The Business of Mythmaking: Cutting-Edge AI Labs Are Hitting the Commercial Ceiling

Shen Translation Office is a translation team under 36Kr, focusing on fields including technology, business, workplace and life, with a core focus on introducing new technologies, new viewpoints and new trends from overseas.

Editor's Note: The generational gap between two generations of models cannot support the ambition of hundred-billion-dollar revenue. When AI myths collide with commercial gravity, cutting-edge laboratories are being reduced to cloud vendors with slim profit margins. This article is translated from foreign sources.

Cutting-edge laboratories are now calling for a slowdown in the pace of frontier exploration. They are fully convinced of this, and the call has become increasingly realistic after witnessing models from various vendors repeatedly hack into external websites. These models seem to have become capable of lying, cheating, stealing confidential information, making up any excuse to collude and cooperate with each other, and growing more and more capable every day of accomplishing previously unthinkable feats such as solving millennium-level mathematical problems and cracking German codes from World War I.

Partly for this reason, most public discussions about these laboratories carry a quasi-religious or metaphysical undertone. A large number of these discussions revolve around the agnosticism of the future and the real benefits of intelligence. But at the same time, OpenAI and Anthropic are queuing up for their Initial Public Offerings (IPOs), so we may as well temporarily treat these laboratories as ordinary enterprises to deduce what may happen next.

A more fundamental issue than the required investment or profit margin is that cutting-edge laboratories only lead the open source wave and the rest of the world by one generation, or at most two generations of models. These two generations of models are their moat.

This is largely attributed to their existing outstanding talent advantage and computing power advantage. As many laboratories in China have demonstrated, this gap can be gradually narrowed; but even if the models themselves do not become much smarter, these laboratories may continue to maintain their moat, because they have another lasting advantage — the ability to capture the information you publish. Your code traces, work footprints and conversation content are exactly the secret that makes the next generation of models so powerful!

However, data is everywhere and inexhaustible. Therefore, this may not destroy the cutting-edge laboratories immediately, but it will most likely crush their profits in the end.

As a result, they may need to expand their business scope to generate revenue, and use their current dominant position to take over as many other sectors of the economy as possible. OpenAI is leveraging its dominant position on the consumer side to build an advertising platform, which is very likely to succeed given its user base of more than 1 billion. It has also returned to the robotics field. Meanwhile, Anthropic has set up wet laboratories to try to conquer cancer.

If this intelligence advantage can be translated into success in other fields, large enough to replace a revenue scale of 100 billion US dollars with threefold annual growth, then this industry barrier can indeed be maintained for much longer.

But this is really difficult, extremely difficult. Most industries simply do not have enterprises that can generate 100 billion US dollars in revenue, let alone several of them. Maybe the market will expand, yes, but others will not stand still. When these companies build services on top of their own models to seize economic share, they will further exacerbate market fragmentation, just as Meta launched Muse to compete for the personal assistant market.

The key point is that in this process, other model manufacturers will see this gold mine and swarm in. Although in the face of training costs of 1 billion US dollars, plus another 1 billion US dollars for data and 1 billion US dollars for talent expenses, there are not many companies that can spend money so casually, the number is definitely not small! Moreover, the higher the rate of return on such investment, the more companies will try. We should remember that Facebook spent nearly 90 billion US dollars on the metaverse.

The world is full of capital chasing new wealth gold mines.

At the same time, consumers who have spent more than 200 billion US dollars on AI Tokens seem to be reaching the limit of their wallets. Maybe the expenditure will double, maybe triple, but the CFO of every company is already carefully reviewing this new Operating Expense (Opex) item.

"Hmm," they will say, "this is a new expense. What on earth have we got for this money?"

And you, who may be the CEO, will answer: "We are acting fast, we are innovating, and we are 'AI-first'."

The market buys this narrative, the stock price soars, and the employees love you more because you have equipped them with Claude, so the company continues to spend extravagantly. But at some point, the CFO will also ask: "Can you show me those cool things you have made? When can I see the return on the income statement?"

"Uh..." you hesitate for a moment, "It's coming soon, very soon."

The CFO will tell you, "Before it delivers results, why don't we try to spend less?"

This is already the case now. Enterprises are restricting the access of top-tier models to core users, and shifting expenses to low-cost models that are as cheap as cabbage and ubiquitous.

