The Floors of Public AI. The Facade and the Floors Below

When someone says AI in the street, on the radio or over lunch, they almost always mean a chat. A white box where you type a question and an answer comes out. That surface is the first floor. Above it there's nothing. Below it there's a whole building, and almost nobody goes down.

The Sign at the Entrance

Sometimes the idea is stretched a bit and in come the image generator, the voice assistant, the automatic translator. The set is still the same: a surface the user talks to. That's where the public debate has settled. That's where the features get published, where the open letters get signed, where politicians pose with the laptop open and where the columnists decide whether they're for or against something called "artificial intelligence."

Kate Crawford argued it in Atlas of AI (Yale UP, 2021): AI isn't a user application but an extractive industry, terrestrial and material. To paraphrase her, AI is a planet, not an app. It has a crust, a mantle and a core. The chat is the crust. The part that's visible is the part that no longer decides anything.

So that the sentence isn't just one more slogan in a bloated bibliography, it's worth looking one by one at the floors that hold up the chat. Each works by a different logic, serves a different public and bears different consequences. That the public conversation stays entirely on the first one isn't an accident. It's functional. The building holds up better the less people go up.

The Assistant That No Longer Forgets

The second floor is the assistant with memory. Here there's no longer a use-and-toss chat; there's a system that retains who you are between sessions, what projects you're running, what it corrected wrong last time, what tone you prefer. The decisive piece isn't the conversation, it's the file.

What the assistant knows about you stops being a transcript and starts being a profile. That profile lives on the servers of the company that builds it, adjusts itself with each interaction, and becomes, without anyone signing anything new, part of the model.

The user still believes they're on the first floor, that they ask and get an answer, while the system operates on one where the informational asymmetry between the two parties has already begun to be structural. Shoshana Zuboff described the full mechanics in The Age of Surveillance Capitalism (PublicAffairs, 2019). The pattern she identified for advertising applies with very few changes to this. The difference is that this time the merchandise isn't sold to an advertiser. It's reincorporated into the product the user keeps using.

The One That Acts for You

The third floor is the agent. The word has begun to circulate over the last two years with an almost domestic tone, as if it were the natural evolution of the chat, a slightly more obliging version. It isn't.

The agent is a different piece. A chat generates text; an agent (a program that executes actions in the world on the user's behalf, not only drafts answers) buys tickets, sends emails, contracts services, accesses your bank account if you've given it permission, writes code and deploys it, opens tabs, fills in forms, answers calls.

What on the first floor was a risk of bad wording becomes here a risk of bad operation with consequences in the physical, legal and economic world. The academic state of the art on agents was mapped by Park and others in Generative Agents: Interactive Simulacra of Human Behavior (arXiv:2304.03442, 2023); since then the field has moved faster than the liability regime that should cover it.

When a chat gets it wrong, you delete it. When an agent gets it wrong, you have to undo the operation. Sometimes it can't be undone.

The Concrete Plumbing

The fourth floor is the compute infrastructure. Here the human conversation disappears entirely and buildings appear. Data centers (industrial facilities housing thousands of specialized servers to train and run AI models) with footprints the size of small cities, cooled by water circuits that in some regions are starting to compete with the municipal supply, powered by electricity contracts of figures nobody would have agreed to sign five years ago.

The Energy and AI Report of the International Energy Agency (IEA, 2025) quantifies the order of magnitude. Data centers' global electricity consumption is projected to double to some 945 TWh in 2030, a figure comparable to Japan's total electricity consumption today, and the centers optimized specifically for AI would more than quadruple their consumption in that same window. That's not an abstract metric. It's a concrete territorial negotiation, municipality by municipality, over who gets the available power and at what price.

The companies that sign those contracts aren't the ones offering the chat to the end user. They're the capacity providers, almost always the same three or four companies, almost always with soft state support, almost always in places with cheap water.

The user who opens the white box sees none of this. They wouldn't have to see it if it worked well. The problem is that the decisions made on this floor, about where it's built, with what consumption, buying water from whom, in turn condition everything that can be done on the three floors above. If there's compute, there's a product. If not, no.

The Material the Model Is Made Of

The fifth floor is the most opaque and the oldest. It's the data and surveillance infrastructure.

