Arrow and Debreu posited the idea of complete markets, where every state of the world could be traded. https://en.wikipedia.org/wiki/Complete_market That's an aspiration, not a possibility.
But enough states can be approximated that we benefit from the widest "wisdom of crowds" in history.
Your idea of compute vs. chips has merit, but will face the same challenge as mortgage backed securities, 50 years ago. With MBS, it immediately ran into "cheapest to deliver." That means the lowest value collateral always appeared, no matter how detailed the definitions.
For chips, we should discuss better metrics/ Best we can do is minimize the variance and structure offsets.
You’re right to flag the “cheapest-to-deliver” problem. GPUs would face an echo of that, e.g. providers delivering the lowest-value slices of compute. The wrinkle is that unlike mortgages, you can at least try to anchor contracts in audited throughput benchmarks that reduce (though never eliminate) variance.
That’s why I take Don Wilson’s Compute Exchange seriously: once you have an auction platform and transparent spot/forward trade, the pressure to extend into futures is almost irresistible.
As for the Trump AI Action Plan, it’s only advisory, but it shows Washington is at least thinking in market-structure terms about compute. That intellectual direction of travel matters, even if the execution is still uncertain.
Arrow and Debreu posited the idea of complete markets, where every state of the world could be traded. https://en.wikipedia.org/wiki/Complete_market That's an aspiration, not a possibility.
But enough states can be approximated that we benefit from the widest "wisdom of crowds" in history.
Your idea of compute vs. chips has merit, but will face the same challenge as mortgage backed securities, 50 years ago. With MBS, it immediately ran into "cheapest to deliver." That means the lowest value collateral always appeared, no matter how detailed the definitions.
For chips, we should discuss better metrics/ Best we can do is minimize the variance and structure offsets.
You’re right to flag the “cheapest-to-deliver” problem. GPUs would face an echo of that, e.g. providers delivering the lowest-value slices of compute. The wrinkle is that unlike mortgages, you can at least try to anchor contracts in audited throughput benchmarks that reduce (though never eliminate) variance.
That’s why I take Don Wilson’s Compute Exchange seriously: once you have an auction platform and transparent spot/forward trade, the pressure to extend into futures is almost irresistible.
As for the Trump AI Action Plan, it’s only advisory, but it shows Washington is at least thinking in market-structure terms about compute. That intellectual direction of travel matters, even if the execution is still uncertain.