Kush Bavaria

The through-line

Kush Bavaria's fixed point is that compute is the commodity the next economy runs on, and that it needs the market machinery any commodity of that importance eventually acquires. He states the company purpose flatly: "Mission of the company is build markets for compute," resting on the belief that "compute will power every single enterprise the same way oil did in the 1900s" [1]. Everything else he says follows from treating that analogy seriously rather than rhetorically: if compute is oil, then it gets priced, traded, shorted and margin-called, and the interesting questions are financial ones [1][2][3].

On compute as the new oil

The comparison to oil in the 1900s is the load-bearing claim [1]. It is not a statement about energy consumption but about economic position: the input that every enterprise, regardless of sector, ends up depending on, and therefore the input whose price becomes a matter of general concern rather than a line item for specialists [1]. The framing extends to the idea that the world's most important commodity is no longer oil at all, with compute taking its place as a tradable asset [3][2].

On putting a price on intelligence

Building markets is the operative verb. The point of a market is that intelligence acquires a ticker, and once it has one it behaves like any other traded commodity [1]. Bavaria engages directly with the consequences: what happens when a hedge fund takes a short position on the price of compute, and what happens when a GPU shortage triggers a margin call, both of which he treats as live scenarios rather than thought experiments, reading them against market dynamics over the April to August period [1][2]. The through-line here is that compute scarcity is no longer only an engineering constraint. Once compute is financialised, a hardware shortage transmits into balance sheets and forced liquidations, which is precisely why the infrastructure for pricing and trading it needs to exist ahead of the stress rather than after it [1].

On simulated populations beating survey research

Drawing on his exposure to Aru, a company running large-population simulations built from AI agents, Bavaria argues that synthetic respondents already outperform real ones for forward-looking questions [1]. His worked example is a stroller company that wants to know what new mothers will actually choose: rather than survey them, run the simulation across the agent set and read off which product is preferred [1]. The finding he singles out as most interesting is that the advantage is greatest where human self-report is weakest. "If you ask humans like hey like do I prefer this or this in two or 3 months from now the humans tend to be more wrong compared to the AI," he says, attributing the gap to human bias, and noting the company has published studies backing the result [1].

Takeaways

• The company's stated purpose is singular: "build markets for compute," on the premise that compute will power every enterprise the way oil did a century ago [1].
• Compute is best understood as a tradable commodity displacing oil as the world's most important input, not merely as a technical bottleneck [1][3].
• Financialising compute brings hedge fund short positions and GPU-shortage-driven margin calls into scope as real market events, read against April-to-August dynamics [1][2].
• Large-scale agent simulations can replace conventional consumer research, letting a company test a product against a synthetic population instead of surveying real customers [1].
• Humans are unreliable about their own future preferences, so AI-simulated respondents outperform them on two-to-three-month-ahead questions because they avoid that bias [1].