuse site in Clarksville went [from 500 kW live in September 2023 to 24 MW by June 2024](https://tecfusions.com/wp-content/uploads/2024/09/Case-Study_Clarksville-Proving-Ground-1.pdf). That's nine months from first power to 24 MW.

The asset here is the interconnection agreement, and the building happens to sit on top of it. Bitcoin miners were the first class of this asset to get repriced because a miner is already a building that turns megawatts into computation. The next class is less obvious, which usually means cheaper: old factories and retired plants, often owned by manufacturing companies whose market caps reflect what they make and ignore the power contracts they sit on. "Buying manufacturing companies is a good trade too," I wrote to him the same evening.

I made the general argument about compute as capital in [Positive feedback eats the world](https://future-seems-so-good.com/blog/positive-feedback-eats-the-world). This is the physical version: if compute is the capital, the grid connection is the land it stands on.

## Somebody priced it

The title of this essay is wrong in the literal sense. Aschenbrenner priced the physical trade, and a lot of money followed him. At the end of June his five largest disclosed holdings were Sandisk, Micron, Bloom Energy, TSMC and Nebius. Then, [CNBC reported](https://www.cnbc.com/2026/08/14/situational-awareness-filing-shows-ai-bets-before-forced-portfolio-sale-to-citadel.html), Sandisk "plunged almost 47% in July," Micron "tumbled nearly 29%," Bloom fell 32% and Nebius 31%. With leverage and losses on its bets against software companies, the fund had to sell its leveraged positions, SK hynix and CoreWeave among them, to Citadel at a discount. A fund that had held roughly $45 billion at its peak held about $10 billion by July 30.

Nothing physical got less scarce that month. The queue was still five years long, turbines were still sold out, and HBM was still short in September. What got priced was the liquid expression of the thesis, a handful of tickers everyone could buy, many of them with borrowed money. I think the remaining asymmetry has two parts.

The first is duration. A margin call takes a month and an interconnection takes 61 months. The physical thesis pays on the queue's clock, and levered public money runs on the market's. Whoever can hold through a 47% month gets the positions the forced sellers gave up.

The second is legibility. What got priced was labelled AI: memory makers, fuel cells, neoclouds, miners that had already rebranded. The assets I care about are still filed under something else, like a textile plant's substation.

The question I use to find them came up in a chat with an agent on September 24: "If intelligence becomes nearly free, what remains scarce? ... Who owns those scarce things?" My list was infrastructure, lithography, packaging, HBM, regulation, distribution, data and relationships. In [Money is method times compute times data](https://future-seems-so-good.com/blog/money-is-method-times-compute-times-data) I argued that an edge is the gap between a thesis and what the market expects, not the thesis itself. "Physical AI is scarce" is now what the market expects. The gap left is in who gets recognized as owning the scarce thing, and in who can afford to wait.

## The bets

Later that evening he asked what the general bets would be if we ran a fund. My answer, translated: "HBM (even more now), CPU, old factories in the US, shorting neolab bonds (they got screwed by H100 depreciation and by Nvidia launching B300), long natural gas, short the real, barbell (top and bottom make money, the middle gets screwed)."

He added the one I hadn't weighted enough: "Long natural gas and very long nuclear. Like 100x long nuclear."

**HBM.** High-bandwidth memory sits next to every frontier accelerator, and its suppliers are sold out. SK hynix [sold out its 2025 supply and most of its 2026 supply](https://www.tweaktown.com/news/104307/sk-hynix-sold-out-of-hbm-for-2025-soon-taking-customer-orders-2026-supply/index.html). Micron's business chief said in August 2025 that the company was ["confident it will sell out"](https://www.trendforce.com/news/2025/08/13/news-hbm-battle-heats-up-micron-reportedly-hints-2026-sell-out-sk-hynix-yet-to-confirm/) its entire 2026 allocation. Aschenbrenner shows up here again. In the first quarter of 2026 his fund held [Micron puts worth $584 million](https://finance.yahoo.com/markets/options/articles/situational-awareness-lp-discloses-5-155553432.html). By June its two largest positions were [Sandisk at 28% of the book and Micron at 27.5%](https://filingsflow.com/manager/0002045724/situational-awareness-lp). The stocks crashed in July and the chips didn't. In September [TrendForce reported](https://www.trendforce.com/news/2026/09/01/news-hbm3e-spot-prices-said-to-be-4-5x-lta-levels-with-samsung-reportedly-locking-70-of-memory-capacity-through-2031/) a 36GB HBM3E stack at about $2,100 on the spot market, "roughly four to five times its LTA price," and Samsung reportedly locking about 70% of its memory capacity through 2031 into long-term agreements. That is also the catch for the equity: such contracts "can limit windfall gains when spot prices surge." I told my friend the main bottleneck right now is data, which I argued in *How to achieve superintelligence*. When labs solve it, every extra trajectory of training and inference runs through memory bandwidth.

