Recently, I watched a startup demonstration where they started 500 GPT-4 inquiries in the user session. 2023. It would cost them $ 5 per user. Today it costs less than 50 cents. And they still complain that the costs of the API are too high.
When the costs collapse, new worlds appear. This is not just a reduction in costs. It’s a phase change.
Think about what happens when the water turns into a pair. You have hot water at 99 ° C. At 100 ° C you have a completely different substance with different properties that can do different things.
Technology costs operate the same way. There are points of threshold where quantitative changes in the price lead to qualitative changes in what is possible.
AI inversion of costs
A year ago, the AI economy was clear: the features had to deliver at least $ 50 values per dollar of computer costs to be built. This limited what the companies decided to do.
Now that rule is irrelevant. Last year, the costs of the token Openi fell 90%. Open -code models like Llala, Mistral and Deepseek can run locally. Small companies can finely adjust models for less than the recruitment of developers.
An interesting thing is not just that the AI has become cheaper – it is sample the costs of the reverse.
In the old world, the decision looked, “Can we afford to use AI here?” In the new world it is: “We can afford not use AI here? “
I know the startups that literally spent the first $ 500,000 only on the opening loans last year. The same startups now spend $ 50,000 and get better results. This is not an incremental improvement. It’s a different game.
What actually changed in AI
Three technical innovations caused this collapse.
We first learned that smaller models can be almost as good as those if you train them properly. Deepseek proved this with his 8 billion distilled models of parameters (Deepseek-R1-Destill-Jelma-8b), which reaches 89.1% at the Math-500, surpassing the GPT-4o (estimated at more than 1 trillion parameter) to 88.6% in mathematical reasoning. Size, it turns out, it’s not all the efficient training sessions on stage.
Second, companies realized how to generate models their own training information. Instead of paying thousands of human labels, they use existing models to create examples. Companies have shown that this reduces data costs by 90%. Their models are now improved in a continuous loop.
Third, the conclusion became more effective. In 2023, running GPT-4 demanded a top-notch end Nvidia GPU. Now optimized models can be started on laptops.
This means that the costs of the AI switched from the prohibition for negligible.
What AI consumes collapse
Implications are massive, but they are not obvious.
The approach is democratized. Little teams can now build products that only technological giants could afford last year. A student in Bangalore can build and implement a specialized model of financial analysis for less than the cost of their textbooks.
Integration wins raw ability. When everyone has access to a good enough model, it is important to integrate them into the work of work that people actually use. Understanding a particular domain is significantly more than the largest model.
Experimentation is just beginning. When costs fall, people try strange things. Most fail, but some act in surprising ways. I see the startups try out AI approaches that would be economically absurd six months ago. Some reveal that what seemed wasteful in the amount of US $ 1 upon inquiry transformed to 1 percent.
Competitive landscape is reset. The companies that have invested millions in the ownership of AI infrastructure last year see their advantage to evaporate that similar options become available as API calls or open code models. Meanwhile, new startups are designing for a new economy from the foundation.
In Zankkha, we see this: our current group reaches a product market with 60% less capital than last year’s cohort. They do not build less; They build more effectively due to cost collapse.
Deepseek R1 Broken Open Code
The cost was dramatically accelerated with Deepsek’s exit his R1 model. This model has moved the economy of arranging AI in a way that is little expected. Deepseek R1 delivered the performance comparable to the GPT-4 with only 15% of operational costs, while it was available as APIs and for local implementation.
What did Deepseek R1 revolutionary did is not only its price, but also its new architecture that enabled the effectiveness to be in proportion. Companies that have previously consumed millions on the model adjustment could now achieve comparable results for tens of thousands. One company customer reported to reduce the cost of AI infrastructure by more than 80% after the Deepseek R1 switch, at the same time improveing the quality of response to domain -specific tasks.
The ability of the model to effectively run on a consumer hardware has further democratized access. Little teams can now implement the possibilities that have previously required specialized infrastructure and deep expertise.
A relationship between open code and cost
There is a strange dynamics in the world of AI. Open source wins in some areas, but loses in others. And that doesn’t happen randomly.
Most consider open code as philosophy, but technological companies use it as a strategy. Large technological companies are not exclusively open or closed; They are strategically both. They are open to comfort the benefits of their competitors and keep the ownership of which differentiates them.
Target Open Lama to comfort the benefits of OPENAI. Openai maintains its training methods owned to maintain its advantage. No approach is “right”, they are just different strategies.
What is interesting is how economically it is played. As one founder recently told me, “production costs are crashed while the distribution remains equally difficult to crack.”
When the models become cheap, the game moves. The hard part no longer builds a model; Gets it to users. And that means that distribution becomes even more valuable than before.
Changes in how information is consumed and processed
This changes the way people consume information. As a researcher AI explained to me: “For 50 years we assume that people have read the documentation directly. We broke into compartments, added the screenshots and built navigation – all for people. This is now history. Today, and early adopters communicate with documentation through AI; they do not read documents; they ask questions and read documents.”
This changes everything about how we design products and share information. Documentation is no longer for people – this is for AI who interprets for people. User interfaces become conversations. Support becomes built -in, not separate.
What we witness is not just cheaper AI. It is completely reviving the way people communicate with technology and information.
Challenges of abundance of AI
Abundance in AI creates new problems.
Paralysis of choice is real. When there were three good models, the decisions were simple. There are hundreds now. I see teams that consume weeks that evaluate models instead of building products.
The quality of the wild is varied. Not all cheap models are good models. We see that companies arrange costly optimized systems that give incorrect information or make bad decisions, only to discover hidden expenses far exceeds what they have saved on calculating.
The distinction race is intensified. When everyone can approach a good AI, to have Ai is no longer special. The value is aware of the domain, the benefits of data, distribution and user experience.
Expectation inflation happens quickly. Users are quickly adjusted to what is possible. Features that delighted people have been considered basic six months ago. One startup I advise that in January he has published an Ai feature for anger of reviews. By March, users complained that this was not accompanied by competitors.
What follows in AI capabilities
Cost crash is not over. We are still early in this transition. Early patterns arise among the companies they succeed.
They assume that Ai is abundant, not scarce. They design by assuming that they can use AI everywhere, not just in key moments.
They build for the economic reality that comes, not the one that exists today. Even when the costs fall further, they are positioned to take advantage.
They are focused on problems that AI still cannot solve. Valuable parts of their bundle are not AI components, but things about AI – data that accumulated that companies that could not have, the flow of work they designed, the user experience they made.
Combine AI with domain expertise. The biggest options are not generally AI tools, but in the application of AI to certain domains in which the founders deeply understand the problems.
The world does not change gradually. This changes in jumps, when a resource exceeds the threshold and becomes abundant.
This is currently happening with AI. The cost threshold has exceeded. Now we see what is possible on the other side.
Opinions expressed in fortune.com Comments are just the views of their authors and do not necessarily reflect opinions and beliefs Wealth.
This story is originally shown on Fortune.com
Source link