Over the same few weeks, I watched the animated series Pantheon and read Leopold Aschenbrenner’s Situational Awareness: The Decade Ahead. I’m writing these notes because I kept bringing both back to the question I follow through technology finance: what has to be built and funded for an idea about AI to become possible?
Pantheon begins with a girl hearing from her dead father, whose mind has been uploaded to a computer. The story makes dependence on computing personal. Keeping that intelligence running matters to someone who loves him, while the companies that control uploaded people can use their skills as labor.
Aschenbrenner starts from somewhere else: not the possibility of uploading a human mind, but progress in AI systems trained on computers. His forecast is that AI could become capable of accelerating its own development, with consequences large enough to demand a national-security response.
I don’t read either work as evidence that the other’s future will happen. What I find useful in the pairing is a question: who supplies the machines on which intelligence would depend? Because Aschenbrenner also runs an investment fund, I can set that question beside a public record of the businesses his fund holds.
Reading the infrastructure forecast
The part of Situational Awareness I want to examine is the construction it implies. Aschenbrenner argues that improvements in computing power and algorithms, along with better ways of using models, could produce systems capable of automating AI research, and that those systems could then help develop their successors. He calls AGI by 2027 plausible.
I’m not relying on that date to evaluate the infrastructure argument. Even a slower expansion would need processors housed in facilities with enough electricity and cooling, and larger clusters would need fast connections between machines. What I want to understand is which of those requirements could become difficult to satisfy as demand grows.
This is where the forecast meets my interest in finance. Buying processors and securing a power connection are different commitments on different schedules. A company can have a promising model and still depend on someone else for enough capacity to run it. I’m interested in the businesses that can meet those requirements, and in what the demand would mean for their earnings.
The fund’s reported holdings
Situational Awareness LP’s filing for the end of 2025 shows several businesses connected to that buildout. I’ve selected ordinary-share positions so that the reported values all refer to the same kind of holding.
| Business | Connection to the infrastructure argument | Reported share value |
|---|---|---|
| Bloom Energy | Power generation systems | $875.5 million |
| Lumentum | Optical networking | $478.6 million |
| CoreWeave | GPU cloud services | $436.7 million |
| Core Scientific | Data-center infrastructure | $418.7 million |
| IREN | Data-center and AI cloud infrastructure | $328.6 million |
| Applied Digital | Data-center infrastructure | $278.0 million |
Values reported for December 31, 2025; these are not the fund’s purchase prices.
The holdings connect the broad forecast to actual businesses, but I can’t reconstruct every investment decision from a quarterly filing. The filing covers certain reportable securities, not the fund’s complete exposure, and this table doesn’t tell me what the manager thought each business was worth.
What interests me is how different the underlying investment questions are. A GPU cloud operator has to pay for equipment and deliver a service before it can collect years of customer revenue. An optical component supplier sells into a different part of the buildout. Both can benefit from more computing demand, but to understand their economics I need to examine their customers and obligations separately.
The company questions I’m taking from this
Pantheon puts the ownership of computing systems in human terms. Situational Awareness makes me consider the scale of systems an ambitious AI forecast would require. The portfolio brings me back to something I can investigate company by company.
For each holding, I want to know what requirement the business can fulfill and what could prevent it from delivering. A signed customer contract would send me to the equipment purchases and payment schedule; a planned facility, to its power arrangements and construction progress. Those details are where I can start assessing a business without first deciding whether to accept a forecast of superintelligence.