I won't argue wether text to speech qualifies as an AI but I agree they must be making bank.
But the people who position themselves to profit from the energy consumption of the hardware will profit from all of it: the LLMs, the image generators, the video generators, etc. See discussion yesterday: https://news.ycombinator.com/item?id=41733311
Imagine the number of worthless images being generated as people try to find one they like. Slop content creators iterate on a prompt, or maybe create hundreds of video clips hoping to find one that gets views. This is a compute-intensive process that consumes an enormous amount of energy.
The market for chips will fragment, margins will shrink. It's just matrix multiplication and the user interface is PyTorch or similar. Nvidia will keep some of its business, Google's TPUs will capture some, other players like Tenstorrent (https://tenstorrent.com/hardware/grayskull) and Groq and Cerebras will capture some, etc.
But at the root of it all is the electricity demand. That's where the money will be made. Data centers need baseload power, preferably clean baseload power.
Unless hydro is available, the only clean baseload power source is nuclear fission. As we emerge from the Fukushima bear market where many uranium mining companies went out of business, the bottleneck is the fuel: uranium.
Trial and error content-creation using generative AI, whether or not it creates any real-world value, consumes a lot of electricity.
This electricity demand is likely to translate into demand for nuclear power.
When this demand for nuclear power meets the undersupply of uranium post-Fukushima, higher uranium prices will result.
More supply and lower prices will result.
Not unlike the recent few years in (say) lithium, anticipated demand surged exploration and development, actual demand didn't meet anticipated demand and a number of developed economicly feasible resources were shuttered .. still waiting in the wings for a future pickup in demand.
Look at how long NexGen's Rook 1 Arrow is taking to develop (https://s28.q4cdn.com/891672792/files/doc_downloads/2022/03/...). Spend an hour listening to what Cameco said in its most recent conference call. Look at Kazatomprom's persistent inability to deliver the promised pounds of uranium, their sulfuric acid shortages and construction delays.
Uranium mining is slow and difficult. Existing demand and existing supply are fully visible. There's a gap of 20-40 million pounds per year, with nothing to fill the gap. New mines take a decade or more to develop.
It is not in the slightest like lithium.
Would two decades in global exploration geophysics and being behind the original incarnation of https://www.spglobal.com/market-intelligence/en/industries/m... count?
> Once you know the facts we can talk again.
Gosh - that does come across badly.
When someone compares uranium to lithium, I know I'm not talking to a uranium expert.
All the best to you, and I'll try to be more polite in the future.
Clearly you're a better expert.
> when someone compares uranium to lithium, I know I'm not talking to a uranium expert.
It's about boom bust and shuttering cycles that apply in all resource exploration and production domains.
Perhaps you're a little too literal for analogies? Maybe I'm thinking in longer time cycles than yourself and don't a few years of lag as anything other than a few years.
You are well-prepared to familiarize yourself with the current supply/demand situation. It's time to "make bank", just like you did in 2007... only more so. The 2007 spike was during an oversupplied uranium market and mainly driven by financial actors.
I invite you to begin by listening to any recent interview with Mike Alkin.
Good night and enjoy your weekend.
Might also be an AI voice-changer (i.e. speech2speech) model.
These models are most well-known for being used to create "if [famous singer] performed [famous song not by them]" covers — you sing the song yourself, then run your recording through the model to convert the recording into an equivalent performance in the singer's voice; and then you composite that onto a vocal-less version of the track.
But you can just as well use such a model to have overseas workers read a script, and then convert that recording into an "equivalent performance" in a fluent English speaker's voice.
Such models just slip up when they hit input phonemes they can't quite understand the meaning of.
(If you were setting this up for your own personal use, you could fine-tune the speech2speech model like a translation model, so it understands how your specific accent should map to the target. [I.e., take a bunch of known sample outputs, and create paired inputs by recording your own performances of them.] This wouldn't be tenable for a big low-cost operation, of course, as the recordings would come from temp workers all over the world with high churn.)