| |
| ▲ | trollbridge 6 hours ago | parent | next [-] | | There isn't much evidence at all that inference is "subsidised" (and by whom?) Training is quite expensive and it does look likely that the American providers have been doing that at a loss. In any case, you can go buy a MacBook Pro M5 48GB or an AMD R9700 and run Qwen 3.6 35B-A3B (a very capable model) and the only "subsidy" is you plugging it in, and 140W is not exactly a huge amount of power (roughly 50¢ per day if you run it 24/7 at 100% load, which it is very unlikely you will). | | |
| ▲ | kergonath 6 hours ago | parent | next [-] | | > There isn't much evidence at all that inference is "subsidised" (and by whom?) None of the big providers are profitable. It’s subsidised by overly enthusiastic VCs. > In any case, you can go buy a MacBook Pro M5 48GB or an AMD R9700 and run Qwen 3.6 35B-A3B (a very capable model) and the only "subsidy" is you plugging it in Right, people could. But they won’t, because that’s a bloody expensive computer and they don’t need that to ask ChatGPT. That war is lost already. Subscription to the big players’ services would need to increase massively for that to happen. And the computational cost is only part of the problem; these models also eat a lot of storage and RAM, which is not exactly getting cheaper. | | |
| ▲ | trollbridge 6 hours ago | parent [-] | | The typical "free" AI or cheap tiers are equivalent in power to a Qwen 3.6 model (which is also much cheaper to run in a hyperscaled situation than on my laptop or PC; a single H200 can host thousands of sessions of a typical chatbot user). There is no evidence the Chinese AI providers are being subsidised either. You can look at API pricing on a service like OpenRouter (which isn't subsidised) and see pretty readily that it's not expensive to provide lower-tier inference. Higher-tier inference like GPT-5.6-Sol or Opus is expensive - $100 a month plan for realistic usage, and only up from there. |
| |
| ▲ | amarcheschi 6 hours ago | parent | prev | next [-] | | I agree with using smaller models, it's just that the majority of people I know feel like they need the biggest, beefier, behemoth model possible (with the longest thought setting) and consume much more than necessary when a flash or smaller model would be OK. I would also like to be able to use a smaller model, but given ram prices I would have to sell a kidney to buy ram now | | |
| ▲ | trollbridge 6 hours ago | parent [-] | | Most people who use a free or $20 a month plan are already using smaller models, and the mainstream chatbot services will route requests to a smaller model often without really telling the end user. You can run Qwen-3.6 on a 32GB card which will set you back about $1400, or $400 of just RAM if you want to run it on a CPU. |
| |
| ▲ | mschuster91 6 hours ago | parent | prev [-] | | > There isn't much evidence at all that inference is "subsidised" (and by whom?) Well... why else would the major providers now tighten the screws on per-token pricing? | | |
| ▲ | blfr 6 hours ago | parent [-] | | Because they thought they could. Turns out the Chinese and Elon had other plans. | | |
| ▲ | mschuster91 6 hours ago | parent [-] | | The Chinese providers are just as much getting subsidies from the CCP, and Musk/SpaceX is (indirectly) raiding retirement funds to fund the bonanza. | | |
| ▲ | trollbridge 6 hours ago | parent | next [-] | | There is no evidence Chinese providers are getting such subsidised, and in fact apparently Jinping (who presumably knows what the CPP is doing) was surprised when DeepSeek and Qwen generated so much buzz. In China, AI inference is just viewed as another basic utility, much like an e-mail provider or a mobile phone network. I'm not aware of how Musk/SpaceX are "raiding retirement funds"? | | |
| ▲ | mschuster91 5 hours ago | parent [-] | | > I'm not aware of how Musk/SpaceX are "raiding retirement funds"? NASDAQ bent their rules to allow SpaceX a (way too early) inclusion into the index and so did MSCI [2] and Russell [3]. Normally, a newly IPO'd stock would have required up to a year of "cooldown" (like the S&P 500 requires) so that stock prices can stabilize. Now though? Billions of dollars in funds are automatically flowing in from retirement accounts into SpaceX and artificially prop up the valuation of this grossly overvalued company. And OpenAI and Anthropic are looking to IPO as soon as possible as well to benefit from the same rules while the markets are still red-hot bullish for anything that can be labeled even remotely related to AI. Assuming that there will be a catastrophic collapse event in the AI bubble - the triggers can be anything from regulatory issues (no matter if in the US, EU or China), new free models from China cutting off the moat of the Big Three, venture capital running out and forcing realistic pricing or a natural disaster/war wiping out TSMC or RAM factories, interrupting supply for the continued outbuild -, this will directly (and massively) impact retirement accounts. In addition, even the sell-offs required in ETF rebalancing can have serious economic consequences. Something has to give when SpaceX, OAI and Anthropic all enter. [1] https://finance.yahoo.com/markets/stocks/articles/nasdaq-che... [2] https://www.reuters.com/business/media-telecom/msci-confirms... [3] https://www.reuters.com/business/media-telecom/russell-rebal... |
| |
| ▲ | blfr 6 hours ago | parent | prev [-] | | The truly cheap Chinese models are usually the open weights ones so while there may be a training subsidy, the inference prices reflect real costs. |
|
|
|
| |
| ▲ | wickedsight 6 hours ago | parent | prev | next [-] | | I get what you're saying, but I'm guessing that people asking how to pee is a drop in the bucket compared to the agentic loops being called to rename some variables across a project. | | |
| ▲ | runarberg 5 hours ago | parent [-] | | I don’t use AI so this comes across to me as a bit of a culture shock, but is asking AI to rename a variable across project really something AI users are wasting their tokens on? In emacs I can do that with `S-l r r` or `M-x lsp-rename`. Using AI to do this seems extremely inefficient and wasteful, not to mention improper and unprofessional, and that is looking past the moral implication of training on stolen code and polluting our climate. |
| |
| ▲ | simianwords 7 hours ago | parent | prev [-] | | The fact is, you wouldn’t be okay with the real prices as well. All indications point to opus not being subsidised but having huge margins. GLM 5.2 is not subsidised - it is an Opus tier model that costs a small fraction. I doubt you would be okay and all problems would suddenly vanish. | | |
| ▲ | ekidd 6 hours ago | parent [-] | | GLM 5.2 isn't quite modern Opus tier, as seen in this comparison where Opus 4.5 scores 4/5 on some coding tasks where GLM 5.2 scores 0/5: https://www.tryai.dev/blog/gpt-5.6-build-off-12-models But yes, GLM 5.2 is cheap. But the real standout on price is DeepSeek V4 Flash, which competes, more or less, with models in between Sonnet and Haiku. From third-party providers, it costs around $0.09/M, $0.18/M out, compared to $3M/in, $15M/out for Sonnet and $1M/in, $5M/out for Haiku. To get the price of DSv4 (Flash and Pro) so low, DeepSeek did a lot of innovative optimization work that will likely show up in other open weight models in the future. | | |
|
|