**Gas turbines.** Gas is what you can build this decade. In December 2025 Reuters reported that GE Vernova had ["sold out of its gas turbines through 2028"](https://www.reuters.com/business/energy/ge-vernova-expects-80-gigawatts-gas-turbine-contracts-by-years-end-2025-12-09/) with less than 10% of 2029 left to sell, and that data centers were about a third of US gas-power transactions. Crusoe ordered [29 aeroderivative units, "nearly 1 GW"](https://www.gevernova.com/news/press-releases/ge-vernova-crusoe-announce-major-29-unit-aeroderivative-gas-turbine-deliver-ai-data-centers), for AI data centers alone.

**Nuclear (my friend's conviction).** His reasoning, translated: "Energy demand will go up at least 20x in the next 10 years, and the only form of energy production with a chance of meeting that demand is nuclear." I haven't found a source that supports a 20x figure, so treat that number as his thesis, not a forecast I can cite. The direction is visible. Constellation is restarting the Crane plant, formerly Three Mile Island Unit 1, under a [20-year power purchase agreement with Microsoft](https://constellationenergy.gcs-web.com/news-releases/news-release-details/one-year-later-crane-clean-energy-center-still-spotlight-and), 835 MW, accelerated to 2027. Google and Kairos are building [Hermes 2, a 50 MW reactor targeted for 2030](https://blog.google/company-news/outreach-and-initiatives/sustainability/google-first-advanced-nuclear-reactor-project-with-kairos-power-and-tennessee-valley-authority/), as part of up to 500 MW by 2035. Nuclear requests in the interconnection queues doubled in 2025. The problem is timing: gas arrives this decade and nuclear mostly arrives next decade. His 100x is a bet on the second half.

**Short neocloud debt.** Neoclouds borrow to buy GPUs and rent them out. The bet is that the collateral loses value faster than the accounting says. CoreWeave extended the useful life of its GPUs [from five to six years in January 2023](https://drops.mts.now/ai-capex/wiki/documents/coreweave-filings/), which reduced 2023 expenses by $20 million, according to a compilation of its filings. In November 2025 Michael Burry [wrote](https://sherwood.news/markets/michael-burry-has-some-concerns-about-ai-accounting/): "Understating depreciation by extending useful life of assets artificially boosts earnings - one of the more common frauds of the modern era." His post was aimed at Oracle and Meta. Jim Chanos pointed the same argument at CoreWeave the next day: "$CRWV's 3Q annualized Adj EBITDA was $3.4B, and annualized interest was $1.2B. Using 10-year life(!) on their $20.0B of est. GPU's ($2.0B) means they are still barely profitable."

Rents say the collateral is already losing value. AIMultiple's rental index put the median H100 at about $3.38 an hour in August 2026, [down from more than $7 in early 2024](https://shattered.io/h100-h200-b200-cloud-gpu-pricing-2026/), with marketplace floors near $1.49. That's a halving in about two and a half years, for a chip CoreWeave depreciates over six.

The debt is real and expensive. In April 2026 CoreWeave had [$2.75 billion of 9.750% senior notes due 2031](https://www.stocktitan.net/sec-filings/CRWV/8-k-core-weave-inc-reports-material-event-0161dfbf1011.html) outstanding, and in June it added [$1.25 billion at 9.625% and €2 billion at 8.500%, due 2032](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000291/crwv-20260618.htm), all senior unsecured. That's close to 10% on unsecured money against an asset that loses rank with every Nvidia generation.

Aschenbrenner held [7.48 million CoreWeave shares](https://filingsflow.com/manager/0002045724/situational-awareness-lp), worth $745 million, at the end of June, and CoreWeave was among the positions he sold to Citadel in July. That doesn't contradict the short. Equity keeps the upside if the boom keeps going. An unsecured bond gets par at best and eats the loss if GPU values collapse before the notes mature. My barbell applies to the capital structure: own the ends, avoid the middle. Polymarket prices the tail at 10.5% (table below).

**Short the real.** Through July 2026, emerging-market equity inflows were [$70.6 billion for South Korea against $11.9 billion for Brazil](https://valorinternational.globo.com/markets/news/2026/08/13/ai-bets-flow-to-asian-emerging-markets-leaving-brazil-behind.ghtml), and Valor blamed Brazil's "lack of a strong AI-related narrative." Rabobank forecasts [USD/BRL at 5.35](https://www.fxstreet.com/news/brazilian-real-fiscal-risks-point-to-weakness-against-us-dollar-rabobank-202608181240) by year-end. I told my friend that Faria Lima doesn't understand AI. His correction was harsher: they understand it, they just trade it like idiots.

