Matt Dratch(@DratchCap):Hearing a lot about how “pacing” (forced or otherwise) lets demand catch supply… despite of course that GPU rents are rising… Sadly, I have some bad news for the people who think that, particularly in light of the ‘fighting city hall’ on Data center fear: you should still be VERY worried. What follows is a bit of a Rorshach test. (Others have covered some of this math I'm sure, but its worth repeating) Let’s say there’s ~40 GW of compute for AI by YE26 with another potential 40GW added each of the next two years. The bears say “whoa” and the pacers say “20 each year might be better!”. Well, 20GW is still A LOT more than they think due to stacked efficiencies. Rubin + serving software probably delivers 2-3x the tokens / watt and model efficiency squeezes out equiv work for 2-3x fewer tokens ish. Adding in efficiency for the existing fleet, and you could be looking at >10x the effective compute in a bull case (in 2026 terms). This is all napkin math, obviously (please don’t shoot me for an off figure). The spenders (dreamers?) think that’s fine bc of similar gains on the training side. Call a recent frontier run maybe 300-400mw, probably less. Moving training clusters to 1GW or beyond, with 1.5x more training compute per watt and 2-3x better algos, and you’re looking at ~10x the effective training compute too (Pre-training gets smaller hardware gains than inference BUT much of RL post-training *is* inference so efficiencies matter there) This is simply a bet that scaling laws and the Bitter Lesson hold. More powerful, generalized intelligence becomes useful for more things and easier to adopt. That’s essentially the entire macro tech debate. ROI / TAM etc all sit downstream. I think the evidence leans that way, and low GPU availability and increased pricing are consistent w/ demand absorbing the gains to date. Skeptics point to circularity, and VC spend. Another glass half empty view. Funding customers can seed demand, which is how a dynamic economy develops new markets. It does still need customers who pay their own way but they can be grown. Taken far enough, the argument they can’t would suggest the whole fractional-reserve banking system has created one big economic fugazzi. Speaking of Fugazzi, back to what’s left of my portfolio.

2026.11.06 01:10

Hearing a lot about how “pacing” (forced or otherwise) lets demand catch supply… despite of course that GPU rents are rising… Sadly, I have some bad news for the people who think that, particularly in light of the ‘fighting city hall’ on Data center fear: you should still be VERY worried. What follows is a bit of a Rorshach test. (Others have covered some of this math I'm sure, but its worth repeating) Let’s say there’s ~40 GW of compute for AI by YE26 with another potential 40GW added each of the next two years. The bears say “whoa” and the pacers say “20 each year might be better!”. Well, 20GW is still A LOT more than they think due to stacked efficiencies. Rubin + serving software probably delivers 2-3x the tokens / watt and model efficiency squeezes out equiv work for 2-3x fewer tokens ish. Adding in efficiency for the existing fleet, and you could be looking at >10x the effective compute in a bull case (in 2026 terms). This is all napkin math, obviously (please don’t shoot me for an off figure). The spenders (dreamers?) think that’s fine bc of similar gains on the training side. Call a recent frontier run maybe 300-400mw, probably less. Moving training clusters to 1GW or beyond, with 1.5x more training compute per watt and 2-3x better algos, and you’re looking at ~10x the effective training compute too (Pre-training gets smaller hardware gains than inference BUT much of RL post-training *is* inference so efficiencies matter there) This is simply a bet that scaling laws and the Bitter Lesson hold. More powerful, generalized intelligence becomes useful for more things and easier to adopt. That’s essentially the entire macro tech debate. ROI / TAM etc all sit downstream. I think the evidence leans that way, and low GPU availability and increased pricing are consistent w/ demand absorbing the gains to date. Skeptics point to circularity, and VC spend. Another glass half empty view. Funding customers can seed demand, which is how a dynamic economy develops new markets. It does still need customers who pay their own way but they can be grown. Taken far enough, the argument they can’t would suggest the whole fractional-reserve banking system has created one big economic fugazzi. Speaking of Fugazzi, back to what’s left of my portfolio.