The Asymmetries of a Great Disruption
While every single technological revolution has created more jobs than it has destroyed, AI will be no exception. However, we are in for a very complex transition to that world.
Simply put, the way AI will develop in the next few years is at real risk of becoming a ‘pay to win’ game, where those with access to the best tools will have an insurmountable advantage over those who do not.
Seems hard to fathom at this point, so let’s drive home this fear that has been growing in me over the last weeks.
I make AI simple, less sexy (no hype), but much more powerful for you. Newsletter written by an AI expert — or at least pretends to be one — with 170k followers and hundreds of thousands of monthly readers with independent thoughts.
Click below and start today for free.
The Cost of Intelligence
While AI has been making steady progress for many years, ChatGPT entered all of us suddenly into a new world that has only recently started taking its real form.
LRM, The Crucial Three-Letter Acronym
When ChatGPT was released, it introduced most of society to a term that, albeit already existing, was largely unknown. This word was LLM, which stands for Large Language Model, an AI model that would take in a sequence of words and output a continuation of the input sequence.
While using an LLM already places a certain degree of advantage for the user against those who don’t use these tools, perfectly exemplified by college students using these LLMs to improve their homework, access to LLMs rapidly became more a matter of being quick on your feet in adopting these tools, as they were always quite accessible even for modest budgets.
Additionally, their usefulness was often exaggerated. They are helpful in speeding up some tasks, but they aren’t tools that define your superiority against non-AI users.
Moreover, their costs have since plummeted several orders of magnitude to the point the more tech-savvy can run powerful models on their computers for free.
Put another way, LLMs have mostly been a ‘be curious to win’ rather than a rich people’s game.But I fear 2025 will mark the moment that curiosity isn’t enough to benefit from these tools. Instead, you are going to need money.
A lot of it.
And the reason behind this is their next evolution, the LRM, an acronym that, unbeknownst to most, changes the game. And the rules.
When Intelligence Became a Luxury
I’m purposefully focusing on text models, as most LRMs (if not all at this point) don’t present multimodal capabilities (processing and specifically outputting anything but text). These are also the most impactful to society, so my points below still stand nonetheless.
Large Reasoner Models (LRMs) play a very similar role on the surface when compared to LLMs; they take in a text input sequence, and they return you one back.
The crucial difference lies in the process. An LLM’s process of generating an answer can be compared to human intuition; it’s fast, automatic, and almost “instinctive,” ideal for the types of tasks where humans perform unconsciousness or non-reasoned cognitive tasks like calling the name of your best friend, singing your favorite song, or simply walking; we do not engage our conscious minds for the task.
This also explains why they have been generally terrible at tasks requiring reasoning, like maths. These types of tasks generally benefit from two crucial aspects:
- Exploration: Humans, upon facing a hard task, will explore a manifold of options to solve the problem.
- Time: The longer we think about the task, the larger the chances we get the correct answer.
These two new aspects imply a really simple-to-understand consequence: the LLM needs to ‘think’ for longer.
Whether you believe AIs think or ‘imitate thinking’ is a mostly unsolved question whose truth currently lies in the eyes of the beholder. But for the sake of the article, the output is a thought response either way.
Consequently, LRMs embody this vision implanted into LLMs. While exploration is largely a work in progress and, crucially, mostly vacant outside the research world, the main attribute that defines LRMs today is that, simply put, they are LLMs that ‘think for longer.’
This thinking for longer on a task usually takes the form of a chain of thought, where the LLM generates a concatenation of linked thoughts, similar to a human taking a multi-step approach to solving a hard problem.
While this has naturally increased the ‘intelligence’ of AI models, the brute force impact that underpins this whole article is that, more importantly, it changes the access to intelligence.
A “small” change that makes intelligence a luxury.
But what do we mean by that?
Don’t Blame the Player, Blame the Rules
In ‘AInomics,’ the unit economics of AI, all boils down to one question: what is the cost of a token?
AI Economics 101
AIs can’t process language, only numbers. Thus, every AI model must perform two actions: encoding and decoding.
- Encoding transforms words (or any other data type) into a vector of numbers that captures the attributes of the word.
- Once these encodings, known as embeddings, are processed (“understood”) by the AI model, it decodes them back into language
Despite honorable efforts by companies like Meta with the Byte-level Transformer I recently covered, most AI models break language into tokens, words or subwords based on their commonality in {insert language of choice}.