**The barbell.** The frame behind all of these is that the ends win and the middle gets squeezed. At the top are the frontier labs. At the bottom is the physical layer: sites, memory, turbines. The middle is anyone renting commodity compute with borrowed money, or wrapping a model with nothing underneath. I also told him this is the best moment for the two of us to start one because, the way I read Nick Land's *Meltdown*, the bubbles will keep getting bigger, and funds that can play that volatility will make orders of magnitude more than before. July showed the other side of that bet.

## The labs drift toward the state

This part is more speculative, and I'll mark where the evidence stops.

My other contrarian bet on August 1 was, translated, that "OAI/Anthropic will be nationalized." My friend asked what would have to happen for that to be 90% likely, or whether it had already happened. I answered: "time." Then, translated: "The American government still hasn't understood the impact of the singularity on sovereignty, and if it has, it already has a plan to sweep in some time from now. Besides, Sama already proposed 5% of OAI to the American government."

The evidence points to something softer than seizure, which I'll call soft nationalization.

**Equity.** In June 2026 CNBC reported that ["OpenAI could donate equity to the U.S. government to seed"](https://www.cnbc.com/2026/06/05/trump-open-ai-altman-stake.html) a public wealth fund, after talks lasting more than a year. The [Guardian](https://www.theguardian.com/technology/2026/jul/02/openai-stake-us-government-ai-sam-altman) followed with the Financial Times' report of a stake around 5%.

**Coercion.** On February 27, after Anthropic refused to drop safeguards against autonomous weapons and domestic surveillance, Defense Secretary Pete Hegseth moved to designate it a supply-chain risk: ["Effective immediately, no contractor… may conduct any commercial activity with Anthropic."](https://techcrunch.com/2026/02/27/pentagon-moves-to-designate-anthropic-as-a-supply-chain-risk/) The same report describes a six-month federal phase-out ordered by Trump. A preliminary injunction blocked the label in March, and on August 27 Judge Rita Lin [granted Anthropic summary judgment](https://cases.justia.com/federal/district-courts/california/candce/3:2026cv01996/465515/250/0.pdf), vacating the designation and the ban on contractors doing business with it: "The empty invocation of national security is not a blank check to punish and retaliate against government critics." A parallel designation under the federal procurement statute is still before the D.C. Circuit and [remains in effect](https://www.joneswalker.com/en/insights/blogs/ai-law-blog/national-security-is-not-a-blank-check-an-ai-model-retaliation-ruling-and-the.html?id=102nzx9). The Foundation for American Innovation called this kind of pressure ["soft nationalization"](https://www.thefai.org/posts/anthropics-position-merits-support), and said invoking the Defense Production Act would amount to "quasi-nationalization of a frontier lab."

**The template.** Aschenbrenner sketched this shape in 2024. In [The Project](https://situational-awareness.ai/the-project/) he wrote that "by 27/28 we'll get some form of government AGI project," and that "this doesn't need to look like literal nationalization... The relationship with the DoD might look like the relationship the DoD has with Boeing or Lockheed Martin."

My friend sharpened it in a way I keep coming back to. "Will OAI + Ant become a total Los Alamos Manhattan Project? And the Chinese models will play the role that OAI + Ant play in the US today. Like 'inference for the people' stays with China while 'research for actual AGI' stays with OAI + Ant inside the government. Which is kind of what already happens with manufacturing in the US."

Aschenbrenner has a version of the split too: "nuclear weapons permanently reserved for the government—but civilian nuclear energy flourished as private projects." In his version the civilian side stays American and private. In my friend's, the civilian side goes offshore, the way manufacturing did. Frontier research moves inside the state, and cheap inference for everyone else comes from Chinese open weights. I wrote about what that inference carries in [The harness is in the weights](https://future-seems-so-good.com/blog/the-harness-is-in-the-weights).

In August, talking to another friend about the sniping between Sam Altman and Dario Amodei, I said the ego battle might have been positive so far, "but eventually this ego will be dissipated by the nationalization of the labs." That's the speculative core. A lab inside the state is a lab whose founder's temperament stops mattering.

## A state that doesn't understand what it depends on

Nationalization assumes a state strong enough to take something. In June I shared an essay in a reading group that argues the opposite. It came without a byline and I don't know who wrote it. Its argument: the American state is being disempowered by an industry it depends on and doesn't understand.

It opens at Camp David, with Eisenhower admitting to Khrushchev that his generals get their budgets by pointing at Soviet programs, and Khrushchev answering that his do the same. Eisenhower's [farewell address](https://www.pbs.org/wgbh/americanexperience/features/eisenhower-farewell/) warned that "public policy could itself become the captive of a scientific-technological elite." The essay reads the Anthropic fight through that lens. The designation looked severe, but "what actually surfaced was an obvious and acute one-sided dependency." It points to the six-month wind-down and says Anthropic's valuation on prediction platforms "barely moved." I haven't checked that against a price series, so treat it as the essay's. The live markets in September lean the same way (table below).

