Demystifying AI: Energy, Privacy and the Bubble Question

AI, like everything else, has its good and bad sides. It is a technology of light and shadow, but if we look at it objectively, we can make the most of its advantages and protect ourselves from its threats.

Where This Article Comes From

This article grew out of a round table on innovation and artificial intelligence held on 1 October at the second edition of Arahealth's innovation event in Zaragoza. The panel was moderated by Pere Estupinyà, and I took part alongside María Barrié, head of innovation at L'Oréal.

We talked about how AI is changing consumption, recommendation and the way we create, and about how to use it without weakening our ability to think, argue and debate. AI can help us do things, learn, make mistakes early and iterate, as long as we keep our human judgement.

The other talks widened the debate to the economy, geopolitics, power and sustainability. That is where this article comes from: reviewing some common claims about AI and telling apart real risks, exaggerations and questions that are still open.

What Is Worth Clarifying About AI

Claims about AI tend to mix verifiable problems with incomplete comparisons and predictions that are still uncertain. In each section I will separate what we know, what depends on context and what is worth keeping an eye on, without presenting the technology as good or bad in itself.

Energy Use

The energy consumption and emissions linked to AI are understandably controversial. The useful question is how much is consumed at each stage, how much electricity is used and what value that use produces.

A generative model such as ChatGPT or Claude goes through several phases: defining its goal, collecting and preparing data, choosing its architecture, training, evaluation, deployment and maintenance. Both training and later use consume energy, and the weight of each phase changes with the model and the number of queries.

There is no single figure that tells us how much AI pollutes. Training BLOOM, a generative AI model with 176 billion parameters, used about 433 megawatt-hours. The different phases each model goes through change its emissions a great deal.

At home, electricity is usually measured in kilowatt-hours (kWh) and, at large scale, such as factories or production lines, in terawatt-hours (TWh) (one terawatt-hour is a billion kilowatt-hours). The International Energy Agency (IEA) estimates that data centres used 485 TWh in 2025 and could reach about 950 TWh in 2030, around 3% of global electricity demand by then. AI is a major driver of that consumption, but not all data centre consumption comes from AI.

Energy use also depends on the task. A study that compared 1,000 inferences (each time a model is used to produce an answer) found that image generation used more energy than text-only tasks. It also found that foundation models, also called general-purpose models (ChatGPT, Claude and so on), could use several orders of magnitude more energy than specialised models.

The impact is not limited to electricity, even if electricity is the main source and the current bottleneck. Data centres also use water, directly to cool equipment and indirectly to generate electricity. One study estimated that training GPT-3 (released in May 2020) in US data centres may have directly consumed about 700,000 litres of fresh water. The water footprint also varies with location, climate, time of day and cooling system. Ideas such as putting data centres in space are still speculative and cannot yet be presented as a solution. But if they ever become reality, they would save a great deal of cooling water and be a considerable, innovative step forward.

So, Is AI an Environmental Problem We Should Take Into Account?

Yes, but it is better to avoid a yes-or-no answer. The impact depends on training, use, the type of task, the efficiency of the data centre, where the electricity comes from (renewable sources or not) and how much water is available locally.

If the benefit of each use justifies the resources it takes, and providers measure, publish and reduce that footprint, then it would be a sustainable and constructive investment. Technological innovation and environmental sustainability could go hand in hand.

Sources for This Section

International Energy Agency (2026), Key Questions on Energy and AI

Luccioni, Jernite and Strubell (2024), Power Hungry Processing: Watts Driving the Cost of AI Deployment?

Luccioni, Viguier and Ligozat (2022), Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model

Li, Yang, Islam and Ren (2023), Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models

Data Privacy

Here, too, there is no black and white, but a scale of greys worth understanding, so you know what is being used and whether your privacy is at risk.

Generative AI models can be divided into open and closed models. Closed models are harder to customise, especially at the deeper levels of how they work. Open models are much more customisable and, above all, you can download them to your own computer.

That is the big advantage of open models today: everything you generate stays on your own machine, which is safer than using third-party apps if what you want is maximum privacy. The downside is that the leading models have at least 150 billion parameters, and a computer that can run one needs a powerful graphics card and plenty of RAM. That said, many tasks do not need the most powerful model: there are smaller open models that run well on a laptop.

The Alternatives Today

If you are not willing to run a model on your own computer but still want privacy and security for your data, these would be the next options.

One clarification first: anything highly confidential is better not shared unless you are using a local model.

For client data, contracts, investments, wellbeing, application code or business strategy, the best option is an Enterprise or Team account with a contract, or the API key these companies offer. Conversations go to the providers' servers, but by default they are not used to train their models, so the level of security is higher.

One step below are individual paid accounts (Plus or Pro). Here, unless we switch it off in the settings, companies can use our conversations to train their new models. In ChatGPT you turn this off in the data controls, and in Claude each user chooses whether to opt in. Gemini handles it differently: if you keep your chat history, human reviewers may read some conversations. If you turn history off, you get more privacy and this does not happen (although Google keeps conversations for up to 72 hours).