Processing the embeddings and generating new ones to form a response requires a huge amount of computation executed by hardware known as a GPU (and others like LPUs, TPUs, or RDUs) that are:
- Absurdly expensive, fueled by an extreme supply/demand asymmetry
- Faulty, they have far-from-ideal life-span and, due to several factors, fail a whole lot.
As the computation amount is very large, you will need many of them, dramatically increasing your total cost of ownership.
Worse off, they require indecent amounts of energy and vast surfaces of land in a collocated fashion (they have to be placed closely one to the other), leading to ungodly amounts of capital expenditures (CAPEX) that, if the trend continues, will comfortably surpass the $320 billion mark in 2025 just considering four companies: Microsoft, Meta, Google, and Amazon (there are many other companies, too).
A rough way to calculate total CAPEX is to take the AI business revenues of NVIDIA, AMD, Broadcom, and Huawei and double them, as the GPUs themselves usually account for 50% of the total TCO of the investment.
My point is that getting LLMs into the market is already absurdly expensive, creating a huge cost burden on these companies:
- OpenAI has reportedly lost $5 billion in 2024 alone,
- The CAPEX-revenues gap of Hyperscalers (the four aforementioned companies) appears to be around 12-to-30 times. And growing.
But to the eyes of the consumer, all this is abstracted into one single metric, cost per token. That is, the way these companies generate money is by charging you:
- A fixed price per input token
- A fixed price, usually 2–3 times larger, per output token
That means that your overall costs as a consumer are determined by how many tokens you send the model and how many tokens it gives back to you. And the good thing is that this price is falling almost every month!
Problem solved… right? Sadly, no. The reason is that, again, LRMs change everything.
The economics of LRMs
LRMs introduce three large cost burdens:
- Cost of data. With LLMs, most data was free, and model providers allegedly did not lose time stealing data, infringing copyright law, if necessary.
- Token amount. LRMs, due to their ‘think for longer’ nature, generate several times more tokens per response, especially through reasoning tokens. As tokens have a unit cost, this leads to an explosion of token generation costs.
- Cost of the verifiers. LRMs use themselves or surrogate models to ‘judge’ the quality of their ‘thoughts,’ thereby adding meta-tokens to the mix.
On the one hand, LRMs require reasoning data to learn, data where not only the input and output are present (like in most data) but also the reasoning in between. Most data in the Internet goes straight to the answer, the writer rarely introduces its chain of thought that led to it, making most existing data worthless for reasoning tasks which LRMs are conceived for.
This implies that the AI labs have to curate the data themselves, hiring PhDs and experts in each field to generate this data for training. This requires time from expensive sources (human experts), increasing the global training costs.
At the token level, using OpenAIs own diagram below, LRMs introduce these two new token types that naturally lead to more tokens: what I call meta-tokens and reasoning tokens.
Meta-tokens, as the name suggests, are reasoning tokens where the model thinks (reflects) about its own thoughts (in the diagram above they are incapsulated inside ‘reasoning tokens’. As the model (or surrogate models) have to generate new tokens to make the critic, these tokens increase the cost.
Reasoning tokens, also as the name suggests, are the tokens these models use to reason over the task and their responses to increase the chances of a good outcome.
Of course, due to the exploration factor we talked about earlier, these models ‘try out’ several resolution paths, generating several additional tokens as the model explores, backtracks, and expands its search to new solution paths.
You get the point; LRMs generate high amounts of tokens. In turn, this implies that the previously-described economics of AI used for LLMs no longer work, and to make LRMs profitable, consumers will soon say goodbye to $20/month subscriptions for good.
The perfect example of this is, you guessed it, OpenAI. Currently, they offer two LRMs to their users: o1 and o1-pro (they also offer o1-mini, a smaller version, but the performance degradation, at least in my experience, isn’t worth it).
While o1 is still offered under the standard, $20/month subscription, o1 is a poor man’s LRM, as it doesn’t perform search. Instead, it behaves like any other LLM (one input, one response), but benefiting from better training data, it has a better problem-solving approach (generating a chain of thought to answer).
Still, o1 is hard-capped at this tier; you can only use it for a certain amount of times per week.