The essay's mechanism is epistemic: "The state doesn't get AI, and ends up having to opt for deference to those who do." Its numbers come from the Institute for Progress, which estimates that the Center for AI Standards and Innovation, the office meant to close that gap, has [about $15 million a year and needs about $84 million](https://ifp.org/funding-for-caisi/). The UK's equivalent has $89.1 million. The country that hosts almost every Western frontier lab funds its window into them at less than a fifth of what Britain spends.

Then it brings in Daron Acemoglu. In [Automation and Repression](https://economics.mit.edu/sites/default/files/2026-06/Automation%20and%20Repression.pdf), Acemoglu, Arda Gitmez and Mehdi Shadmehr model a capitalist state facing a possible worker revolt, which can regulate automation, redistribute, or repress: "Our main result is a complementarity between automation and repression. Unless the threat of revolt is quite weak or the capital stock is very low, the capitalist state prefers repression."

The essay's plain-language version is the part I'd underline. Early in the buildout, repression is too costly, so everyone promises redistribution. Later, redistribution gets more expensive as capital's share grows, and AI makes enforcement cheap. "Cooperation is cheap, while the cost of repression remains high. This makes redistribution the obvious answer to mass dislocation, for now."

That changes how I read my own nationalization bet. Soft nationalization and state disempowerment are two readings of one fact: the state needs the labs more than the labs need the state. An equity stake is the state trying to buy back, with ownership, the understanding it never paid for with an $84 million budget.

The August ruling, two months after the essay, complicates it. The executive lost, but it lost to a court, and a court is also the state. The order took away the power to punish a supplier for its terms and left the one that matters commercially: the Department of War "is undisputedly free to select the AI vendor of its choice." The state is weaker as a punisher and unchanged as a customer.

## 51.84%

On August 29 I sent both of them a probability chain. The claim: "There is going to be a >10x market cap gap between OAI and Anthropic, or the inverse, in 5 years, with both being bigger than today → 51.84% of probability."

The chain has four links.

1. **The US wins the AI race: 80%.** "By 2028 or 2029 the USA will already have millions of autonomous researchers delivering a lot of discoveries that leverage the country even with less GW than China." I'd have said 90%, "but Taiwan is still a thing so I dropped it to 80%."
2. **No new lab breaks the duopoly: 90%.** I gave a 10% chance that "a new lab figure (probably Ilya or whoever)" finds an efficient path to superintelligence, which would break the "both bigger than today" condition.
3. **No hard nationalization: 80%.** "The integral over time of this probability adjusted by me is 20%, considering all the politics and so on."
4. **The gap doesn't stall: 90%.** A 10% chance that neither lab pulls away. I don't think it will stall: "I have a strong belief that model capabilities are going to have a bigger and bigger gap and that this will be visible in inference and in prices (self improvement is a leverage that people kind of have not understood yet and underestimate)."

0.80 × 0.90 × 0.80 × 0.90 = 0.5184.

Two notes, a month later. First, link 3 only makes sense for hard nationalization. The soft version, with equity stakes, defense contracts and a joint venture, leaves both companies with market caps, and might widen the gap by picking a national champion. Second, I told my first friend on August 1 that the strange part "is that the gap grows exponentially." If self-improvement compounds, a 10x gap in five years is the conservative reading.

Prediction markets now price pieces of the chain. These are Polymarket prices on September 25, 2026, read from its public API:

| Market | Yes | Resolves on | Volume |
|---|---|---|---|
| [Anthropic valued above OpenAI on December 31](https://polymarket.com/event/anthropic-vs-openai-higher-valuation-on-december-31) | 91.5% | Nasdaq Private Market price, 2026-12-31 | ~$84k |
| [Anthropic valuation hits $3T by December 31](https://polymarket.com/event/will-anthropics-valuation-hit-by-december-31/will-anthropics-valuation-hit-high-3pt0t-by-december-31) | 24.5% | Nasdaq Private Market price, any day through 2026-12-31 | ~$1.3M |
| [US federal government takes a stake in OpenAI](https://polymarket.com/event/which-companies-will-the-us-take-a-stake-in/will-the-us-federal-government-take-a-stake-in-openai) | 9.5% | By 2026-12-31 | ~$11k |
| [CoreWeave announces bankruptcy](https://polymarket.com/event/coreweave-announces-bankruptcy-by/will-coreweave-announce-bankruptcy-by-december-31-2027) | 10.5% | By 2027-12-31 | ~$12k |

None of these measures a 10x gap, but the first says the market is close to sure about direction this year, and the direction is Anthropic. The third says soft nationalization through equity isn't expected before 2027, whatever the talks. The stake and bankruptcy markets are thin, so read them as a few people's opinions with money attached.

