I read that SAP’s share price had fallen by 47% and that, amongst others, Salesforce (-43%), Adobe (-47%), Accenture (-57%) and Intuit (-65%) fared no better, despite the fact that the profits of these listed companies remain largely positive to this day.

Financial market analysts have highlighted the increasing use of AI by these companies’ customers as the common reason why the market is punishing them.
For my part, I am fairly convinced that AI is one of the main reasons why these companies are losing value, although it is not the only one.

AI: the jack-of-all-trades?

In fact, customers are finding that the opportunity cost of engaging a third-party consultant or purchasing products, services and licences from these companies is higher than that of subscribing to AI subscription plans offered by various providers such as Anthropic, Microsoft or OpenAI.

In truth, I believe that clients have also realised that the goods and services provided by consultants are often paid for at a price that is significantly higher than the potential revenue derived from using those goods and services.
Moreover, these goods and services are very often themselves produced and supplied through the intensive use of AI, which has forcefully entered all production processes.

Paying $12,000 for consultancy on a corporate design system (often a set of graphic components with no material connection to a company’s front-end components) is uneconomical compared to redesigning those components using AI.
And this is all the more true for Salesforce and SAP services, given that these IT service providers employ tactics to force customer loyalty, imposing a compulsory , lifelong partnership on them: a very costly partnership.

There are also those, such as Adobe (with its new AI-powered Photoshop), who are trying to stay afloat. But why use Photoshop when you can use a good generative AI model with effective prompts?

But the hangover might wear off

In short, companies supplying goods and services to other businesses (particularly if these customers are highly structured) do not seem capable of posing a serious alternative to AI.
However, I am convinced that the markets are highly volatile, and above all, I am convinced that they have overestimated the impact of AI.

In all likelihood, these companies will recover; AI won’t replace us all (at least not in the immediate future, at least not until the trend changes); and we should go back to thinking of AI as a tool – one of humanity’s greatest inventions, but still just a tool.

That’s why.

Costs spiralling out of control

AI has been sold to us as the El Dorado, but companies – all companies – are realising that AI subscription plans are rapidly becoming more expensive and that the market for AI providers is an oligopoly, one that is, moreover, heavily concentrated in the US, which does little to keep prices down.

They are also realising that AI is discriminatory, though not for the reasons it is fashionable to cite in sociology departments: rather because data centres are energy-intensive and require vast quantities of water, comparable to those needed to run a small nuclear reactor (or, if you prefer more post-modernist units of measurement, comparable to the water requirements of dozens and dozens of Sudanese villages, because it suits us to mention the Sudanese).

AI models, therefore, discriminate on the basis of the most traditional and ruthless crtierion: price.

The most incredible case concerns GitHub Copilot.
Since the start of June, with the new policies, the consumption of tokens required to readthe input and generate an output ( not including cache tokens, analysis tokens, etc.) has skyrocketed, and unless you take remedial action by setting a company-wide limit per user, you risk burning through hundreds of dollars in just a few days. Per user.

We are therefore witnessing a wave of inflation, but not the traditional kind: token inflation.
The cost of tokens has risen exponentially in just one month.
Models such as Claude Opus 4.7, used by GitHub Copilot Chat, are capable of consuming 5% of tokens for a single request, particularly if they are asked to perform a multi-agent analysis.

The cost in tokens is even higher if the AI is accessed via the GitHub Copilot CLI, which uses around 1% of the tokens just to set up the Markdown instruction file.

The prospect of switching to a competitor (such as Anthropic, given that the best models – which are also used by GitHub Copilot – are still Anthropic’s) may seem tempting, but the truth is that there is no guarantee that Anthropic’s licence prices will remain low, nor that they will always provide the same volume of output, all other things being equal.

Moreover, their budget models (such as Claude Haiku 4.5) are not particularly effective for automating production processes compared to ‘more expensive’ models, such as the latest Claude Sonnet and, above all, Claude Opus.

AI is more tangible than you might think

The most forward-thinking companies today are taking out AI licences costing between $75,000 and $100,000 a year, even for very small IT, Data or Research and Development departments.
Some of these leading companies, driven by security policies or the potential for long-term cost savings, are even moving towards developing their own AI systems, with results that are not always encouraging (particularly in the banking sector).

But everyone, whether highly innovative or not, is coming up against the same reality: that AI, and the web in general, are far more tangible than they realised.
Perhaps we have all underestimated the importance of infrastructure. And we all owe an apology to the backend developers and systems engineers.

Cloud services, AI, external services – all services that seem far removed from reality, yet in fact rely on enormous, bulky, energy-hungry, powerful infrastructures. Robust infrastructures.

And perhaps, as Italy and as Europe, we should start seriously considering building our own infrastructure ; and perhaps using nuclear energy to power it; and perhaps securing the materials needed to build it.

In all this, it is right to ask ourselves ethical questions – and, above all, about the environmental impact of AI – but in the meantime, we need to invest in infrastructure and integrate AI into our production processes to make them more efficient, without sacrificing the essential, critical and decisive human contribution.

AI may well be the key to dominance in the 21st century, but anyone relying on cloud services hosted by servers in data centres in a foreign country (or even an enemy country) is taking an existential risk.

Here too, the Ukrainians have much to teach us: the struggle for survival has driven them to do things that we find hard to comprehend, but which we should – and soon will have to – take as a model.