Is GPU compute a financial engineering problem? Framing AI lab economics around 80% margins

sarahdrinkwater · x · 2026-09-16

A practitioner thread on AI lab economics:

@thdxr proposes a framework: isolate the core number — can you buy GPUs, run software, and sell at a profit (reported 80% gross margin)? Then ask how much cheaper it gets at scale, how much you can charge while still growing fast, and separately whether it's worth giving 20% to channel partners like AWS and Google for added capacity and demand. R&D can be funded from profits, or you can raise money to go faster in a hyper-competitive space.
The reposter agrees GPU compute is becoming a financial engineering problem: prices keep rising, but innovation is faster, and new ways of leveraging compute will make the economics work.
He's bullish on disaggregated compute — splitting the monolith into prefill, decode, storage and optimizing each — noting AMD's Cerebras partnership signals a ready-made solution emerging.

More from Venture