## The missing variable

My other friend answered on September 1, and his reply is better than my chain:

> Makes sense, acho que ta faltando prob of American middle class mass dissatisfaction with ai leading to stop the infra build out and to massive regulate slowing fundamental progress. This is important because this would lead to a break in the acceleration just at the most important time where china is increasing its second derivative. With that in mind I think one of the most important things we can do is better distribute the gains and create a shared narrative of success and blast it all over the world.
>
> This is the first time in history the winning players holders of the technology do not have a shared vision about the future. Ie
> - america wwii - American dream
> - Soviet Russia - tranquil farm life
> - imperial Japan - Japan expansion and dominance
> - Roman empire - Roman glory

("Acho que ta faltando" means "I think what's missing is.")

The anxiety is measurable. In an April 2025 Pew survey, [56% of US adults were extremely or very concerned](https://www.pewresearch.org/2025/04/03/views-of-risks-opportunities-and-regulation-of-ai/) about AI eliminating jobs, against 25% of AI experts, and 64% expected fewer US jobs over the next 20 years.

And it already has a physical form. Data Center Watch counted [$64 billion of data center projects blocked or delayed](https://www.datacenterwatch.org/report) by local opposition over roughly two years to March 2025, then [$98 billion](https://www.datacenterwatch.org/q22025) in the second quarter of 2025 alone. In the first quarter of 2026 it counted [at least 75 projects worth about $130 billion](https://www.datacenterwatch.org/q1-2026-1), "roughly matching the scale of all of 2025 in just three months," while active opposition groups more than doubled, from 396 to 833. The second quarter added [at least 45 projects worth nearly $68 billion](https://www.datacenterwatch.org/q2-2026). That's about $198 billion disrupted in the first half of 2026.

![Data center investment blocked or delayed by local opposition](https://future-seems-so-good.com/blog/assets/the-ai-trade-nobody-priced/charts/datacenter-opposition.svg)

His variable enters the chain as a fifth factor. The numbers that follow are mine, as an illustration, not his: if the probability of a backlash that stops the buildout is 20%, the chain falls from 51.84% to 41.5%. At 35% it falls to 33.7%.

It also fits Acemoglu's model better than anything in my chain does. His backlash is the revolt variable, and the model says automation pushes the state from redistribution toward repression. He argues for redistribution and adds something the model doesn't have, a story people can believe in. Each society on his list had a picture of what winning looked like for an ordinary person. The labs offer a vague promise of redistribution to be worked out later, and a promise like that can't hold a coalition together.

For the trade, the backlash cuts two ways. It hurts every greenfield project that needs a rezoning vote. It probably helps the brownfield thesis, since a site that was already industrial, already zoned and already connected gives a local opposition group less to vote on. That second point is my inference. I haven't seen opposition data broken down by site type.

## Sidebar: a friend's South Korea bet

*This is my first friend's bet, not mine. He put R$1,000 on it on August 1.*

Asked for his contrarian bets, he said, translated: "By 2050 South Korea loses all national sovereignty and either (1) becomes part of the Chinese bloc, or (2) has its economy 100% taken over by the US and its national bank effectively made useless."

His reasoning is Greece in 2015, "forced by the European Union to effectively dilute its central bank and sell all its important assets." Korea, he argues, is "in a population decline that has already passed the point of no return," so its assets will be liquidated to a great power, most likely the US.

The pieces are real. The July 2015 Euro Summit required Greece to set up an [independent fund to privatize state assets](https://www.consilium.europa.eu/en/policies/financial-assistance-eurozone-members/greece-programme/), and 14 regional airports went to Fraport for [€1.2 billion](https://www.rte.ie/news/business/2017/0412/867195-greek-airports-sold/). Korea's fertility rate was [0.72 in 2023](https://www.reuters.com/world/asia-pacific/south-korea-birthrate-rises-first-time-9-years-marriages-surge-2025-02-26/), the lowest in the world. The 2025 trade deal committed Korea to [$350 billion of US investment](https://www.reuters.com/world/china/what-we-know-about-south-koreas-trade-deal-with-us-2025-07-07/), with $200 billion of it in [cash, capped at $20 billion a year](https://en.yna.co.kr/view/AEN20251114008554320).

The evidence against: fertility [rose to 0.80 in 2025](https://www.korea.kr/news/policyNewsView.do?newsId=148970652), Korea leads Valor's comparison of AI-driven equity inflows, and a pledge capped at $20 billion a year is a negotiated squeeze, not a Troika. What I find sharp is how it crosses the HBM long. SK hynix sells out its memory because Korea makes something the whole buildout needs. His bet is that this is exactly why it gets absorbed.

## Where the trade leaks

**Power may already be priced.** GE Vernova's sold-out order book is public, the miner trade is over, and the brownfield version may be crowded by the time anyone reads this. There's a worse leak too. Aschenbrenner's third-largest position in June was [Bloom Energy, at 9.4%](https://filingsflow.com/manager/0002045724/situational-awareness-lp), which makes on-site power, and Crusoe bought its own turbines. If generation behind the meter beats the queue, a grid connection is worth less than I think.