Models aren't trained with material that appears by spontaneous generation. They're trained with text, image, video and audio extracted from public digital libraries, from social networks, from customer-service transcripts, from administrative databases, from medical records, from urban-camera images, from mobile-app traffic.

Part of that material is obtained with a license. Another part is obtained with an ambiguous consent the user accepted years ago in a form they didn't read. And another part is obtained just like that, assuming that the cost of the later legal dispute will be less than the benefit of the trained model.

From Assistant to Population-Management Instrument

On top of that data, the systems that are no longer chats but decision mechanisms have also been built. Credit scoring (an automated score that decides whether a bank grants you a loan), recidivism-risk assessment, résumé screening, allocation of social resources, predictive policing. The AI Now Institute, which Meredith Whittaker co-founded before moving in 2022 to lead Signal, has spent years documenting that transfer, in which the same technical capacity sold to the consumer as an assistant is sold to the State as an instrument of population management.

The one asking the chat about a pasta recipe and the one watching an algorithm deny them a benefit are operating with pieces that come from the same building. The market operation presents them as separate worlds. They aren't.

Doing the Math

It's worth stopping a moment and doing the math.

The user acts on the first floor. Their accumulated profile lives on the second. The actions an agent executes on their behalf happen on the third. The physical capacity that makes all that possible is decided on the fourth. The data that feeds the system and the institutional uses of that same system are spread across the fifth.

When the user protests or applauds something, they do it almost always from the first floor, with a vocabulary built on the first floor, and aimed at problems that are actually generated on the floors they don't see. It's like arguing over the color of the hallway wall without knowing whether the foundation holds.

Who Benefits From You Looking Only at the Sign

That asymmetry doesn't hold up on its own. There are concrete interests behind the debate staying up top.

A company that wants to sell a consumer product would a thousand times rather the public conversation be about whether its chat is polite or not, than about how many megawatts its next data center consumes in a drought-stricken region. A government that wants to deploy predictive tools in social services would a thousand times rather the debate be about whether students do their homework with ChatGPT, than about who audits the model that decides which family gets a child-protection alert. A media outlet that needs to generate traffic would a thousand times rather the feature about the chat that falls in love with the journalist, than the feature about the electricity contract the regional government has signed with a hyperscaler (a large cloud data-center operator, able to absorb compute demand at industrial scale) for fifteen years.

Each of those biases is understandable on its own. Added up, they produce the current result. A public conversation saturated on the showiest floor and empty on the four that matter.

Democratization in Reverse

The usual defense against this asymmetry is the promise of democratization. If the tool is available to anyone, the argument goes, power is distributed. The sentence sounds good and gets repeated a lot.

It's exactly the reverse. The more the use of the tool spreads on the first floor, the more the technical and economic capacity that holds up the floors below concentrates. Every new user who gets used to the chat is a new user who can no longer do without the compute provider underneath. The mass of users finances, via subscriptions or via data, the consolidation of the few players who control the fourth and fifth floors.

Stanford HAI's AI Index Report 2026 documents that concentration: the compute dedicated to training the cutting-edge models has been growing at a rate of several multiples per year since 2020, and the hardware that makes it possible is dominated by a handful of providers, with Nvidia controlling the bulk of the capacity. The ability to train at the frontier fits today in very few hands.

Democratization of use, concentration of decision. They aren't contradictory things. They're the same thing seen from different sides.

The Recognition From Inside

Mustafa Suleyman, who isn't an external critic of the sector but one of its architects, wrote in The Coming Wave (Crown, 2023) an internal version of the same diagnosis. Serious technological waves aren't contained, they're assimilated, and when they're assimilated they reorganize the distribution of power in a way that can no longer be reversed from the user's floor.

The recognition arrives late and from inside, when it no longer commits anyone to anything. The useful thing about the book isn't the warning. It's the confirmation that the asymmetry is structural and not a communication failure that can be corrected with better pedagogy.

The User Manual and the Other Discussion

There's an operational consequence of all this worth looking at slowly.

Any discourse on AI that stays at how to write better prompts (the instructions the user writes to guide the model), at how to use the assistant for personal productivity, at how to avoid hallucinations (invented answers the model presents as true), at how to teach children to use the tool with critical sense, is a discourse useful on the first floor and almost irrelevant on the other four.

