The economic viability of AI: Why Microsoft and Uber are now hitting the brakes

The capabilities of large AI models are surprising in a variety of areas. For complex tasks that are very labour-intensive, the AI model produces an initial result in a very short space of time. Whilst this result still needs to be checked by experts and possibly refined further, it enables massive time savings to be achieved. Examples include summarising or comparing texts, creating graphics or presentation slides, carrying out an initial financial assessment of a company, or providing a wide range of support in software development and quality assurance.

AI as a cost factor: Why even large companies are limiting its use

Now, reports are emerging that companies such as Microsoft and Uber have restricted the use of external AI tools such as Cloude. The reason cited is the costs incurred through their use. This makes it clear that, in a business context, it is not just about the speed at which challenging tasks are solved; it is also necessary to assess how cost-effective the use of such tools is.

Providers of large AI models, such as Anthropic and OpenAI, have changed the billing structure for their models and further restricted their use. Whereas there used to be flat rates for using the systems, billing now takes place on a pay-as-you-go basis according to the number of tokens consumed by users when accessing the providers’ APIs. Tokens can be individual words, parts of words or punctuation marks that the AI model processes. These are therefore tiny units. Even if a single token costs only a fraction of a cent, very high token counts can quickly accumulate for larger tasks. Consequently, the cost of solving a task is correspondingly high. Due to the nature of the tasks, this becomes significantly more costly when using Agentic AI as autonomous systems to solve them.

One might then critically examine what it actually cost to solve a task using AI – taking into account the billing and the necessary manual follow-up work – and compare this with a solution relying purely on human staff, which is, of course, more time-consuming. Added to this is the fact that, as a rule, staff resources are available and therefore incur their own costs. In business practice, we can therefore observe a rebound effect for certain tasks, arising from the new pricing situation.

Cost-effective alternatives: How businesses can use AI cost-effectively

So how should one go about this? There are various alternatives to consider. Firstly, for each specific task, one should check whether there are smaller, lower-cost models – or perhaps even open-source variants – that can achieve similar results. Providers are also beginning to differentiate their offerings in this area. Secondly, one might consider using only European large language models (LLMs), such as Mistral, for these tasks; this model also boasts performance that should not be underestimated, whilst offering even more affordable pricing models and flat rates. Good results are often achieved with it. Finally, at an advanced stage of AI deployment, one can also differentiate between which tasks are best handled by the expensive LLMs and which by the more affordable ones.

Generally speaking, the prices for such solutions also factor in the costs of high computing power (hyperscalers). It is important to note here that energy costs (for servers and air conditioning) account for up to 60 per cent of running costs. Electricity prices at American data centres are around 50 per cent of those at German or European data centres. This, too, is a disadvantage when attempting to operate large LLMs in Europe.

Testing different LLM variants naturally involves a certain amount of effort. This must be taken into account. Furthermore, data protection considerations come into play for many tasks. As a result, American hyperscalers are often ruled out. One must therefore rely on European or local solutions. In practice, this means that, on the one hand, the use of LLMs is subject to many constraints, whilst on the other hand, cost-effectiveness must be considered alongside time-related factors. Staff who have previously carried out these tasks are still there, and costs are incurred. In many cases, rapid changes are therefore not the preferred option. In the medium to long term, there will be a shift towards LLMs, and demographics will also contribute to this. And the most powerful LLMs will not always be the preferred option, as they are expensive to use. However, the focus will always be on areas where economic success can be directly measured. This certainly also explains the rather hesitant adoption of AI models in many sectors.