**Depreciation may be fine.** CoreWeave's CEO said its first 10,000-plus H100 contract was [renewed at a price "within 5% of the original agreement"](https://sherwood.news/markets/michael-burry-has-some-concerns-about-ai-accounting/). Nebius said it was reselling Hoppers at renewal "often at better pricing than they were previously priced." Old chips find inference work, and a contracted renewal is not a marketplace rate. And even if I'm right, shorting a bond that pays 9.75% costs you that coupon every year you wait.

**The backlash may be priced, or partial.** About $198 billion was disrupted in the first half of 2026 while turbines stayed sold out through 2028 and Samsung reportedly locked most of its memory capacity into contracts through 2031. The gap between public and expert worry in Pew's data may mean policy lags the public for years. Local opposition can move projects to another county without stopping the national buildout.

**Nationalization may never come.** A court vacated the supply-chain label, and Polymarket gives a US stake in OpenAI this year 9.5%. Aschenbrenner himself expected a joint venture rather than a seizure. The essay quotes Trump in June saying he doesn't want to shut Anthropic down because "we're beating China by a lot." If the relationship settles into contracts and minority stakes, markets handle that fine, and link 3 of my chain is too pessimistic.

**Memory can turn.** The fund that made memory more than half its book in June held Micron puts in March and was a forced seller in July. Sold-out capacity invites new capacity, and suppliers who sold forward at contract prices don't collect the spot premium.

## The bets, and what would kill them

Not investment advice. This is a watch list.

| Bet | Thesis | What would falsify it | What to watch |
|---|---|---|---|
| Long US industrial sites with existing grid interconnection | New supply takes about 5 years to connect; sites with substations skip the line and are priced on what they used to make | Behind-the-meter power becomes the default; interconnection reform cuts queue times | LBNL's next Queued Up; large-load queue data; lab or neocloud purchases of manufacturing sites |
| Long HBM | Suppliers sold out; training and inference run through memory bandwidth | New capacity overshoots; architectures that need far less HBM | 2027 sell-out statements; HBM3E spot-to-contract gap (4 to 5x in September 2026); capacity under long-term contracts |
| Long gas turbines | Firm power you can build this decade; GE Vernova sold out through 2028 | Cancellations; a hyperscaler capex pullback; cheap storage | GE Vernova backlog; gas share of the queue (+86% in 2025) |
| Long nuclear (a friend's) | Only nuclear can meet demand at scale | Restarts slip; small reactors stay far above gas on cost | Crane (2027); Kairos Hermes 2 (2030); nuclear in queues (10.4 GW) |
| Short neocloud unsecured debt | GPUs lose value faster than 6-year schedules assume; ~10% unsecured coupons | Hopper lease rates hold; renewals near original prices | Median H100 rent ($3.38/hr in August 2026, from $7+ in early 2024); CoreWeave note spreads; Polymarket bankruptcy odds (10.5% by 2027) |
| Short BRL (mine) | Brazil is off the AI capital map and carries fiscal risk | Carry or an election squeezes shorts; Brazil finds an AI story | EM flows (Brazil $11.9B vs Korea $70.6B through July 2026); USD/BRL vs 5.35 |
| OAI and Anthropic diverge more than 10x in 5 years (51.84%) | US lead holds, no new lab, no hard nationalization, gaps compound | Capability and price convergence; a third lab reaches the frontier | Funding-round valuations; Polymarket (91.5% Anthropic above OpenAI at year-end); price per agentic task |
| Soft nationalization | The state needs the labs and buys in with equity and contracts | Courts and politics keep the state at arm's length, as in August | Polymarket US stake in OpenAI (9.5% for 2026); the D.C. Circuit ruling; CAISI's budget |
| Middle-class backlash (another friend's) | Opposition stops the buildout at the worst moment | Redistribution and a shared story arrive first | Data Center Watch totals ($130B in Q1 2026, $68B in Q2); opposition groups (833); state moratorium bills |
| South Korea loses sovereignty by 2050 (a friend's) | Demographic collapse forces asset sales to a great power | Fertility keeps recovering; chip supply becomes bargaining power | Korean fertility; disbursements under the $20B-a-year cap |

If I had to pick one row to watch, it's the first. A five-year queue creates a scarce asset, the liquid version of that trade already got priced and blew up, and the illiquid version is still priced by the building instead of the megawatts.