Moreover, they also have a $200/month subscription to o1-pro, which adds majority sampling to the mix; it tries several times at solving the task and returns you the best try. Yet, o1-pro doesn’t perform search, meaning that despite it’s still not a fully-fledged LRM, it’s value is already priced at $2,400 a year, ten times more than the standard tier, as the costs that OpenAI endures to serve o1-pro are eating them alive. This is already a prohibitively value for most, but we are just getting started.
Enter o3.
o3, The Great Inequalizer
Although I welcome you to read my extensive review on the detailed differences between o1, o1-pro, and o3, OpenAI’s new, unreleased model, o3, is the real game-changer.
The Impact of LRMs
For starters, this is a real LRM with search capabilities, meaning it thinks for longer by generating a thought tree that mimics how humans solve problems: exploring, iterating, and backtracking over different possible paths until the answer is reached.
And here is where the economics break completely.
According to reliable sources, o3 required millions of dollars spent by OpenAI to obtain results on a single benchmark, the ARC-AGI, whose results were the main source of bragging rights for OpenAI to prove the intelligence increase that this model implies.
But the model is so unbelievably expensive that it can’t be released. Once it’s released, we will probably be looking at a $2,000/month subscription, or $24,000/year, numbers that alienate most of society.
And here is where I reach the main point of the article. Based on the results obtained by the model, I genuinely believe that o3 is the first model that can literally change your life.
For the first time, an AI model obtains human-level performance in reasoning tasks across several domains, not just one like AlphaGo, Google’s Go-playing model; a model that, in your hands, can allow you to create amazing things on demand.
And that virtue is actually a problem for most, as most of society will not be able to afford the price of running this model. This means that those who will have that opportunity, mostly rich people, will have an obscene advantage in many areas, such as:
- Maths
- Coding
- Scientific discovery
- Finance
- Business
You name it.
Simply put, o3 will be a model you can give a humongous task that would require you weeks, months, or even a lifetime, and the model will serve you the solution on a silver platter.
The problem? You won’t be able to afford it, leaving its use to those with pockets deep enough to handle the monthly charges.
Over time, as token costs continue to decrease, these AIs have the potential to push our entire society forward with new discoveries in all these areas.
As LRMs democratize intelligence, making it cheaper than human intelligence, Jevons Paradox will do the rest (deflationary pressures in the production costs of goods and services lead to greater demand that far surpasses the revenue losses due to easier access to these goods and services), but that will take a decent while.
In the meantime, most of us will be left in a state of purgatory, incapable of competing with rich people armed with LRMs, until the costs of these models drop to a point where they are affordable to most.
That’s why I believe AI will not only not decrease inequality but increase it at first, as the rules of the game switch from intellectual curiosity to pocket depth.
And is there anything our leaders can do about it? My optimistic nature says yes.
There are several things I believe governments need to focus on to prevent this seemingly inevitable situation from being too damaging:
Increase power generation:
Through decades of poor choices, especially in Europe, Western countries have suffered from an allergy to energy generation, trying desperately to push green energy sources, which are fundamental to preserving our planet but severely unreliable, in lieu of options like nuclear.
And the key thing here is that we don’t have to choose between one or the other; we need both. Decreasing energy costs can help infrastructure and model providers considerably reduce operating costs, translating into lower customer prices.
AI and education.
Governments need to rethink their approach to education completely; our newer generations aren’t remotely prepared for what’s coming. Instead, we have told them that they have a right to everything, and our kids have grown into young adults who are avoidant of hard work (or aren’t even remotely interested) and that things will stay the way they are.
But AI is literally the land of opportunity; it’s a tool for empowerment, a tool for creation that changes the working paradigm from ‘this is what I want to be, this is what I study’ to ‘here’s the problem that needs solving, this is what I need to accomplish.’
AIs democratize access to knowledge and execution of tasks that, previously, would have been impossible to that person: carpenters building a website for their business, a burger-joint entrepreneur building a new ERP for his/her company, or researchers that accelerate the speed of data analysis from months to hours.
How many students are listening to such ideas? None.
Competition and innovation.
Lastly, we need more competition from incumbents.
AI is already massively concentrated, and things are only getting worse. The absence of competition can very well lead to monopolistic practices, which eventually affect the end customer.
Regarding compute, China is currently challenging the status quo, even having released a model, DeepSeek v3, that was trained with a 10x cost reduction compared to Western models; if the ‘need for compute’ decreases, prices will do too. Also, we need to push for more open-source innovation, as open-source models correlate with strong price decreases as they give free access to tools that rival the proprietary models.