It's not that it's done badly. It's that it's done for a layer of the problem that doesn't touch the decisions that generate the problem. The user manual is a legitimate genre. The political discussion about the infrastructure is another genre. To confuse them, to present the first as if it were enough to solve what only the second can touch, is the rhetorical operation that sustains the status quo.

The Price of Going Up a Floor

Going up a floor costs. It costs reading time, it costs new vocabulary, it costs following copyright lawsuits, it costs understanding why an electricity-supply contract in Aragón or in Iowa matters to the conversation, it costs reading Crawford and Zuboff instead of productivity influencers.

And then, once you've gone up the floor, there's no visible reward. It doesn't produce likes. It doesn't produce clients. It produces, at most, a sustained discomfort with the friendly version of the product everyone is using.

The attention economy disincentivizes the trip. That's why almost nobody makes it. That's why the public conversation has spent three years talking about the neon sign while the building's plumbing is designed with no witnesses.

From Which Floor the Answer Comes

The question that opens this text isn't whether AI is good or bad. The question is from which floor you're looking when you answer.

Whoever answers from the first responds about a white box that writes answers. Whoever answers from the fourth responds about a territorial negotiation over water. Whoever answers from the fifth responds about who decides which child gets taken from you or which loan gets denied to you.

The three answers use the word AI. The three refer to phenomena partly connected and radically distinct. A mature public conversation on the matter begins the day we stop pretending they're the same. That day, today, hasn't arrived.

Definitions

Agent. An AI program that doesn't limit itself to generating text but executes real actions on the user's behalf: sending emails, contracting services, buying tickets, writing and deploying code, handling accounts. The difference from a chat is that the agent's mistake has material consequences that can't always be undone.

Data center. An industrial facility that houses thousands of specialized servers to train and run AI models. It consumes very large quantities of electricity and cooling water, which turns its location into a political decision about the distribution of a territory's resources.

Hyperscaler. A large cloud data-center operator, able to absorb compute demand at industrial scale. In practice they're three or four global companies that concentrate the bulk of the sector's training and inference capacity.

Prompt. The instruction the user writes to the model to guide the answer. The culture of "well-written prompts" belongs to the first floor and rarely touches the decisions made further down.

Hallucination. An invented answer the model presents with the same confident tone as a correct one. It's not a wording error or a deliberate lie: it's a consequence of how the answers are generated, word by word, from probabilities.

Credit scoring. An automated score a system assigns to a person to decide whether they get a loan, a mortgage or an insurance policy. It's one of the classic institutional uses of the fifth floor and one of the most opaque to the citizen being assessed.

References

Crawford, K. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence (Yale University Press, 2021). The general framework of the article: AI understood as an extractive industry with physical layers, not as a user application. The formulation "AI is a planet, not an app" is a paraphrase of her argument, not a literal quote. https://yalebooks.yale.edu/book/9780300209570/atlas-of-ai/.

Zuboff, S. The Age of Surveillance Capitalism (PublicAffairs, 2019). The reference for the second floor, on how the user's accumulated profile is incorporated into the product without renegotiation of terms.

Park, J. S. et al. Generative Agents: Interactive Simulacra of Human Behavior. arXiv 2304.03442 (2023). The reference for the third floor, on the academic state of the art of generative agents.

International Energy Agency (IEA). Energy and AI Report (April 2025). The reference for the fourth floor, on the electricity and water consumption of AI-associated data centers: projection of a doubling of global consumption to some 945 TWh in 2030 and of consumption by AI-optimized centers more than quadrupling in that window. https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works.

AI Now Institute. The institute's work on the concentration of power and the institutional use of AI systems. The reference for the fifth floor, on the transfer between consumer assistant and population-management instrument. The institute was co-founded by Meredith Whittaker, president of Signal since 2022. https://ainowinstitute.org/.

Stanford HAI. AI Index Report 2026. The reference for the section on the concentration of training compute: growth of several multiples per year since 2020 and domination of the hardware by a handful of providers, with Nvidia leading the capacity. https://hai.stanford.edu/ai-index/2026-ai-index-report.

Suleyman, M. The Coming Wave (Crown, 2023). The reference for the sector's internal diagnosis on the irreversibility of the asymmetry once the technological wave is assimilated.

También te interesa

En otros sitios

#concentration-of-power #data-centers #surveillance #energy-cost #open-models