Revenue growth has been dramatic. OpenAI generated $3.7 billion in revenue in 2024 before jumping to $13.07 billion in 2025. By the end of that year, monthly...
The difference being they had a plan to monetize YouTube which offset the operating costs and even could make a profit, in addition to the stuff they over invested in having long term benefits. Aka they roughly could spent a lot to save money later off of what they built.
LLMs are not the same. They cost a shitload of money to run. Actual token based usage without being subsidized by investors would make any LLM cost users so much money that the actual value of it would immediately become major problem. Sure one can, currently, get decent code output by using hundreds or thousands of tokens, or using multiple LLMs / loops, or having agents go burn however many on iteration! But if we paid actual costs this shit would rapidly be shut down.
The AI infrastructure is also not saving money long term. Training is unbelievably expensive. Compute costs are all about gigantic data servers with video cards running, and the economics of all that is way tighter than anyone gives credit for. Those cards last like 3 years. The cooling costs are crazy. You have to have constant use to be efficient. None of these things are able to be covered by AI economics nor does it even make sense to be.
It costs too much, it’s just being covered by your money being pissed away by tech investors for a technology that cannot survive.
In theory costs could come down with each new hardware generation if the we dont keep pushing models the to max extent of what the hardware can do while pushing size.
E.g Claude Opus today, only trained in a similar size and manner as today, will be cheaper to run on whatever the next GPU that comes out with higher speeds and processing capabilities, unless of course NVidia raises the cost substantially. Given the current situation I think nvidia might do that which would hamper this lowering of costs, but it should possible, if not slower.
E.g 10 years from now it will be cheaper to run a opus similar model. But 10 years from now everyone will want the mythos of today, then. That wont be cheaper.
Sure one can, currently, get decent code output by using hundreds or thousands of tokens, or using multiple LLMs / loops, or having agents go burn however many on iteration!
This always stumps me. Because if that was true, Anthropic products (API, Cache, Claude Cowork, Claude Code) would not be shitty.
They are shitty. They are coded shittily, and Anthropic is unable to solve some bugs for years now. E.g. there is the console flickering bug that they tried to fix 3 times and rollbacked or failed all of them.
The difference being they had a plan to monetize YouTube which offset the operating costs and even could make a profit, in addition to the stuff they over invested in having long term benefits. Aka they roughly could spent a lot to save money later off of what they built.
LLMs are not the same. They cost a shitload of money to run. Actual token based usage without being subsidized by investors would make any LLM cost users so much money that the actual value of it would immediately become major problem. Sure one can, currently, get decent code output by using hundreds or thousands of tokens, or using multiple LLMs / loops, or having agents go burn however many on iteration! But if we paid actual costs this shit would rapidly be shut down.
The AI infrastructure is also not saving money long term. Training is unbelievably expensive. Compute costs are all about gigantic data servers with video cards running, and the economics of all that is way tighter than anyone gives credit for. Those cards last like 3 years. The cooling costs are crazy. You have to have constant use to be efficient. None of these things are able to be covered by AI economics nor does it even make sense to be.
It costs too much, it’s just being covered by your money being pissed away by tech investors for a technology that cannot survive.
In theory costs could come down with each new hardware generation if the we dont keep pushing models the to max extent of what the hardware can do while pushing size.
E.g Claude Opus today, only trained in a similar size and manner as today, will be cheaper to run on whatever the next GPU that comes out with higher speeds and processing capabilities, unless of course NVidia raises the cost substantially. Given the current situation I think nvidia might do that which would hamper this lowering of costs, but it should possible, if not slower.
E.g 10 years from now it will be cheaper to run a opus similar model. But 10 years from now everyone will want the mythos of today, then. That wont be cheaper.
This always stumps me. Because if that was true, Anthropic products (API, Cache, Claude Cowork, Claude Code) would not be shitty.
They are shitty. They are coded shittily, and Anthropic is unable to solve some bugs for years now. E.g. there is the console flickering bug that they tried to fix 3 times and rollbacked or failed all of them.
Or maybe we define decent differently.