Anthropic lost a cool $80 billion in the first half of the year. I see that they have about 20 million users. That works out to $8000 that they’re burning per user per year in order to get you to use their product.
Of course, the AI gold rush isn’t looking to break even. What will you do if your AI costs go up to, say, $20k a year?
At that price I’d have no choice but to use whatever free/cheap alternative AI is available, and if there is no such AI, I lived perfectly well from age 0-35 during the era before AI. There’s always plain Google search.
If the AI was any good it would have made me rich enough to afford it.
Reminds me of this meme that’s been floating around for a while now.
It’s a different question I think – “Will you stop using it, or will you pony up?”
I pay for 3 AI subscriptions. I will cancel 2 of them and only keep 1.
Also AI keeps advancing. The most advanced programs today are the cheap, free programs a year from now. AI has gotten so advanced that the cheap programs meet my needs now.
Which is why the big players are trying to get laws passed that would effectively shove those free models out of the market, leaving themselves as the only option.
I dispute the premise of the question. The driving impetus behind the current AI frenzy is the hope of reaping enormous profits by displacing low- to medium-tier knowledge workers with AI. It was never about earning big money from home subscribers using AI as a question-answering tool. If that basic service doesn’t continue to be free, it will certainly be very low-cost as a continuing demo.
At which point you’d get chip-runners selling folks on the street 4 megsterr of hot RAM. A functional AI model that can be run locally isn’t all that large.
I don’t want to be part of any reality that ends up with Linda Lee lasered to death. 
I think there’s a parallel thread for what companies will do – my previously company started talking about going all-in on AI, cutting employees, etc, and I kept asking the execs this question and nobody seemed to appreciate it. Microsoft was basically giving Copilot away for free and they were too short-sighted to even consider what would happen if they fired half their staff and then their AI bill surged past what they were paying employees.
But I don’t think for a second that anyone is going to offer a free product “for demo purposes.” Big tech will take every dollar they can from a user, even if it’s “watch 3 minutes of adds before you can submit your prompt.” Enshittification will come for everyone.
I simply won’t be able to afford it and will discontinue use. Right now I pay $100/mo for one AI service, which is already a lot, but worth it.
In a world where that happens but somehow hardware (GPU and memory) becomes affordable again, then there is the alternative of running AI on your own machine. You can already do that even today for less than $20k/year, e.g. with a good Mac Studio or a Nvidia Spark… anything with a good GPU and fast RAM.
There are some free and open-weight models that are maybe, I dunno, 70% as capable as the frontier models? There is a noticeable capability gap (and they’re much slower on home computers than in big data centers), but they’re still a LOT more AI capability than the world had in 2018.
The thing is… inference itself (using the AI) isn’t really all that expensive compared to continuously training the AI. The methodology is out there, the models are out there, the weights are out there, and even if some of the frontier companies jack up their prices and start making uber models for the uber-rich, I think the global market will find other niches and price points to compete in, with more affordable everyday assistants for everyday people. We already see that today with the Gemini and Siris of the world; far less capable than the frontier models that you can pay for today, but free for everyone to use on their phones. Same with the Google Search AI — it’s really pretty crap compared to the paid models, but most people don’t know or care.
So, I think the realistic answer is: Even if you can’t afford $20k/year, you can probably get an “almost as good” service for a lot cheaper, if not from American labs, then Chinese ones. There are some EU upstarts too, including some that have nationalized/socialized government subsidies, I believe (and I don’t just mean in the corruption sense, but rather “AI as a public utility”). (EDIT: Sorry, maybe that was South Korea, not Europe)
Now, all of that is assuming they stay the same basic underlying LLM technology with small improvements every few months, not a sudden, drastic leap into AGI or sentient robotics or whatever.
Worst comes to worst, if ALL AI everywhere suddenly disappears… well, cool. Personally I think the Internet peaked in the 90s and 2000s anyway, and I wouldn’t really mind if all the big tech companies today went bankrupt, AI or not, and we went back to the early hacker era. Despite all the conveniences we have today, I don’t think the world is a better place in any meaningful way compared to 20 years ago. Much worse in many ways.
A lot of home users would not pay that, but it wouldn’t be a problem for corporations. Giving a dozen employees an $8k/yr AI budget would cost less than a single person’s salary.
Also, today’s subscription prices are subsidized on purpose. $20/mo gives you far more than $20 worth of AI compute. If they ever really felt the heat, they could either decrease the usage you get with that (like OpenAI recently did, halving the allowance on their $100 plan), or just move everyone to pay-as-you-go, usage-based pricing. That would more accurately reflect actual costs anyway. “Hey Siri, what’s a good fried chicken recipe?” tends to naturally cost less than “Hey Bixby, hack into the DoD Temple of War for me.”
That’s probably not far from what companies are paying per employee right now. Their bill is going to be much, much higher.
I don’t know if they do. My understanding is that the cheapest flash models now only cost about $0.50 for a million input/output tokens. Prices drop dramatically as hardware scales up and software gets better.
A functioning flash chatbot may only cost a penny per million tokens in a couple of years.
The question is how are AI companies going to charge $2000 a month for a drop in model when a virtually free AI program can do most of the same thing. Also the $2000 a month model today will be the $0.01 per million tokens model in a few years. That’s my impression. The flash programs today are better than the pro models a couple years ago.
I want AI to advance, I just don’t see how private software companies are going to make money on it. The companies that will make money on it are the companies that make the hardware IMO. There’s too much competition and too much depreciation for any one company to corner the market and charge high prices.
I feel like its like GLP-1 meds. They try to recoup the cost of R&D by charging massive prices, so people just buy compounded peptides for a fraction of the price of the official stuff.
shrug I’ll download one of the many open-weight models from HuggingFace and run it locally for free. I won’t have the newest frontier models of course, but I probably don’t need these much as I did when I was less experienced.
Any more questions or is that it?
Which AI product did you use for this calculation?
Yes and no, I think. The costs of using the models do go down with generations, but the costs of training them are still high, not just in dollars but also watts, limited chip capacity, and talent. There’s still a shortage of everything, even as the frontier labs juggle expansions against profitability as they eye IPOs.
And that’s a separate issue from whether the subscriptions are subsidized. They are, because you can compare the same prompts on the same providers and models against API pricing, which tend to be much higher.
But you’re right that the software-only labs have a tough road ahead. As the long tail of “good enough” models keeps getting better, frontier prices become less appealing for the everyday household. Probably there will always be a small percentage willing to pay frontier prices for frontier capabilities, but most people won’t need them. A good AI that can replace a $300k software developer would still find a place in the enterprise, but home users wouldn’t need that.
I don’t imagine most of today’s AI labs surviving very long. We’re just in the gold rush phase right now.
Eh?