A few things stood out to me in this piece on compute markets, where AI companies buy access to someone else’s GPUs to train models or run their services ↓ → GPU deals still happen in group chats A lot of that computing power is still bought through brokers, private deals, and group chats. Comparing prices across providers is a mess. TBH, that’s sorta crazy considering how far this tech has come. → Compute could trade like electricity The article argues GPU rentals could become a commodity market, with pricing benchmarks and contracts to manage future costs. Electricity markets started going through a similar shift in the 1990s as they opened up to competition. → The same chip doesn’t mean the same deal Location, reliability, and rental terms all affect the price. Cheap GPU access that can get interrupted is way different from guaranteed availability. An AI chatbot needs to stay available for users, while a company using AI to label images overnight might accept pauses to pay less. → NVIDIA is basically the central bank for compute Its release schedule influences how quickly older GPUs lose value, and it also benefits from keeping existing GPUs busy, since demand for that computing power gives data centre operators a reason to buy more chips. → The opportunity goes beyond renting GPUs Companies are also working on pricing tools, trading platforms, lending, and insurance. The author’s bet is that companies combining reliable GPU access with those financial services could have an edge. Silicon Data says it tracks 95% of neocloud GPU providers and 100% of major hyperscalers. That rental data is already being used to build daily pricing benchmarks for compute. So I think that gives us a glimpse of how GPU rentals could grow into a wider financial market.