The Bad AGI Business
Can AGI actually turn a profit?
Published: 2025-11-16
Companies are trying to sell everyone on AGI
The AI hype train cannot be more jammed with big corporations like OpenAI, Anthropic, Google, Microsoft, and many others that are racing to train and serve their own models. The marketing pitch is quite straightforward: Artificial General Intelligence (AGI) is just around the corner and it can replace human workers in almost all occupations. AGI doesn’t complain, it is infinitely scalable, and can be upgraded as it improves. If you are sold on this vision, then there is no reason you don’t buy into their product and hop onto the AGI bandwagon. However, even assuming AGI is really achievable in the next five years, is it really a good business model?
What is our AI capable of
Let’s be clear on the definition, this is just AGI not Artificial Super Intelligence (ASI) that can self-improve. I have more doubts about ASI, but it is not too out of my imagination that some sort of AI-powered integrated system can replace a lot of human jobs. Before we go into why the AGI big AI startups are trying to sell could be a bad business, we have to define what is the AI we are talking about:
- It is a robust version of the current AI model. One of the main limitations of current AI models is they are not reliable enough. Say it is only right 90% of the time, and you don’t know where that 10% error is. Let’s assume this is resolved through some mixture-of-experts technique that you just need to pit enough AIs against each other and the majority vote is essentially at or better than human level.
- It is not self-improving in a run-away manner.
- It can be deployed into every industry related to software and documents, but we do not have physical robots yet.
AGI here is basically a function that turns electricity into highly morphable Intelligence that can be easily deployed to a wide range of tasks. Even without being ASI and having physical bodies, AGI provides incredible value to all industries. There is no doubt about that. However, this doesn’t mean it is a good business.
Price depends on supply and demand, not value
The famous diamond-water paradox is a classic example to illustrate this: Water is way more important to human survival than diamonds. However, water is way cheaper than diamonds. This is because water is way higher in supply than diamonds, and even though we can argue diamond has a much lower demand compared to water, the supply and demand interaction still makes diamond way more expensive.
Because of the tremendous value AGI can provide, we can assume the demand for AGI is going to be high. However, let’s not forget about the supply side of the equation. If the supply of AGI is also abundant, then the price of AGI will remain low, which may not be enough to provide the margins needed for a profitable business, espeically currently all the investors are expecting astronomical returns.
Why is state-of-the-art AI model so cheap?
When you ask a question on the ChatGPT website, have you wondered why is this amazing tool free? Interestingly, people have asked the same question about Wikipedia in the past, and practically every great inventions in the modern internet era. The answer is just like every other thing that is unreasonably cheap: someone else is paying for it. In the case of Wikipedia it is a foundation funded by donations, and in the case of AI companies it is the investors that are bullish on making a profit out of the growth of AI in the future.
There are two main costs of producing an AI model: training and inference. Training cost includes R&D costs, hardware, electricity, and inference cost is similar to training but it is tied more to how many customers one is serving. Training cost is like a down payment which the investors and the startups are willing to pay for, hoping the inference cost can be covered by the revenue generated from customers. However, if the inference cost is too high compared to the revenue, then the business model is broken. The whole reason why AI companies are trying to sell everyone on AGI, it is because they are trying to jack up the demand, such that later on consumers will continue to produce the much needed revenue to convince investors to keep growing the company. This means the cheap AI models we can use today is mostly a transient strategy companies employ to get people hooked, and once demand is up and industries rely on AI infrastructure to run, we can expect the price of using these hosted AI models to raise as demand increases. As we can already see in the last year or two, these companies starting to roll out increasingly expensive pricing plans, and promoting solutions that can burn through a large number of credits that get end-users to pay more. However, this may not be enough to save them.
The three scenarios of road to AGI-as-a-Service
Let’s consider three possible scenarios of how AGI will be rolled out into the market, and we can start to see why companies betting solely on AGI could potentially be a bad business.