## Sources

- Situational Awareness LP 13F history: [13f.info](https://13f.info/manager/0002045724-situational-awareness-lp) (Q4 2024: $254.8M, six holdings, top MRVL, VST, VRT, TLNE); [FilingsFlow](https://filingsflow.com/manager/0002045724/situational-awareness-lp) (Q2 2026, filed 2026-08-14); Blockspace via Yahoo Finance, [Situational Awareness LP discloses $5.5 billion portfolio with large semiconductor puts and bitcoin miner longs](https://finance.yahoo.com/markets/options/articles/situational-awareness-lp-discloses-5-155553432.html), 2026-05-18; CNBC (Yun Li), [Situational Awareness filing shows AI bets before forced portfolio sale to Citadel](https://www.cnbc.com/2026/08/14/situational-awareness-filing-shows-ai-bets-before-forced-portfolio-sale-to-citadel.html), 2026-08-14
- Core Scientific, [rejection of CoreWeave proposal (SEC EX-99.1)](https://www.sec.gov/Archives/edgar/data/1839341/000162828024027044/ex991-corzunsolicitedpropo.htm), 2024-06-06; CoreWeave, [CoreWeave to Acquire Core Scientific](https://investors.coreweave.com/news/news-details/2025/CoreWeave-to-Acquire-Core-Scientific/default.aspx), 2025-07-07; Reuters, [Core Scientific terminates $9 billion merger deal with CoreWeave](https://www.reuters.com/business/core-scientific-terminates-9-billion-merger-deal-with-coreweave-2025-10-30/), 2025-10-30
- Lawrence Berkeley National Laboratory, [Queued Up: 2026 Edition](https://eta-publications.lbl.gov/sites/default/files/2026-06/queued_up_2026_edition.pdf), June 2026
- US EPA, [Reuse Considerations for Data Centers on Brownfield Sites](https://www.epa.gov/brownfields/reuse-considerations-data-centers-brownfield-sites); GridCensus, [Bayonne Plant Holding](https://gridcensus.com/brownfield-sites/new-jersey/bayonne-plant-holding-llc-6f9b4828); TECfusions, [Clarksville Proving Ground case study](https://tecfusions.com/wp-content/uploads/2024/09/Case-Study_Clarksville-Proving-Ground-1.pdf), 2024
- TweakTown, [SK hynix sold out of HBM for 2025](https://www.tweaktown.com/news/104307/sk-hynix-sold-out-of-hbm-for-2025-soon-taking-customer-orders-2026-supply/index.html); TrendForce, [Micron reportedly hints 2026 sell-out](https://www.trendforce.com/news/2025/08/13/news-hbm-battle-heats-up-micron-reportedly-hints-2026-sell-out-sk-hynix-yet-to-confirm/), 2025-08-13; TrendForce, [HBM3E spot prices said to be 4–5× LTA levels, with Samsung reportedly locking 70% of capacity through 2031](https://www.trendforce.com/news/2026/09/01/news-hbm3e-spot-prices-said-to-be-4-5x-lta-levels-with-samsung-reportedly-locking-70-of-memory-capacity-through-2031/), 2026-09-01
- Reuters, [GE Vernova expects 80 gigawatts of gas turbine contracts by year's end](https://www.reuters.com/business/energy/ge-vernova-expects-80-gigawatts-gas-turbine-contracts-by-years-end-2025-12-09/), 2025-12-09; GE Vernova, [29-unit aeroderivative order for Crusoe](https://www.gevernova.com/news/press-releases/ge-vernova-crusoe-announce-major-29-unit-aeroderivative-gas-turbine-deliver-ai-data-centers), 2025-07-22
- Constellation, [One year later: Crane Clean Energy Center](https://constellationenergy.gcs-web.com/news-releases/news-release-details/one-year-later-crane-clean-energy-center-still-spotlight-and), 2025-09-23; Google, [First advanced nuclear reactor project with Kairos Power and TVA](https://blog.google/company-news/outreach-and-initiatives/sustainability/google-first-advanced-nuclear-reactor-project-with-kairos-power-and-tennessee-valley-authority/), 2025-08-18
- MTS, [CoreWeave filings compilation](https://drops.mts.now/ai-capex/wiki/documents/coreweave-filings/); Sherwood News, [Michael Burry has some concerns about AI accounting](https://sherwood.news/markets/michael-burry-has-some-concerns-about-ai-accounting/), 2025-11-11; Shattered.io, [H100 rental halves to $3.38/hr as B200 hits $5.66](https://shattered.io/h100-h200-b200-cloud-gpu-pricing-2026/), updated 2026-09-22 (citing AIMultiple's Cloud GPU Rental Price Index of 2026-08-25)
- CoreWeave 8-K, [9.750% senior notes due 2031](https://www.stocktitan.net/sec-filings/CRWV/8-k-core-weave-inc-reports-material-event-0161dfbf1011.html), April 2026; CoreWeave 8-K, [9.625% and 8.500% senior notes due 2032](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000291/crwv-20260618.htm), 2026-06-18
