TL;DR - Recursive self-improvement will drive AI labs towards identical products and zero profit margins.

The last few years have made one thing clear: intelligence is valuable.

This value extends beyond the raw number-crunching capacity of computers. The sparks of creativity seen in LLMs and the nascent understanding emerging in world models can be sold before either technology approaches a true general understanding of the world. Non-general, domain specific intelligence can be sold.

Frontier AI labs are not content to sell this domain specific intelligence. They are research labs, with head-counts in the thousands and eleven-figure funding rounds, with a clear unifying goal. They are here to create general intelligence.

General intelligence is clearly extremely valuable, being the basis of the entire knowledge labour market. The economic value generated by AGI is comfortably estimated to be in the trillions, and investors are pouring capital into these labs with the expectation of immense returns once this technology is realised.

However, there is a sense in which general intelligence is fungible by design. Being suitably high-performing across domains, running on the same hardware, and trained on broadly the same data results in relatively little room for differentiation. If the underlying software becomes too similar across multiple model providers, then the firms become locked in a competitive equilibrium where profit margins are driven to zero. Even if the technology is transformative, the firms developing it and their investors may see little of this windfall in their annual profit figures.

Recursive Self Improvement and Homogeneity

Recursive self improvement (RSI) occurs when a model gains sufficiently high awareness and skill in the domains needed to develop itself, that it begins to meaningfully assist in that development. Once its capabilities grow as a result of its own effort, it is even more able to grow its own capabilities. Until bottlenecked by compute or data, this can be expected to result in an exponential increase in model capabilities.

The path to AGI and beyond to ASI likely involves RSI at some level. Human researchers and developers are required to get a system to a level where it can meaningfully contribute to its own development. Then, it helps these researchers accelerate their work, until the system itself is good enough to do entirely autonomous AI R&D. Unless there is no intermediate capability where AI is a useful companion researcher to humans from current models all the way up to AGI, RSI will play a role in this development. It seems likely that later on, when the AI is doing much of the research, RSI will be the driving force behind changes to the model.

Looking at the current model training stack, there is a lot to be improved. We almost certainly don’t use the optimal C compiler, the best ML backend libraries, or the best statistical inference techniques at a high level. But while a single perfect optimum may not exist, there is likely a set of approaches that achieve essentially equivalent performance given the same hardware, data, and tasks^[1].

All frontier labs are currently running similar hardware and have access to the same small number of high-end hardware providers. The datasets they are trained on are largely publicly available or otherwise similar to one another. If task-specific models are the profitable angle of differentiation, a sufficiently advanced AGI after undergoing RSI is presumably capable of creating these too for any firm with access to such an AGI^[2].

With enough compute and enough time, any approach to AGI capable of achieving RSI will converge on a meaningfully undifferentiated set of approaches to model development assuming shared hardware and data.

RSI is presumably highly capital intensive, requiring very large amounts of model inference to get meaningful increases in performance. Using expensive reasoning models, debate frameworks and other ways of further eliciting the capabilities of models can give impressive outputs at very high cost. It may take a long time and great expense before this software convergence has fully set in. Regardless of this, with similar enough hardware and data, it seems likely that this software convergence will eventually occur.

Hardware Availability is Largely Uniform

If AGI labs are, to a first approximation, selling access to identical general intelligence, then what differentiates them?

They aren’t necessarily offering access to the models (in many cases, the models themselves are a closely guarded trade secret). They are offering inference of those models. Here, there is space for innovation. If I can offer access to general intelligence inference for less than a competitor, then I can undercut competitors and open up a profit margin.

When pricing inference, there are three components: the hardware the model is run on, the software used to facilitate this inference, and physical factors of operating the server (such as electricity and data interconnect costs).

The hardware being used is approximately the same for every lab. Hardware providers aren’t AGI labs and AGI labs aren’t hardware providers. Very little differentiation can be assumed here^[3].

The software being used in the inference is presumably being written (at this point of the development of AGI) by the AGI itself. If the AGI which separate firms have access to has converged to a similar form under RSI, then the software output will be largely the same (given a compute budget).

We are left with only the physical factors of the data centre operation. Once all of the AGI labs have moved their servers to Iceland for cheap energy and free cooling, it seems there is very little innovation to be done. Even if there are some innovations in cheap data centre operation which don’t come about clearly from AGI development, such as Deepmind’s Alpha Evolve, the profit margin comes from cheaper inference and not from AGI itself.

Anatomy of a Bad Investment

An investment into an AGI company, assuming that money doesn’t lose all meaning after the creation of general intelligence, is a bet that the discounted cash flows from that company outperform standard market returns.

This requires not only that the company makes something that people want, but that it is differentiated enough to demand a profit margin while doing so. Water companies, agricultural goods providers, and electricity generators are all providing critical services to the functioning of the economy and to human survival. This is not in itself grounds for earning a profit. The offering must also be differentiated enough to open up a profit margin.

The path to perfect competition and diminished profit margins can be summarised as follows:

1. Intelligence is valuable: This is clear both from the existence of the knowledge labour market and from the explosion in AI revenue over the past few years. You can sell intelligence.
2. Hardware Availability is Standardised: Labs have access to the same hardware for the most part. The largest exception to this is Chinese frontier labs such as Deepseek not having access to some high-end chips.
3. RSI Leads to Homogeneity: Recursive self-improvement, when run on approximately the same hardware, with the same data, and optimising for interactions with the same economy, will converge on a number of approaches with very similar results.
4. Perfect competition drives margins to zero: Identical products rarely sell for a profit. If another firm can provide general intelligence at the same price you can, prices will decline until one firm is selling the general intelligence at cost. If inference costs are the same between providers, perfect competition dictates that no firm will make a long-run profit^[4].

