Recently I watched a startup demo where they were running 500 GPT-4 queries per user session. In 2023, that would have cost them $5 per user. Today it costs them less than 50 cents. And they’re still complaining about API costs being too high.
When costs collapse, new worlds appear. This isn’t just cost-cutting. It’s a phase change.
Think about what happens when water turns to steam. At 99°C, you have hot water. At 100°C, you have an entirely different substance with different properties that can do different things.
Technology costs work the same way. There are threshold points where quantitative changes in price lead to qualitative changes in what’s possible.
AI cost inversion
A year ago, the economics of AI were clear: Features had to deliver at least $50 of value per dollar of compute cost to be worth building. That limited what companies decided to do.
Now that rule is irrelevant. In the last year, OpenAI’s token costs fell 90%. Open-source models like Llama, Mistral, and DeepSeek can run locally. Small companies can fine-tune models for less than hiring a developer.
The interesting thing isn’t just that AI got cheaper—it’s that the pattern of costs inverted.
In the old world, the decision looked like: “Can we afford to use AI here?” In the new world, it’s: “Can we afford not to use AI here?”
I know startups who literally spent their first $500K just on OpenAI credits last year. The same startups now spend $50K and get better results. That’s not an incremental improvement. It’s a different game.
What actually changed in AI
Three technical innovations caused this collapse.
First, we learned that smaller models can be almost as good as big ones if you train them right. DeepSeek proved this with their 8 billion parameter distilled model (DeepSeek-R1-Distill-Llama-8B), which scores 89.1% on MATH-500, surpassing GPT-4o (estimated at over 1 trillion parameters) at 88.6% in mathematical reasoning. Size, it turns out, isn’t everything when efficient training takes the stage.
Second, companies figured out how to make models generate their own training data. Instead of paying thousands of human labelers, they use existing models to create examples. Companies have shown that this cuts data costs by 90%. Their models now improve themselves in a continuous loop.
Third, inference got way more efficient. In 2023, running GPT-4 required high-end Nvidia GPUs. Now, optimized models can run on laptops.
What this means is that AI costs have shifted from prohibitive to negligible.
What the AI cost collapse means
The implications are massive, but not obvious.
Access has been democratized. Small teams can now build products that only tech giants could afford last year. A college student in Bangalore can build and deploy a specialized financial analysis model for less than the cost of their textbooks.
Integration beats raw capability. When everyone has access to good-enough models, what matters is how you integrate them into workflows people actually use. Understanding a specific domain matters more than having the biggest model.
The experimentation is just beginning. When costs drop, people try weird things. Most fail, but some work in surprising ways. I’m seeing startups try AI approaches that would have been economically absurd six months ago. Some are finding that what seemed wasteful at $1 per query is transformative at 1 cent.
The competitive landscape has reset. Companies that invested millions in proprietary AI infrastructure last year are watching their advantage evaporate as similar capabilities become available as API calls or open-source models. Meanwhile, new startups are designing for the new economics from the ground up.
At Upekkha, we’re seeing this: Our current cohort is reaching product-market fit with 60% less capital than last year’s cohort. They’re not building less; they’re building more efficiently because of the cost collapse.
DeepSeek R1 open-source breakthrough
The cost collapse accelerated dramatically with DeepSeek’s release of its R1 model. This model shifted the economics of AI deployment in ways few anticipated. DeepSeek R1 delivered performance comparable to GPT-4 at just 15% of the operating cost, while being available both as an API and for local deployment.
What made DeepSeek R1 revolutionary wasn’t just its price point, but its novel architecture that enabled efficiency at scale. Companies that previously spent millions on model customization could now achieve comparable results for tens of thousands. One enterprise customer reported reducing their AI infrastructure costs by more than 80% after switching to DeepSeek R1, while simultaneously improving response quality for domain-specific tasks.
The model’s ability to run efficiently on consumer-grade hardware further democratized access. Small teams can now deploy capabilities that previously required specialized infrastructure and deep expertise.
Relationship between open source and cost
There’s a strange dynamic in the AI world right now. Open source is winning in some areas but losing in others. And it’s not happening randomly.
Most think of open source as a philosophy, but tech companies use it as a strategy. Big tech companies aren’t purely open or closed; they’re strategically both. They open source what commoditizes their competitors’ advantages and keep proprietary what differentiates them.
Meta open-sourced Llama to commoditize OpenAI’s advantage. OpenAI keeps its training methods proprietary to maintain its edge. Neither approach is “right,” they’re just different strategies.
What’s interesting is how this plays out economically. As one founder told me recently, “Production costs are crashing while distribution stays just as hard to crack.”
When models get cheap, the game shifts. The hard part isn’t building the model anymore; it’s getting it to users. And that means distribution becomes even more valuable than before.
Changes in how information is consumed and processed
This is changing how people consume information. As one AI researcher explained to me: “For 50 years, we assumed humans read documentation directly. We broke content into sections, added screenshots, and built navigation—all for humans. That’s history now. Today, AI early adopters interact with documentation through AI. They don’t read docs; they ask questions and AI reads docs for them.”
This changes everything about how we design products and share information. Documentation isn’t for humans anymore—it’s for AI to interpret for humans. User interfaces become conversations. Support becomes embedded, not separate.
What we’re witnessing isn’t just cheaper AI. It’s a complete rewiring of how people interact with technology and information.
Challenges of AI abundance
Abundance in AI creates new problems.
Choice paralysis is real. When there were three good models, decisions were simple. Now there are hundreds. I’m seeing teams waste weeks evaluating models instead of building products.
Quality varies wildly. Not all cheap models are good models. We’re seeing companies deploy cost-optimized systems that give incorrect information or make poor decisions, only to discover the hidden costs far exceed what they saved on compute.
The race to differentiate intensifies. When everyone can access good AI, having AI isn’t special anymore. The value moves to domain knowledge, data advantages, distribution, and user experience.
Expectation inflation happens fast. Users quickly adapt to what’s possible. Features that delighted people six months ago are now considered basic. One startup I advise released an AI feature to rave reviews in January. By March, users were complaining it wasn’t keeping up with competitors.
What’s next in AI opportunities
The cost collapse isn’t over. We’re still early in this transition. Early patterns are arising among the companies that are thriving.
They assume AI is abundant, not scarce. They design with the assumption that they can use AI everywhere, not just at key moments.
They build for the economic reality that’s coming, not the one that exists today. Even as costs fall further, they’re positioned to take advantage.
They focus on the problems AI still can’t solve. Valuable parts of their stack aren’t the AI components but the things around the AI—the data they’ve accumulated that model companies can’t have, the workflows they’ve designed, the user experiences they’ve crafted.
They combine AI with domain expertise. The biggest opportunities aren’t in general AI tools but in applying AI to specific domains where the founders deeply understand the problems.
The world doesn’t change gradually. It changes in jumps, when some resource crosses a threshold and becomes abundant.
That’s what’s happening with AI right now. The cost threshold has been crossed. Now we’re seeing what’s possible on the other side.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.
This story was originally featured on Fortune.com
