Is AI still cheap or just temporarily subsidized?
AI pricing today doesn’t fully reflect real cost.
Some reports and industry estimates suggest that real inference cost can be 5–10x higher than what companies currently charge, depending on model size, context length, and infrastructure load.
So the market is not really stable yet.
It’s still in a growth phase, supported by subsidized pricing and scale.
And usage is already massive.
Most of you probably saw this report:
Meta employees reportedly consumed 73.7 trillion AI tokens in a single month, which is estimated to be around $221M per month (~$2.65B per year) in compute cost.
That’s just one company.
Now multiply this across the industry. Even with “cheap” tokens today, total consumption is already reaching extreme levels. The issue is not only pricing, it’s how fast usage is scaling at enterprise level.
📌So the real questions become:
Will AI become significantly cheaper in the coming years through better hardware, model optimization, and efficiency gains?
Or will we see the opposite — prices increasing once companies stop absorbing the real cost?
Will businesses eventually require paid AI accounts for employees just to control spending and usage quality?
And what happens to companies built entirely on top of APIs — products like Lovable, Cursor, Highsfield that simply layer GPT, Claude, Gemini, and others?
Because their whole model depends on one assumption:
That foundation models stay cheap and widely accessible.
So the real question is simple: when AI pricing stops being artificially low, what breaks first, pricing models, API-based products, or the entire ecosystem built on top of them?
Replies
Really thoughtful breakdown. The API-dependent product risk is real we feel it building auriko.pro. We use AI for curriculum generation and the cost math only works right now because pricing is subsidized. The products that survive long term will be the ones that use AI as an accelerator but don't need it for every single interaction. We've been deliberately building toward that 1M+ questions already generated and cached, so we're not paying per question forever. But you're right that a lot of "AI products" are really just API wrappers with a margin squeeze coming. The ecosystem shakeout will be brutal for pure wrappers.
@auriko_app This is a great point. I think the companies that move fast now and prepare their systems for scale will have a big advantage.
Using AI where it creates the most value, while reducing unnecessary API calls through caching, optimization, and smart workflows, will probably become a key difference.
The biggest risk is not using AI. It’s building a product that depends on every single interaction being powered by AI forever.
@orkhan_ilyas Not saying you're wrong, but I am curious to learn more about that and see numbers. Can you give me links to the reports referred in, "Some reports and industry estimates suggest that real inference cost can be 5–10x higher than what companies currently charge, depending on model size, context length, and infrastructure load."?
@drabiger Honestly, this is not a real research based report, which is why I specifically noted the words “some reports.” I forgot which channel it was, but there was a podcast where an ex Google employee from the AI team talked about this and mentioned this point. Besides that, it was not the first time I saw this. I have seen several sources suggesting that companies are currently not making profits but mostly losing money because they are selling tokens cheaper than the actual cost.
Again, this is not a confirmed fact, just a discussion that came up.
P.s No worries, I am always open if someone asks for a source or proof.
I think AI is both cheap and subsidized.
The cheap part will get cheaper: small models, embeddings, summaries, classification, caching, local models, basic automation.
The expensive part will stay expensive: frontier models, long context, agents, coding workflows, multimodal work, and enterprise-scale usage.
So the whole ecosystem will not break. What breaks first are thin API wrappers with no moat — products that are basically “UI + GPT call.”
The winners will be products that add real value: workflow, data, distribution, trust, integrations, model routing, caching, security, and better UX.
AI pricing will probably become like cloud pricing: seat price + usage limits + enterprise budgets + expensive premium models only when needed.
So the real risk is not that AI becomes unusable.
The real risk is that companies built only on cheap tokens discover they were not building a product — they were reselling someone else’s infrastructure.
@konstantin_gerasimenko Exactly. The companies that survive will be the ones building real value around AI, not just reselling access to models.
running an AI company, the thing that doesn't get talked about enough is that price per token actually has crashed hard the last two years, not because of subsidy alone but real efficiency gains (distillation, better serving infra, competition between labs). so it's not purely "artificially cheap," some of it is genuinely cheaper now than it was. what is still expensive is the stuff that scales badly - long context, agentic loops with lots of tool calls, anything that reruns the same context repeatedly. we route aggressively, cheap model for classification and routing, expensive model only for the step that actually needs it, and that alone cuts spend more than waiting for prices to drop ever would
@galdayan Good point. I agree that price drops are not only because of subsidies. Better infrastructure, distillation, and competition are making AI genuinely cheaper.
The interesting part for me is whether efficiency gains can keep up with how fast usage is growing.
And yes, routing different tasks to different models will probably become a standard approach for keeping costs under control.
@orkhan_ilyas my honest read is no, efficiency won't fully keep up, but it doesn't need to. usage growth is mostly people doing MORE with AI, not the same task getting more expensive. so the per-task cost curve bending down and the total-usage curve going up aren't really racing each other, they're just two separate things happening at once. the actual squeeze will be on whoever built their product assuming today's token price is permanent