Hacker News Reader

Comment by Kuyawa | Hacker News Reader

AI investment will crash but AI itself (the technology) will continue thriving, learning, improving and there is absolutely no way to stop it. The only reading on the crystal ball is if US companies fail, China will take the lead by leaps and bounds, so the only solution is to keep pushing the cart until the wheels come off or we all cross the finish line, together.

View on HN

> absolutely no way to stop it.

Unless if they take down the flat rate subs and you have to pay api prices, usage will drop a lot. Right now I freelance using gpt, if I had to pay api prices I'd probably lose 50% of the income, if not more, with the ~40% tax on top I might as well spend my time doing something else

I don't necessarily buy that API prices represent "real" prices of the models.

Well of course they're higher than marginal cost. But these providers have to also generate a fat return on investment.

What do you think the real price is? Subscriptions are heavily subsidised, I don't know anyone who would deny that.

The general model for subscriptions is that power users are subsidized by subscriptions of casual users, like a gym membership or whatever.

This is a little dicier in post-agent AI, because it's easier for casual users to automate power-user consumption, but the providers have done decently in discouraging that.

There's people here saying they're obviously subsidized, there's people here saying they're obviously profitable. I think they're probably subsidized, but I would hold back on saying it's obvious.

It's probably a safe assumption that openrouter prices for Kimi K3 are "real".

Not in the slightest as there are providers selling K3 for half of what OR has them listed for. (And maintaining profitability)

If they're hosted in China next to a coal power plant and maintained by people getting paid 1/10th of a US engineer sure.

I doubt opanai and anthropic will go that road

I moved one of our daily workflows over to Kimi K3 on Fireworks. It replaces two sales assistant positions, and was about $50/day for 14 million tokens total (in/out). They surely are not subsidizing this price as it is just inference only and other providers are even less money.

Yes, they are ludicrously profitable - even when future training costs are taken into account!

[citation needed]

That would just create incentive to build better consumer HW for larger open weight models. And that would make me very happy. But I think the poster in another thread who stated there's a netflix subscription phenomena going on with the fixed rate pricing was onto something. My agents run 24/7 until I hit all my limits. I suspect my results are not typical. And as of yesterday I have issues with both major CEOs yet I thank them both for subsidizing my tokenmaxxing.

You're unlikely to ever be priced out of tokens, at least if you'd be willing to settle for a model closer to Sonnet 4.5. That level of model certainly isn't as efficient as Fable 5, but it can crank out CRUD apps and other consulting mainstays quite well, with some supervision.

To give you an example of model in this class, the DeepSeek V4 Flash preview is a 284B A13B model, with a native quantization mixing 4 bit and 8-bit values. You can easily run it on an RTX Pro 6000 Blackwell (or 2) at a reasonable quant, especially if you offload the MoE weights to 48-64GB of system RAM. This costs US$11,800 at Microcenter right now, and it will work in any gaming box with decent cooling and a modern 1000W power supply. Over the lifetime of the card, an entire system would cost you under $4,000/year. Power is about 300W for the card (either a blower model, or a workstation model with the power cap), and another 150W or so for the rest of the server.

Or you could buy it on Open Router from dozens of different commodity vendors, starting around $0.09 per million tokens input, $0.18 per million tokens output. This is a competitive market price, so some of the providers might be losing money or reselling surplus capacity. But given the underlying hardware costs, the numbers are in the ballpark. In other words, if you're willing to settle for lower-quality tokens, you can get roughly Sonnet 4.5 for close to free, or as a modest capital expense for a successful freelancer. Halfway decent tokens are cheap, and you can generate them in-house!

The problem is I cannot imagine ever making a single dollar with AI. I have built about 16 SaaS apps with AI, in the time it took to build by first old fashioned human-created SaaS app. The one thing the agentic coded apps have in common with my old one is that none of them have ever won a single paying customer.

> You can easily run it on an RTX Pro 6000

Sure, and people can just build their own dropbox for like $250 too.

How many people will bother though

Well, since you "freelance using GPT" and you expressed concern that you'd stop being able to make a profit if you had to pay API prices, I figured that swapping an off-the-shelf graphics card into a gaming rig might be a reasonable way to stay in the black. If it came to that. And the software setup on Linux is just compiling and installing llama-server, which shouldn't be enough to stop any programmer trying to make a living.

Or you could just take your credit card and spend $20 on credits at https://openrouter.ai/deepseek/deepseek-v4-flash. I'm not sure that I could manage to spend even a $1/day at those rates.

