In junior high, I wanted a specific pair of Nike sneakers. A bunch of the girls at school had them and, in junior high, that is enough. But this was around the time my dad’s business began to go south, and we were barely getting by. So when I finally got a pair of Nikes, I wore them until they fell apart. Now that I can buy my own, I’ve got a mini-collection going. I thought of this as I learned about recursive self-improvement (RSI) in AI. RSI means the models are discovering better algorithms for themselves, without human researchers in the loop. What makes recursive self-improvement qualitatively different from past algorithmic efficiency gains is the feedback loop structure. If RSI is working, then the algorithmic efficiency curve stops being a slow human-driven process and becomes an automated loop running continuously. 

And if you can get 10x smarter models for the same compute through better algorithms, you might not need 10x more compute. Your sneakers (compute) take you further.

Recent history shows a couple of examples supporting RSI’s existence:

• The Anthropic Institute shared that their engineers on average ship 8x as much code per quarter as they did from 2021-2025.

• DeepSeek's models are matching US frontier model performance at a fraction of the training cost through architectural innovations. 

• So-called Chinchilla scaling laws (DeepMind, 2022), showed that most large language models of the era were undertrained relative to their compute budget. The paper found that model size and training tokens should scale roughly in proportion, meaning a smaller model trained on significantly more data could outperform a much larger one trained the conventional way. This effectively cast doubt on the efficiency of large-scale investments in oversized, undertrained models.

There are, of course, no guarantees. RSI could simply not truly come to fruition. We’ve outlined 3 scenarios anyway, as a thought exercise for ourselves:

Scenario 1: Compute scaling remains primary: Intelligence continues to scale primarily with hardware. More GPUs and more power equal smarter models. With this, the physical bottleneck in transformers, power, cooling, InP, and packaging continue to be a multi-year that benefits the related companies, particularly in the factory layer.

The evidence for this scenario: frontier model training runs are still getting bigger. Vera Rubin is NVIDIA's next GPU architecture generation continuing the generational step-up from Hopper to Blackwell. 

Leopold Aschenbrenner, co-founder of the fund Converge, and deeply connected in Silicon Valley, made several predictions in his 2024 manifesto, Situational Awareness. This included assuming intelligence scales primarily with compute. If this scenario is true, the physical buildout is essentially demand-certain meaning there is a need to continue to build faster than the intelligence curve rises.

Scenario 2: RSI compresses compute needs:

In this scenario, RSI fully fruits, drives algorithmic gains such that the intelligence per unit of compute rises faster than demand for intelligence. This could mean less hardware is needed to achieve the same capability. The physical infrastructure companies get hurt, but not because demand for intelligence disappears, but because the compute intensity per unit of intelligence falls faster than intelligence demand grows.

The evidence for this scenario: The Chinese open-source model release where intelligence is getting cheaper, faster than usage is growing. This is what you'd expect in a scenario where algorithmic efficiency is outrunning compute scaling.

Scenario 3: Jevons Paradox, where both scenarios 1 and 2 happen + my view:

In this scenario, algorithmic efficiency rises through RSI, but the demand for intelligence expands faster than efficiency rises. Cheaper intelligence enables new use cases that weren't previously viable. Eventually, you end up needing more compute, not less, even as each unit of compute becomes more capable. In other words, the total addressable market for intelligence increases faster than efficiency gains compress the hardware needed per unit.

Besides my personal parallel of having less sneakers to have more later, historical evidence is in electricity. US electricity consumption grew 10x between 1950 and today, even as the real cost of power fell 60% because of efficiency gains. The refrigerator is 75% more efficient than in 1975, but we bought more of them, stocked them fuller, and kept them colder. 

A similar dynamic also played out with computing: per-operation energy cost has dropped by more than 10,000x since the 1970s, yet data center power demand grew around 3x in the last 20 years alone.

Looking ahead, it’s possible to have a period of a few quarters where RSI-driven algorithmic efficiency temporarily reduces the perceived need for incremental hardware before Jevons demand expansion catches up. If one or more hyperscalers reduce or delay capex guidance on an earnings call, this could exacerbate some of the worries that were in the markets last week, and impact the physical infrastructure trade. 

If Jevons Paradox is right, and you have investments in the potentially impacted areas, it becomes a choice of trying to time it right by selling and buying back, or holding. 

To me, timing it perfectly isn’t easy. So instead, questioning how much of the physical infra/factory layer to hold now, in exchange for the long-run thesis, is the consideration, and the key question behind any risk tolerance stance.

We, in Robinhood Strategies, have trimmed a small amount ourselves within the last few weeks, but we’re not walking away (in our Nikes).