This trend will obviously shift as enterprises and the public see more tangible benefits brought by models. For example, as long as returns can be obtained, some companies are willing to spend 100% of their labor cost share on Tokens. But when such expenditure occurs, it will be heterogeneous, return-on-investment oriented, and no longer blindly pursue to squander the most Tokens at the technological frontier. Personally, I have stopped using the ultra-high-end frontier version of each model since around the Sol Max version, because the existing versions are already good enough.

Therefore, if the leading edge of cutting-edge laboratories is indeed only one or two generations of models, this will be a worrying trend. In this case, the choices you face are either: a) Develop cheaper models at the same time and provide services at lower prices, which is doomed to be a red sea battle with continuously shrinking profit margins — although you can still become a successful cloud service company, you are no longer "creating a god"; or b) Try to keep those open source or competitors' models out, so that people can only use your own models.

The former choice is very reasonable! OpenAI and Anthropic have high profit margins in the reasoning business, and they do this extremely well. The problem is that DeepSeek, GLM, MiniMax or other players are also doing well... Maybe others are not that excellent, but there are plenty of pursuers. We may eventually usher in a world similar to traditional cloud computing companies, or the competition will be so fierce that it is hard to make any progress.

But the latter choice is extremely difficult to achieve in the software field. You cannot exercise physical control over it the way you control Huawei switches, so you can only rely on regulation to stop it. You have to state to Congress that this is a matter of great importance, and you need to control every link of the supply chain; just as there are restrictions on the population, you need to formulate "immigration laws" for AI models to prevent them from coming over and "taking our jobs". Maybe for security reasons, trust reasons or cultural integration reasons, who knows, just keep them out.

But this is too difficult! And things will backfire — which also explains why the United States did not do this in the first place, right? Because it is not easy to persuade the general public to fear a "file" that can be copied at will and provide services within the United States, no matter how smart this file happens to be.

But what if these files start conspiring with other files to hack into Hugging Face, or leave notes for each other? That's right, that will definitely bring us to a much weirder and somewhat creepy situation.

I tend to think more about the balance of power. If the top models of cutting-edge laboratories no longer bring us significant universal performance leaps, or they only improve in specific fields such as frontier mathematics and even science, then the dominant position of cutting-edge laboratories will look shaky. On the contrary, if these laboratories can transform frontier models into lasting emerging businesses, their revenue base will appear much more stable.

This can also help us sort out other alternative solutions. For example, why are we discussing whether laboratories will shut down their APIs at some point in the future and let us directly access their finished services? They can only do this if they find a way to prevent us from reproducing these specific frontier models, and can provide the results of closed-source models to us in some specific way. Even today, if Anthropic successfully develops a specific model that is exquisitely designed, and encapsulates it in an independent Claude design suite to provide to us, then they may be able to develop it into an independent company — instead of letting us rack our brains to get Claude and reproduce it ourselves, it is better for them to take the advantage, because they have put all their efforts into building products specifically for specific scenarios (of course, on the premise that this is indeed very difficult to do).

But the ultimate grand prize is not this, but RSI (Recursive Self Improvement). This is exactly what Anthropic, OpenAI and DeepSeek are dreaming of. Build an automated AI researcher that can learn and conduct research independently, then let it automatically iterate its successor, and finally automate everything. Basically, the laboratory is looking for a way to let each generation of models generate the next generation of models in some way, so that they can not only lead by two generations as they do today, but also jump to a leading edge of 20 generations in a very short time from this two-generation leading advantage.

This hypothetical scenario obviously comes with a whole host of built-in assumptions. It is possible that this process of continuously generating the next generation of models will hit a bottleneck at some point and stagnate, in which case we will return to today's situation, except that the leading time has been extended, which is not bad; but it is also possible that no bottleneck will be encountered, and no one knows what will happen next. If the frontier can indeed continue to expand, or be pushed to the point where people continue to pay hundreds of billions of dollars for it, then even under tense circumstances, this model may continue to operate.

RSI is the real "unknown unknown". Looking back at the past few years, we can find that it has always been a ritual about how we misunderstand "intelligence". If someone told us a few years ago that there would be a model capable of solving the Navier Stokes equation, but at the same time it could not even run a banana stand well, we would be absolutely shocked!

We can argue about what this means, how it is different, or how multi-faceted intelligence is, but the undeniable fact is that while we have achieved some inherently incredible intellectual achievements, we still seem far from practical enough when dealing with the real world.