My larger point is that while frontier tokens are a near-monopoly and who knows what they really cost, many real-world workflows can be run using commodity tokens, or even served in-house by anyone who can afford to hire US or EU programmers. And if you're willing to settle for what would have been a state-of-the-art coding model in October 2025, commodity tokens are close to free.

Software brained developers can't comprehend the idea of capital expenditures required to operate a business. Too spoiled by cheap laptops, free software and coffee shop wifi.

It can't 'learn' on its own. Models only get better with mountains of RnD for data, training, and lots of fine tuning.

So the moment investment dries up, models stop improving.

However, it's likely we'll get good 80/20 solutions where you get most of the performance of the then-unsustainable high end models for significantly less compute.

>It can't 'learn' on its own.

Nearly there today. https://www.anthropic.com/institute/recursive-self-improveme...

Do we really believe marketing copy from Anthropic re: RSI considering their recent model releases seem to be markedly regressing in real world performance vs. past models?

and lots and lots and lots of cash

There is "absolutely no way to stop" the thing that requires hundreds of billions of dollars of cash injected every year just to avoid falling over?

it's a bit unfair to say this is the only thing that can happen. we can just as easily assume in like over a few years, US leadership identifies that AI is an arms race, and continuing down this path leads is going to lead to a zero-sum-game. and so instead of just mindlessly pushing the cart forever, we see government intervention

However, who will pay for the very expensive training if the ROI isn't there? Absent government subsidization, then likely the frontier models will be sold off to the hyperscalers by the bankruptcy administrators to wring a few more dollars out of them as they become out of date and more and more irrelevant - assuming that inference alone is profitable at a price the market is willing to pay of course, if not then yes the lights will indeed be turned off.

Just like every AI cycle that came before. At every cycle, we get a new tool. But we also get an extended "AI Winter" where investment dries up. This has been happening since the 1980s, we just keep redefining what "AI" is. At every cycle we move the line a bit; at one point voice recognition was firmly in the AI camp(and before that, science fiction). Now we have a machine that can do that in our pockets and nobody cares.

Does anyone remember expert systems?

Douglas Hofstadter talked about these moving goalposts for AI in his book Gödel, Escher, Bach. I remember when a program that could win at chess was considered impossibly strong AI. Then, the game go. Etc.

Heinlein in his book The Moon Is A Harsh Mistress imagined that people in the future would find it worthwhile, even necessary, to learn an artifical language LogLan so they could talk with computers. LogLan (Logical Language) would be designed to be totally unambiguous, a necessity for speaking with a computer. He didn't imagine that we'd simply throw so much compute at the problem that we'd teach computers to understand our vague, mumbled natural languages in all their diversity.

Ah, Lojban!

https://en.wikipedia.org/wiki/Lojban

Sadly this does not let you speak to computers, just other language nerds with a lot of time.

a friend of mine did a high school project where he and a friend learnt it well enough to have a short conversation with each other. That's the only time it has come up in life until now. I hope he sees this.

I didn't know there is a finish line.

Yup. If there is a crash, those with cash and understanding will clean up. AI is eating the world, period.

what exactly is this finish line? AGI?

That's what they'll sell you. But fundamentally current LLM implementations seem to be the wrong methodology for 'AGI' and I don't think we're remotely close to having this.

Yes.

Investors for sure don't care about people. Replace Accenture with OpenAI and you replace 800k people immediadly.

Geoffrey Hinton said it in his Videos: a normal person costs A LOT of money to train, educate, onboard, keep etc. they forget things etc.

AGI you train once, teach once then clone and copy and just run it.

We even might already crossed the line were it is already more cost efective to train 1-5 frontier models something it doesn't know yet than teaching this to a 1000 humans.

Economists have a term of art for this phenomenon: "automation".

What exactly was the finish line for the Internet?

> AI investment will crash but AI itself (the technology) will continue thriving, learning, improving and there is absolutely no way to stop it.

No, there's is a way to stop it. As obvious counterexample: go all Butlerian Jihad: death sentence for anyone caught "making a machine in the likeness of a human mind." It's doable with a few laws and treaties. I'm not saying that's the only way it could be stopped, it's just a way it could.

I'm fucking sick of TINA, especially with arguments from vigorous assertion.

HN consistently, IMO, underrates the degree to which technology adoption is “political” rather than inevitable. I was going to write a long jokey bit here but instead; counter examples - nuclear power, eugenics, the AC vs DC wars during electrification, the great firewall of China…

It's techno determinism which assumes a libertarian view of markets where politics and culture don't matter.

Reply on news.ycombinator.com