Rechenleistungs-Reality-Check
The general solution for the future is: more capacity, more chips, more data centers. All of this is essentially happening because its use in a single professional field or a singular use case has gained momentum. If all white-collar jobs are to be replaced by AI in the future (i.e. hundreds of millions of jobs), this will require a level of computing power that we are still miles away from today. That’s why AI and cloud providers are also promising astronomical investments in new data centers that have not yet been built.
Even if all this capacity is realized, prices must be kept low. After all, all of these data centers need electricity – and energy is not an unlimited resource. OpenAI also recently had to realize this and temporarily shut down its “Stargate” data center in Great Britain because electricity costs were becoming too high.
Looking to the future, this means for AI providers: They must make AI so cheap that it is profitable to replace human labor. At the same time, it is also expensive enough to be able to finance the massive infrastructure investments – and probably also the operating costs. Otherwise nothing will come of the bold plan. The financial service provider Citadel Securities also laconically states: “If the marginal costs of computing power for certain tasks exceed those of human labor, there will be no substitution.”
On a larger level, the main interest of AI companies is not to make their business operations more cost-efficient. Rather, this is a necessary evil in the race to create a digital god – Artificial General Intelligence (AGI).
The day of reckoning
A counter-argument: AI chips are becoming increasingly resource-efficient and therefore more cost-effective to operate. That’s also true: Gartner, for example, predicts that AI inference costs per token will fall by 90 percent in the next few years. However, this does not automatically lead to lower prices; after all, token consumption for AI agents increases even more. “As token consumption increases faster than token costs decrease, the overall cost of inference is expected to increase,” the augurs conclude.
Another counter-argument: just looking at the frontier models of AI laboratories is not enough. Finally, open and specialized, smaller AI models are changing calculations. This cannot be fundamentally disputed either. The only problem is that the best open models today are primarily developed in China, which is known to have its disadvantages. So for this point to make a real difference, the development of open AI models would also need to gain momentum in other parts of the world. In addition, just because a model is open, the inference costs do not simply disappear into thin air.
