Comment on Will AI Startups End Up Like Blockchain Startups?
chicken@lemmy.dbzer0.com 6 days agoPeople can easily self host email, file backup, etc but pay for service anyway
Who pays for email? Who pays subscriptions for file backup? Why would you when you can just use another companies service that is free? Self hosting AI is increasingly viable, but that isn’t even the problem for companies hoping to make billions on it, the problem is that as soon as they try to put the squeeze on their customers they will just go somewhere else that offers the same thing. Look at what happened with Deepseek; OpenAI can’t maintain dominance.
AI will be prohibitively expensive to self host for a very very long time.
It already isn’t, there are tiny models that are practical for some things that will run on basically anything, and there is a lot you can do with a mid to high end graphics card. Nvidia is artificially limiting vram but that’s not going to remain the limitation for long. But even if AI running on datacenter hardware maintains a big advantage, that’s not enough for these companies to make huge profits.
Blue_Morpho@lemmy.world 6 days ago
That you don’t pay for gmail doesn’t mean it isn’t monetized. Google makes $220 Billion a year from ads.
Google drive, Onedrive, and iCloud are forced down all users throats. Millions pay. A quick google says Apple makes $96 Billion a year from iCloud alone.
If it were that easy Google wouldn’t make $350B a year. Google is just a search index and not very good.
Yes, and Google displaced Yahoo. That doesn’t mean search stopped being monetized.
It’s not VRAM, it’s ram in general. Deepseek needs ~700GB. Ram prices have been dropping about 30% every 5 years (and that’s ideal numbers. Moore’s law is dead. Current improvements are smaller than looking back 5 years). It’s about $6k to build your own Epyc for 6 tokens/sec. No average person is going to spend $6k when a service does it for $10 a month and is 10x faster. Do the math for how long it will be until $6k of ram costs $600.
But the real killer that will keep AI out of consumer hands for forever is that it is reliant on new models to include updated information that was scraped from the Internet. If you ask Deep Seek about anything that’s happened in the past year, it can’t give any answers because the model was built with older data. So Google, Apple, Microsoft and anyone capable of indexing the entire web will always have an AI model that is updated with the latest information. Home users will be lucky to get models with information that is several years out of date. Building models requires 100x the resources as running the trained model and you need the data to train it.