The expectation is of course that that will not be required. Competent models will become smaller and the compute to run them will become cheaper, once the market stabilizes. That will take some time of course.
replyEventually yes. You can buy greeting cards today for $5 that have more compute power than the Eniac.
replyYes because the models will be etched straight into silicon.
replyNot today. But just like the average person can now afford a smartphone which is more powerful than a workstation not long ago, we will have enough computing power for good enough local inference, at an accessible price.
replyWe're not in the 2010s anymore, consumer hardware is actively regressing. Smartphones are actually a great example to bring up, because I believe 2026 was the first year in which smartphone hardware did not advance at all (and arguably declined) relative to price, due to AI-related shortages.
replyNobody can predict the future, of course, but I personally believe this pattern will hold. A decade from now, the idea of a consumer being able to purchase a personal device with more than 8GB of RAM will be a thing of the past, all significant compute will be done in the cloud using the massive amounts of hardware being hoarded by corporations as we speak.
A decade from now, the idea of a consumer being able to purchase a personal device with more than 8GB of RAM will be a thing of the past
replyThe only way this is remotely possible is if the copyright cartels get their way and make it illegal to have unlicensed local compute. Assuming that doesn't happen, I'd gladly bet a large amount that computing will be much more affordable 10 years from now.
I think one of two things is more likely: either at some point the bottom falls out on the $1T/year datacenter capex we have today, or it doesn’t and eventually supply will enormously expand as new fabs being planned today eventually come online.
reply