My gut feeling: compute might not be the first thing for which futures markets naturally work—even if the theoretical case is compelling.
Two reasons come to mind—though I know I’m not strictly proofing anything and wonder whether I might be simplifying too much:
A) Large share of demanders don’t know their demand beforehand
Futures markets make the most sense when both sides have clearly foreseeable needs:
Supply = OpenAI et al.: they can plan and invest now—makes sense.
Demand = much fuzzier. Individual firms don’t know their quantity demanded in advance; their demand curve as (qty, price) relations can fluctuate widely as fct. of future developments they don’t control: Big compute demand tends to rise and fall with the whims of the broader data economy (and data tech advancements specifically even for given final demands for goods and services).
By contrast, think of the electricity market: Power plants know their output years in advance. Industrial buyers and data centers often know large parts of their demand 5–10+ years out. ==> We have deep futures markets. Where demand isn’t clearly known, buyers rationally, complacently just act as price takers.
B) Buyers face no strong local constraints or monopolies
Firms (or individuals) may typically seek long-term contracts when they’re locked into a particular supplier—e.g., food or housing markets with low local competition. But for compute, there’s global supply and lots of competition. If prices rise in one spot, you can often source capacity elsewhere. That reduces the pressure to hedge and makes price-taking a pretty reasonable default.
==> Might compute simply remain a exactly not natural candidate for futures markets? I don’t think ‘my’ demand for compute (as individual or as firm) will become really long-term predictable. Maybe it’s because if I’m big in the compute business, I mainly do things that esp. these days anyway are every few months or years rewritten how they’re done in terms of compute.
Maybe there’s also something of ‘I don’t really know my compute demand before I have computed my answer’; at least in my computational modelling jobs that’s really quite true; my eventual compute demand can vary x-fold (x easily 10+) compared to my original predictions, even for an already precisely defined project and only few weeks or months ahead. (That said I recently bought some more compute hardware myself because cloud felt so expensive, which might speak against my own argument)
For compute futures specifically, AI labs who are building their own data centers are on both sides of the equation: they supply the compute (from their data center), but they also consume the compute (their big training runs). So, implicitly, these labs are already forecasting their compute demand, and they are building data centers in anticipation of that demand. So we have empirical evidence that firms can forecast compute needs.
I do agree with the general sentiment that maybe demand for compute/intelligence feels less predictable than say electricity, but it feels like the uncertainty is mostly right tailed, e.g. I know I will need at least this much compute / intelligence, but there is a chance that I need 10x more than that.
My gut feeling: compute might not be the first thing for which futures markets naturally work—even if the theoretical case is compelling.
Two reasons come to mind—though I know I’m not strictly proofing anything and wonder whether I might be simplifying too much:
A) Large share of demanders don’t know their demand beforehand
Futures markets make the most sense when both sides have clearly foreseeable needs:
Supply = OpenAI et al.: they can plan and invest now—makes sense.
Demand = much fuzzier. Individual firms don’t know their quantity demanded in advance; their demand curve as (qty, price) relations can fluctuate widely as fct. of future developments they don’t control: Big compute demand tends to rise and fall with the whims of the broader data economy (and data tech advancements specifically even for given final demands for goods and services).
By contrast, think of the electricity market:
Power plants know their output years in advance. Industrial buyers and data centers often know large parts of their demand 5–10+ years out. ==> We have deep futures markets.
Where demand isn’t clearly known, buyers rationally, complacently just act as price takers.
B) Buyers face no strong local constraints or monopolies
Firms (or individuals) may typically seek long-term contracts when they’re locked into a particular supplier—e.g., food or housing markets with low local competition. But for compute, there’s global supply and lots of competition. If prices rise in one spot, you can often source capacity elsewhere. That reduces the pressure to hedge and makes price-taking a pretty reasonable default.
==> Might compute simply remain a exactly not natural candidate for futures markets? I don’t think ‘my’ demand for compute (as individual or as firm) will become really long-term predictable. Maybe it’s because if I’m big in the compute business, I mainly do things that esp. these days anyway are every few months or years rewritten how they’re done in terms of compute.
Maybe there’s also something of ‘I don’t really know my compute demand before I have computed my answer’; at least in my computational modelling jobs that’s really quite true; my eventual compute demand can vary x-fold (x easily 10+) compared to my original predictions, even for an already precisely defined project and only few weeks or months ahead. (That said I recently bought some more compute hardware myself because cloud felt so expensive, which might speak against my own argument)
For compute futures specifically, AI labs who are building their own data centers are on both sides of the equation: they supply the compute (from their data center), but they also consume the compute (their big training runs). So, implicitly, these labs are already forecasting their compute demand, and they are building data centers in anticipation of that demand. So we have empirical evidence that firms can forecast compute needs.
I do agree with the general sentiment that maybe demand for compute/intelligence feels less predictable than say electricity, but it feels like the uncertainty is mostly right tailed, e.g. I know I will need at least this much compute / intelligence, but there is a chance that I need 10x more than that.