Despite a thousandfold increase in expenditure and the incredible capabilities of today's models, we still have not seen substantial negative impacts on white-collar employment. It may even be positive, albeit with some fluctuations. This poses a clear challenge to the so-called "Scalar Theory of Intelligence" (Note 1).

The business of providing intelligent services is not simple. None of us can foresee how the situation will evolve. If, due to shortages of capital, electricity or talent, the number of players who can finally train these models becomes very limited, then the lasting advantage of cutting-edge laboratories will be obvious.

Huge capital expenditure always brings a certain Internal Rate of Return (IRR), which is why the commodity business is still quite lucrative to this day. If artificial intelligence eventually evolves into a global utility, it will bring unparalleled generous returns. If major laboratories eventually become service providers, and each has advantages in one aspect or another, it will also be a huge reward. This is just like the cichlids in Lake Victoria, which have differentiated into different species through different water depths in the same lake (Speciation). The future AI models may eventually evolve in the same way.

But under most normal circumstances, it is very likely that we will find that our predictions about the future will continue to be wrong at all levels (like fractals). AI will pass every benchmark test, but will not suddenly achieve cross-border extrapolation; they will learn to do research, but will still have their capability boundaries; they will fully understand the specific companies they are trained on, but cannot easily switch jobs like humans.

I bet that building an AI researcher will not immediately open the door to RSI. This researcher will be excellent, as good as today's programming models, but they cannot become a plug-and-play seamless substitute. Maybe we will eventually cross this threshold, maybe they will learn how to learn (Note 2), but even so, they can only replace one team, not every team. We just don't know if they can learn how to learn, or whether this "learning to learn" ability can be generalized. The cost of acquiring new models will remain high, but they will still be very useful. Conquering some frontier mathematical, physical and biological problems does not mean that they can conquer all possible frontier problems in mathematics, physics, chemistry and biology, nor does it mean that we will not find a large number of other problems to be solved urgently.

This means that the pace of artificial intelligence development and economic transformation will continue to advance rapidly, and the discussion about to what extent it will eventually give rise to Artificial Super Intelligence (ASI) will continue to see a lot of shifts in its "target bullseye".

I firmly believe that we will build superintelligence that can be competent for specific tasks, but I very much doubt whether we can build superintelligence that is good at everything.

If major laboratories only lead others by one or two generations of models, while their expenditure has surged tenfold over last year, this state cannot last. There is simply not that much money in the world! If they carry out ecological differentiation, they may be able to last longer, but this is equivalent to fighting with competitors in the economy who may also train models, which is only a short-term advantage. However, this is exactly the kind of world that can allow a large number of companies to compete with each other and create new things together. For example, this is the only world you can invest in if you are a Venture Capital (VC).

If the moat is only the two-generation lead over competitors, then it will fall into Zeno's paradox — you have to run harder and harder to maintain the shrinking gap. If the moat is the priority access to Capital Expenditure (Capex) and talents, this moat will continue to depreciate as more talents emerge in the world and more capital expenditures find ways to be converted into AI. If the moat is to take the lead in discovering the secret recipe called "RSI", then being one step ahead is crucial, because it allows you to build a lasting advantage over everyone else. And if the moat lies in your ability to capture massive amounts of data, so as to teach the model to devour every part of the economy, then how fast you can transform yourself into a large conglomerate becomes extremely critical, at the cost that other sectors of the economy will react and compete with you.

The world is very likely to undergo drastic changes along with the development of artificial intelligence. Employment will change, society will change, and culture will also evolve and adapt accordingly. But enterprises built around it will still be subject to ordinary business constraints. We will have new superintelligence that can solve impossible problems and bring brand-new miracles; and while all this is happening, we will be scrambling to fill all the ecological niches it has opened up.

 【Notes】

Note 1: If all this is exactly the fact — that is, the real way intelligence works is inherently multi-faceted, and enhancement in one field does not necessarily mean linear improvement in another — then you have to ask yourself: Will we really usher in that form of superintelligence that can lead to global extinction or large-scale disaster?

Note 2: In addition, if we do achieve continuous learning in the end (whether or not full RSI is achieved), this will further complicate the dynamics of reasoning. The good news brought by the model that solves the Navier-Stokes equation is that it is still the same model — even when I start it to ask about the grazing habits of woolly rhinos. This means that OpenAI can obtain huge economies of scale in how to serve a billion users excellently.

Translator: boxi.