- Valor International, [AI bets flow to Asian emerging markets, leaving Brazil behind](https://valorinternational.globo.com/markets/news/2026/08/13/ai-bets-flow-to-asian-emerging-markets-leaving-brazil-behind.ghtml), 2026-08-13; FXStreet, [Rabobank on Brazilian real fiscal risks](https://www.fxstreet.com/news/brazilian-real-fiscal-risks-point-to-weakness-against-us-dollar-rabobank-202608181240), 2026-08-18
- CNBC, [OpenAI stake talks](https://www.cnbc.com/2026/06/05/trump-open-ai-altman-stake.html), 2026-06-05; The Guardian, [OpenAI stake for US government](https://www.theguardian.com/technology/2026/jul/02/openai-stake-us-government-ai-sam-altman), 2026-07-02
- TechCrunch, [Pentagon moves to designate Anthropic as a supply-chain risk](https://techcrunch.com/2026/02/27/pentagon-moves-to-designate-anthropic-as-a-supply-chain-risk/), 2026-02-27; US District Court for the Northern District of California, [Anthropic PBC v. U.S. Department of War, No. 26-cv-01996-RFL, order on cross motions for summary judgment](https://cases.justia.com/federal/district-courts/california/candce/3:2026cv01996/465515/250/0.pdf), 2026-08-27; Jones Walker LLP (Andrew R. Lee, Michelle Ramsden, Christina Baptista), [National Security Is "Not a Blank Check": An AI Model Retaliation Ruling and the Designation It Leaves Standing](https://www.joneswalker.com/en/insights/blogs/ai-law-blog/national-security-is-not-a-blank-check-an-ai-model-retaliation-ruling-and-the.html?id=102nzx9), 2026-09-03; Foundation for American Innovation, [Anthropic's position merits support](https://www.thefai.org/posts/anthropics-position-merits-support)
- Leopold Aschenbrenner, [The Project](https://situational-awareness.ai/the-project/), Situational Awareness, 2024
- PBS American Experience, [Eisenhower's Farewell Address, 1961](https://www.pbs.org/wgbh/americanexperience/features/eisenhower-farewell/)
- Institute for Progress, [Funding for CAISI](https://ifp.org/funding-for-caisi/), 2026
- Acemoglu, Gitmez and Shadmehr, [Automation and Repression](https://economics.mit.edu/sites/default/files/2026-06/Automation%20and%20Repression.pdf), June 2026 (also [NBER w35336](https://www.nber.org/papers/w35336), [CEPR DP21594](https://cepr.org/publications/dp21594))
- Pew Research Center, [Views of risks, opportunities and regulation of AI](https://www.pewresearch.org/2025/04/03/views-of-risks-opportunities-and-regulation-of-ai/), 2025-04-03
- Data Center Watch, [$64 billion of data center projects have been blocked or delayed](https://www.datacenterwatch.org/report), 2025, [Q2 2025 update](https://www.datacenterwatch.org/q22025), [Q1 2026 report](https://www.datacenterwatch.org/q1-2026-1) and [Q2 2026 report](https://www.datacenterwatch.org/q2-2026)
- Polymarket, prices read from the public Gamma API on 2026-09-25: [Anthropic vs OpenAI, higher valuation on December 31?](https://polymarket.com/event/anthropic-vs-openai-higher-valuation-on-december-31) (Anthropic 0.915); [Will Anthropic's valuation hit (HIGH) $3.0T by December 31?](https://polymarket.com/event/will-anthropics-valuation-hit-by-december-31/will-anthropics-valuation-hit-high-3pt0t-by-december-31) (Yes 0.245); [Will the US federal government take a stake in OpenAI?](https://polymarket.com/event/which-companies-will-the-us-take-a-stake-in/will-the-us-federal-government-take-a-stake-in-openai) (Yes 0.095); [Will CoreWeave announce bankruptcy by December 31, 2027?](https://polymarket.com/event/coreweave-announces-bankruptcy-by/will-coreweave-announce-bankruptcy-by-december-31-2027) (Yes 0.105)
- Council of the EU, [Greece programme](https://www.consilium.europa.eu/en/policies/financial-assistance-eurozone-members/greece-programme/); RTÉ, [Greek airports sold](https://www.rte.ie/news/business/2017/0412/867195-greek-airports-sold/), 2017-04-12
- Reuters, [South Korea birthrate rises for first time in 9 years](https://www.reuters.com/world/asia-pacific/south-korea-birthrate-rises-first-time-9-years-marriages-surge-2025-02-26/), 2025-02-26; Korea.kr, [2025 births statistics](https://www.korea.kr/news/policyNewsView.do?newsId=148970652), 2026; Reuters, [What we know about South Korea's trade deal with the US](https://www.reuters.com/world/china/what-we-know-about-south-koreas-trade-deal-with-us-2025-07-07/), 2025; Yonhap, [US-Korea investment MOU](https://en.yna.co.kr/view/AEN20251114008554320), 2025-11-14
- Private conversations with two friends (2026-08-01, 2026-08-24, 2026-08-29, 2026-09-01), quoted and translated; my own chat with an AI agent, 2026-09-24; the June 2026 essay on the state's disempowerment, author unknown (no byline), which I shared in a reading group on 2026-06-28