1. AGI is expensive to train and expensive to run.
This is the most likely scenario, at least in the near future. The expensive cost to train and deploy AGI means only the biggest players can provide the service. In order to serve the AGI to a wide range of industries, the companies have to build massive data centers as well as power plants to support the operation. This means the cost of providing AGI-as-a-Service is going to be very high. The main question for end-users is whether AGI will be worth the cost, or it is cheaper to employ human workers to do it.
Speaking of human workers, don’t forget the heavily integrated global economy essentially out-sourced jobs to low-cost labor countries. It is not impossible to imagine a world that human workers else where benefit enough from the education up-lift by AI, then in turn outprice the AGI services which big AI companies is trying to jack up the demand for.
2. AGI is expensive to train and cheap to run.
This is also a likely scenario, and it is probably the best scenario for big corporations. The model being expensive to train means not that many competitors can enter the market due to the upfront cost investment, and being cheap to run means companies have a better margin to make profit. Still, there are multiple companies big enough to reach AGI. If one company reaches AGI, it is likely other big players can reach AGI soon after. And once again they will start undercutting each other to gain market share, which means the profit margin is going to be thinner than expected.
3. AGI is cheap to train and cheap to run.
If AGI is relatively affordable to train and run, then we can expect to see smaller players entering the market, or companies opt to train their own models. Obviously this undermind the whole business model of big AI companies trying to sell AGI-as-a-Service, because if companies can just train and host their own models at a low cost, then there is no reason to pay for hosted services. In the end, it will be cloud infrastructure providers and hardware manufacturers that are going to benefit from this scenario, rather than the AI model providers themselves.
[Insert your hero] coming down from the top turn buckle
And what is the absolute worst case scenario for big AI companies? If a model is cheap to run, it just take one entity, whether it being a start-up, a country, or an open-source community, figure out how to make a decently functioning AI model that business can adpot at a low cost, then open source/weight the model, these big AI companies basically cannot stop people from hosting their own AI models and serve it to a community of users that is trying to get away from their dream expensive AGI-as-a-Service plans. Companies are always looking to cut cost, and it wouldn’t be long until they figure out they could hire a team of consultants to deploy an in-house AI model that is good enough for their needs, rather than paying for hosted AGI services from one of the existing big AI companies.
This is doesn’t mean big AI companies are completely out of business, and in fact they may still have a decent market share by providing the best-in-class AI models that are continuously improved and maintained. However, having uncontrolled market competition (which imo, is great for consumers) means the profit margin is going to be thinner than they hoped for.
Some concluding thoughts
Cheap energy will be the key to affordable AGI
Remember AGI is a function that turns electricity into solutions? This means cheap and hopefully clean engery is going to be the deciding factor of how much AGI is going to cost in the long run. If I were to start a company today, it won’t be a company that build AGI, since I think it is coming anyway, but I would build a company that optimizes the infrastructure AGI will run on, so when AGI is here, it is cash printing time.
Self-hosted AGI could be great for consumers
Aside from the usual privacy issues, self-hosting AI models could avoid vendor lock-in and the potentially arbitrary price hikes that big AI companies may impose on their customers. This doesn’t mean you have to build the AI, just be ready to deploy the infrastructure at a scale that is profitable. Isn’t that crazy to think about the situation which these AI companies are basically paying out of their investors’ pockets to prove AGI is useful and possible, just for people to get access to a decently functioning open-source model that they can host themselves, and seeing big AI companies left with an empty basket?
My bet is on Google again, unfortunately
Out of all the big lab, Google (and maybe Microsoft) probably once again may be the only one(s) that can survive the AI bubble burst that is likely to happen in the next few years. Google’s integration and ecosystem is just too strong. They have their own cheap compute engines, their data centers are relatively independent, and their software ecosystem is deeply integrated into many people’s daily lives. Google doesn’t have to make the best AGI, they just have to keep it the cheapest and most convenient to use for the average consumer.