For AGI labs hoping to make money from selling access to their models, there are three broad approaches to earning a return on investor capital.

1. Hope that the other labs never make it, and you remain the only provider of general intelligence.
2. Make enough money while you have a monopoly on AGI to recoup the entire investment.
3. Establishing enough political influence in their short period of market domination to obtain power which never disappears.

Option 1 seems very unlikely, given the current distribution of compute and the lack of major frontrunners among the labs.

Option 2 is possible. OpenAI is looking at a roughly $500 billion valuation, which is a plausible amount of money to make with a 3-6 month monopoly on AGI.

Option 3 is the most concerning. That the military, political, and societal dominance that even a short period of access to leading AGI without meaningful competition allows for centuries of monopoly profits to be brought in. This is certainly not the world I want to life in, but it’s not implausible to me.

Monetisation Beyond Software

The argument I have presented states that AGI labs will not profit substantially from AGI, not necessarily that they won’t earn a profit.

AGI labs will have user bases with switching costs, good corporate relationships, and brands which keep customers coming back. These are small draws when the underlying product converges on being identical, but it might be enough to keep the businesses alive.

The elephant in the room regarding model differentiation is data collection. Codex, Claude Code, and the Cursor Tab offer a constant stream of feedback signals for refining coding models. ChatGPT still has far more monthly active users for its chat service than competitors than other providers, helping it refine a better chat model that its competitors. ElevenLabs is the go-to provider for TTS services, giving it better real world data about the applications of TTS models than anyone else.

There is space for differentiation, and therefore a profit margin, with specialised models excelling at particular tasks. While the value of these models may increase when they are created by fine-tuning or otherwise post-training an AGI-level model, the profit available to the AGI labs is not from the AGI itself but from the proprietary data used to specialise it.

As an example, take OpenAI’s superior proprietary data regarding conversational chat models built up by the large user base of ChatGPT. In the case where RSI has resulted in multiple firms having access to homogeneous and capable AGI, OpenAI can use this data to create a specialised model better able to chat to users. Or, they could sell this data and the constant feedback signals made available to them by their user base to other model providers. These other firms would need to purchase the data to remain competitive, and would then have to sell the resulting model at cost to properly compete once other firms had also integrated the data into their models.

The profit margin comes from selling data and crucially not from selling access to AGI. Even if AGI increases the economic value of that data and the models created with it, the firms selling AGI-inference remain trapped in perfect competition with margins dragged to zero.

This is somewhat analogous to what happens in the hedge fund space where Bloomberg and other financial data vendors make large profit margins even if their customers are all locked in ruthless competition. The difference here is that hedge funds are differentiated by their software and analysis, whereas AGI labs selling the product of RSI on the same hardware, with much of the same data, are not so differentiated.

Conclusion

At present, labs are differentiated by their software and their capacity for data collection.

If advanced AI is achieved which is capable of recursive self-improvement, after sufficient time and compute is expended the AI will converge on similar software. This is assuming that approximately the same hardware and pre-training data is available, which appears true across frontier labs. It applies regardless of the starting point, whether the model is an auto-regressive transformer, a diffusion model, or some other architecture so long as it is capable of recursive self-improvement.

Having the same hardware and software leads to labs being unable to provide a differentiated offering regarding AGI. Perfect competition drives margins to zero, leading to none of the AGI labs making an appreciable profit from the technology (so long as none of the labs obtain enough of a lead to enjoy a monopoly).

Labs will still have access to proprietary data and user feedback. They will be able to profit from this either by training their own models to incorporate the information or by selling it to other labs. Either way, there will be a profit margin from data availability and approximately none from selling access to AGI.

Investing in a lab with the stated goal of creating AGI has always been a strange investment. Success could remove any meaning of money, destabilise the world enough that returns cannot be claimed, or even destroy the world entirely. If the output of these labs is a true general intelligence, created from the same hardware and data by a continuous improvement cycle, then it is doubtful that this by itself will provide a meaningful return on that investment even if the world remains stable enough for an investor to try to claim one.

1. ^{^}

   More likely, there are a very large number of approaches with approximately the same performance when scaled up sufficiently. Different starting points will, under RSI, converge on subtly different implementations. For a sufficiently advanced AI with enough time and compute, these will be meaningfully competitive enough with one another that the starting architecture should be largely unimportant to eventually realised capabilities.

2. ^{^}

   The intelligence made available by labs creating AGI remains undifferentiated even if specialised models are the primary offering. A suitably advanced AGI could create the software required for those more specialised models. If this underlying model is largely the same across different model providers, then the more specialised models it creates could be expected to share those similarities.

3. ^{^}

   If this separation changes, and either a hardware provider creates AI capable of RSI or an AGI lab starts creating its own hardware, the firms may begin to appear differentiated. But such a firm could sell its superior hardware to other labs, which would then sell inference on that hardware of their mostly identical models at cost. The profit margin still comes from the hardware production, and not from the AGI inference being sold.

4. ^{^}

   This is only true in the case of perfect competition, and in the long run. In real competitive markets, there are small amounts of profit available for multiple firms in the short term.