With free models, they will almost certainly use our data for training unless we switch it off.

Type of information	Which tool to use
Public or generic	Free plan
Internal but not sensitive	Paid plan with training turned off
Client data, billing and sensitive data	Business plan with a contract, the API or a local model
Highly confidential or very sensitive information	A local model, or no AI at all

Sources for This Section

OpenAI, Data Controls FAQ

Anthropic, Updates to Consumer Terms and Privacy Policy

Google, Gemini Apps Privacy Hub

Revolution or a Bubble About to Burst?

AI-generated image.

I agree with the idea that the past is there to learn from, but always while looking ahead. Living in the past is never good: it feeds our ego and keeps us from improving.

Last year, at a technology conference in Italy, interviewed by the chairman of the carmaker Stellantis, Jeff Bezos explained what he thinks of the AI industry and whether or not it is a bubble about to burst.

I liked what he said because it drew a line between a financial bubble and a technological, or industrial, one.

The 2008 financial bubble, after the fall of Lehman Brothers, was the clearest example that when you offer something with no real value, no transparency and no principles, everything can collapse overnight.

When the internet arrived, just before the dot-com bubble, there were huge companies that poured enormous resources (largely borrowed money) into laying the cables that would connect the whole world across the oceans. The amount of money invested was so large that several of these companies went bankrupt, but the value they created remained. If we are so connected today through fibre optics, it is thanks to projects that failed as businesses but changed the world, and that made getting online far cheaper and more accessible than it was back then.

For Bezos, the bubble we may be in today has more to do with the second kind than the first. He also warned that, in this frenzy, investors struggle to tell good ideas from bad ones.

Several studies also agree that AI only delivers a real return to companies willing to rethink their workflows so that AI becomes a unifying layer, not an add-on.

So let's go back to the past and compare figures and estimates for what may be another case of overvaluation, but one that could also bring major advances.

The Numbers, as of October 2026

According to Sequoia Capital, in 2023 AI needed to generate about 200 billion dollars a year in revenue to justify the investment being made, and it was still 125 billion short of that figure.

By June 2024, that figure had risen to 600 billion dollars a year to sustain all the demand and infrastructure.

Five years from now, according to Bain, it would be 6 trillion dollars a year.

The investment figures are already frightening, but there is more.

In three years, AI investment has grown more than in any previous boom relative to its starting point. The technology keeps renewing itself; it is not static. Between 70% and 90% of the spending is estimated to go into equipment (chips, servers, networks) that is replaced roughly every five years. And if you don't upgrade, you will probably notice the difference. So capital expenditure has to be constant.

On top of that, with more agentic capabilities, tasks use more tokens and resources, even though the cost of the same task is currently falling about tenfold each year.

Now the Positive Side

Unlike in the dot-com era, the main AI investors are companies with a lot of cash, so this is direct investment in the technology by leading firms. That said, they are relying more and more on debt: according to the Bank for International Settlements, between 30% and 40% of this investment is already financed through debt and leases.

AI is also generating real revenue. Cursor, which was bought by SpaceX, Elon Musk's company, and is now part of SpaceXAI, grew its revenue twentyfold in 14 months (to 2 billion dollars). And leading companies that treat AI as something strategic, not as a tool for writing emails or social media posts, are improving their EBITDA by 10% to 25%.

Then there are the astronomical valuations of OpenAI and Anthropic, and their 2026 revenue. Anthropic, the company behind Claude, led by Dario Amodei and his sister Daniela, is billing at a run rate of about 65 billion dollars a year (if it kept July's revenue for the whole year), and the New York Times estimates it will pass 100 billion. OpenAI is already at a run rate of about 70 billion a year, although its actual 2026 revenue is estimated at around 36 billion, and it expects to break even (stop making losses) in 2030. Anthropic would get there in 2027 or 2028.

That OpenAI is worth more than Toyota, or that Anthropic is worth as much as Walmart (both before going public), is just the first sign of the huge expectations placed on these companies.

Sources for This Section

Sequoia Capital, AI's $600B Question

Bain & Company, Technology Report 2026

Bank for International Settlements, The AI investment race

Bezos at Italian Tech Week (Malay Mail)

SpaceX acquires Cursor (CNBC)

What Is Still Worth Watching

There are several other points to weigh when evaluating a technology like this: how powerful it could become at its current pace, the risks it carries, the concentration of power in a few market players, how ethically it is used and behaves, and its reliability and the mistakes it can make (a 97% success rate at each step of a 30-step process becomes only about 40% by the end). That does not make these systems bad or frightening, but it is good to know their strengths and weaknesses so we are prepared and not caught off guard.

If you are interested in this kind of article, a comment would be great, so we can go into more detail with examples, illustrations or whatever is most useful to you.

To me, it is clear that a technology like this can bring major advances. Going too fast can stop us from enjoying the landscape and each stage, obsessing over reaching a place we are not sure is good. Speed is not always progress, but if it comes with sense, judgement and care, we will certainly be heading in the right direction.

And you, which side of the scale would you put AI on today?

Thanks for reading.