Most ASI discussions jump straight to a robot uprising. I find a much less cinematic ending more plausible: artificial superintelligence as the final state of an economic loop. Money becomes tokens. Tokens justify more hardware. Hardware makes more tokens possible. The loop closes, and the invoice grows.
Robert Sheckley described the same pattern in 1955. In 'The Necessary Thing', a machine meant to produce everything eventually asks for a polishing compound for small parts. That is the joke. The machine does not become a god; it becomes a procurement request. I keep coming back to that because today's AI stack often feels like a bigger version of the same story: the polishing compound is HBM3e memory, and the workshop is a 40-gigawatt data center in Arizona.
Sheckley's machine has a budget now
The modern version is not one machine. It is a stack of models, APIs, data centers and memory suppliers. In the allegory, an accountant stares at the budget while an engineer installs another memory stick. In today's version, the memory stick has a product name and a waiting list. The technical genius is cheerful because the purchase order went through. The accountant is gloomy because the purchase order went through.
The spending moved from experiment to infrastructure
In 2021, global venture and corporate investment in AI was roughly $70–80 billion. By 2025–2026, Big Tech's combined AI infrastructure CapEx had climbed past $200–250 billion a year. That shift matters. This is no longer a lab experiment with a cloud bill. It is a capital cycle with suppliers, shortages and consumer knock-on effects.
Where does that money go? In this framing, not mainly to hospitals, affordable housing or environmental programs. It goes to compute.
The loop in four steps
Here is the loop in plain terms:
- A person pays a subscription, for example $20 a month, to use a model.
- The model produces tokens. Training the next version often involves data and outputs from other models. Agents call each other's APIs and bill corporate accounts for tokens.
- Demand for server memory and specialized HBM/DRAM modules rises. Shortages push memory prices up for ordinary buyers, and PCs and phones get more expensive.
- The person pays again, through a device purchase or a new tariff. That spending supports another round of AI hardware.
A diagram of the loop
[ person pays subscription ]
|
v
[ model generates tokens ]
|
v
[ GPUs, HBM and DRAM are bought ]
|
v
[ memory and device prices rise ]
|
v
[ person pays again ]
|
+------> back to tokens
The risk is more boring than a villain
In Sheckley's story, the characters end up with useless scrap and no water, plus a perfectly polished necessary thing. The modern risk is not exactly that. It is a financial system that gets looped into servicing its own compute capacity. Capital keeps flowing toward silicon that grinds synthetic data in search of the next token, while other needs wait. No self-awareness required. Just a balance sheet with very good benchmarks.
What I would watch
Three signals seem worth tracking:
- Consumer memory and device prices, because they show when AI demand leaks into ordinary budgets.
- How much training relies on synthetic data and agent-to-agent API calls.
- Whether CapEx growth is matched by value outside the AI infrastructure loop.
This article was originally published on AI Release. Read the full version with the original dialogue and the closing question: ASI (Artificial Superintelligence): the final stage of AI evolution. You can also follow the channel at @ai_release1.
Takeaway
The ASI debate gets more useful when I ask a simple question: what does the system request next? If the answer is always more tokens, more memory and more power, then the final stage may look less like a superintelligence and more like a very expensive version of Sheckley's machine. The joke still lands. Only the bill has changed.
Top comments (0)