95% will fail, 4% will become stable projects that throw out enough cash to support their developer (plus a team of 2 - 4), 0.99% will become venture-scale companies, and 0.01% will grow to Google/Microsoft/Apple-scale.
The only relevant question for VCs is whether or not that top 1% has already launched or not. If they have they're too late. If they haven't then continued investing in AI makes sense from their perspective. Given the feeding frenzy it looks as though the general consensus is that that 1% hasn't launched yet (nor an appreciable fraction of it).
In other words, VC analysts are expecting one or more key breakthroughs are still in the pipeline, and given how broad the investments are they haven't a clue as to where those breakthroughs will happen. I'm privy to some of these conversations (but in a European context) and even though I think that in principle they are right I'm not necessarily bullish on the idea that the larger part of that 1% is still to come. What GPT-5 will do when it is released will determine to what degree free money will still be thrown at AI start-ups, if it underwhelms I expect this source of funding to be reduced fairly quickly. It's the .com boom all over again: everybody wants to get on the train but nobody has any idea where the train is going to.
But in the very long term (beyond the horizon defined by the typical VC cycle) the impact will be massive.
The IBM PC was what everyone outside a few niche companies copied. IBM Simon beat the iPhone by over a decade and probably could have been significant with a few iterations. Almost every modem used to connect to the early commercial internet slotted into an IBM clone. They built the wagon, but every time they had to make a choice on how to proceed, they chose poorly.
Microsoft spent a long time resting on its laurels before starting to innovate again, but then they fell back into their old ways. Every smart move is hampered by the continuing need to subsume innovation to serve a desktop OS monopoly with declining relevance. LLMs are headed away from the silos that birthed the revolution toward on-board purpose-built LLMs that are entirely within the user's control, so Copilot's days are numbered. Even integrated GPUs will have enough power and memory to run the better LLMs in 5-10 years.
Apple: back to selling overpriced PCs. Bringing SOCs in from mobile to the desktop is an interesting move, but Microsoft and Windows PCs are likely to beat them on price and features once that world catches up. They're in for a period of mistaking profit growth for innovation the way all big companies do. The pivot to services will be fraught, hindered by corporate inertia, and possibly kill the company. They've had on-board ML cores in all their mobile devices for years, and the best they could do was Siri.