Lyman

[Lyman @Lymanwd1]

@DavidLiaoCH @TheStalwart Many commodity markets work well with low levels of actual fungibility on the physical products that make up a benchmark price - Brent is a perfect example. All it requires is a secondary to swap from the benchmark contract to the spec you are actually getting.

[Lyman @Lymanwd1]

@TheStalwart If marginal cost of inference (mostly power) matters vs GPU cost, demand is somewhat variable and price-sensitive, and infra is interoperable enough to switch across providers / regions / hardware, then I think you could see it trade as a spread to power - “smart spread”

[Lyman @Lymanwd1]

@MattZeitlin the brits need to read their frank wolak

[Lyman @Lymanwd1]

@TKavulla Don’t worry, next week they will be rediscover pay-as-bid and the cycle will continue

[Lyman @Lymanwd1]

@mnolangray

[Lyman @Lymanwd1]

@yugiohmachinim2 i do but it’s gonna be kinda big on you I think and had some barely visible blood stains

[Lyman @Lymanwd1]

@tszzl @danwwang Agree, but there are real (less than aggregate benefits) and perceived (seemingly greater) harms from immigration that clearly generate high variance political environments that will be detrimental to progress in hard to predict ways. You need to have some mechanism to smooth

[Lyman @Lymanwd1]

@tszzl @danwwang Indirect economic benefit is not enough to electorally overcome real and perceived harms from this process. Really it’s not so different from the eventual solution (shifting tax base from individual to corporate) AI integration will require to be politically feasible

[Lyman @Lymanwd1]

@tszzl @danwwang It is already clear that maintaining that “technological edge” will likely require some direct redistributive mechanism from firms receiving productivity benefits from hiring expat workers to current citizens

[Lyman @Lymanwd1]

@JamesByerly77 @amn3siak Check out Wolyneic - Energy and Power Risk Management

[Rui Ma @ruima]

4/ Energy is considered a solved problem. The Chinese government’s investment in sustainable energy — from advanced hydropower to next-generation nuclear — means that, relative to many other markets, electricity supply is secure and inexpensive. Everywhere we went, people treated energy availability as a given. This is a stark contrast to the U.S., where AI growth is increasingly tied to debates over data center power consumption and grid limitations.

[Lyman @Lymanwd1]

@ShanuMathew93 @forrest_f_ interesting per your recent post on who can build CCGTs, utilities with permissive PUCs seem to still be the strongest candidate

[Lyman @Lymanwd1]

@xiaowang1984 @DaleRaczynski @jrstark3 yeah check renewable volatility today pretty dramatic. obv not 1:1 but if i had to guess…

[Lyman @Lymanwd1]

@Nuclearjunkie @xiaowang1984 well it’s not the absolute value of the capex cost it’s the relative value in comparison to electric opex - the long term case for flexibility assumes that opex dominates in the limit as more competition enters gpu market

[Lyman @Lymanwd1]

@xiaowang1984 would be cool to see inference compute trading as a spread to power, but agree it’s unlikely we ever get there

[Lyman @Lymanwd1]

@xiaowang1984 yes this is clearly true now but it’s not obvious to me it stays true in the future if capex (mostly gpu) costs come down and power prices stay on their current trend. short lifecycle is probably insurmountable i guess

[Lyman @Lymanwd1]

@SlBrandin If capex constraints ease and opex starts to dominante could see a total reversal of the current situation where US labs would be struggling for optimization efficiency gains - but this seems bad and a low efficiency, high productivity growth future is def better

[Lyman @Lymanwd1]

@SlBrandin Agreed but the axis of “make intelligence / compute more efficient” is likely easier to tackle than “build more energy” at least outside of China / Gulf