The past few months have left a strange feeling in the tech ecosystem: on one hand, an almost irrational enthusiasm for artificial intelligence; on the other, investment numbers that are starting to seem directly incompatible with any sustainable business logic.
If you connect three pieces — IBM’s data center numbers, Geoffrey Hinton’s warnings, and McKinsey’s cold data — the puzzle of the “8 trillion” (trillions in Anglo-American terminology) paints an uncomfortable picture: a colossal AI infrastructure is being built without a clear economic model to make it viable in the long term.
1. The AI bill: from hype to basic arithmetic
In a recent speech widely circulated in the sector, IBM CEO Arvind Krishna put numbers to what had so far only been an intuition: a single 1-gigawatt megacenter dedicated to AI could cost around $80 billion in CAPEX. Scaled to the ~100 centers some analysts forecast for the next decade, the figure skyrockets to $8 trillion in accumulated investment.
While the specific details of that calculation can be debated, the order of magnitude is what matters:
- AI data centers are not “a little more expensive” than current ones: they operate in a different league of power, cooling, electrical design, and internal networking.
- The critical hardware (GPUs, accelerators, switches, fast storage) has an economic lifespan of 3–5 years before becoming obsolete against new generations.
- At current interest rates, financing several trillion dollars in CAPEX involves hundreds of billions annually just in debt service.
All of this happens while major cloud providers announce annual CAPEX of tens of billions tied to AI and advanced data centers. This spending pace could be manageable… if AI were already generating massive, measurable returns. But here comes the second reality check.
2. McKinsey cools the celebration: almost everyone uses AI, few make money
According to McKinsey’s The State of AI in 2025 report, about 88% of organizations already report using AI regularly, but only 39% claim to see a significant economic impact on their results, measured in profit improvement or EBIT.
Put less elegantly: almost everyone experiments with AI, but very few are truly profiting from it.
The study also highlights several notable patterns for a tech-driven environment:
- Most projects remain trapped in the “pilot purgatory”: shiny POCs, spectacular internal demos… but zero scaled deployment.
- “Top performers” — less than 10% — don’t just automate tasks: they redesign entire processes and business models around AI.
- AI agents (capable of executing actions on systems, not just generating text) appear as the next wave, but remain experimental in most organizations.
Meanwhile, infrastructure investments are being made as if everyone is already a “top performer.” And the numbers don’t add up.
3. The piece Hinton adds: incentives to replace, not just increase
Geoffrey Hinton, one of the pioneers of deep learning and a Turing Award winner, has been warning for months about a less technical and more economic risk: if AI infrastructure becomes as costly as sector calculations suggest, the temptation will be enormous to justify it through massive replacement of human labor, rather than incremental productivity improvements.
Hinton isn’t just talking about “robot apocalypse” scenarios but a more prosaic issue: market pressure to turn these billion-dollar investments into cuts in labor costs, even if that clashes directly with corporate rhetoric about “AI empowering employees.”
Translating into business language:
- If you spend billions on GPUs and data centers, the ROI “PowerPoint” can’t be based on “saving 3% in administrative time.”
- The structural incentive is to automate entire functions, not just assist them.
- This incentive is reinforced when energy and cooling costs of data centers soar and regulators begin scrutinizing AI’s electricity consumption more closely.
The pattern is familiar from other sectors: massive CAPEX + margin pressure = aggressive employment decisions. The difference now is that this pattern applies to infrastructures capable of automating cognitive tasks, not just physical ones.
4. Three blind spots in the “AI = efficiency” narrative
From a tech perspective, the issue is no longer whether the model works technically. We know that LLMs and agents can write code, generate contracts, summarize meetings, or monitor incidents. The real problem is how projects and architecture are being approached:
- Efficiency without process redesign
Many organizations are using AI as a “faster engine” within already flawed processes. Automating redundant paperwork, absurd workflows, or fragmented decisions. The result: errors at greater scale and speed, without creating new revenue streams or competitive advantages. - Leadership disconnected from technical reality
McKinsey finds that companies where senior management is personally involved in AI usage are up to 3 times more likely to capture value.
In too many firms, AI remains “an IT or innovation thing,” while the executive committee signs off on giant CAPEX checks without understanding the necessary operational model to make them profitable. - Overdimensioned infrastructure for eternal pilots
Platforms, data lakes, and GPU clusters are built assuming massive use cases that never leave the lab. The result: underutilized infrastructures draining finances, just as capital and energy costs rise.
5. What are those who profit with AI doing differently?
The small group of companies that achieve sustainable returns share several strategic choices that particularly interest the technical audience:
- Agents before chat toys
They go beyond “cute chatbots”: they connect models to business systems (ERP, CRM, ticketing, observability, etc.) with clear controls, thorough logging, and defined actions. It’s AI that does things, not just generates text. - Architecture designed for TCO, not just demos
They seriously evaluate the mix between:- On-prem GPUs, hosting, or colocation for recurring, predictable loads.
- Public cloud for demand spikes and experimentation.
- Specialized smaller models that are more efficient than huge general-purpose LLMs.
The goal isn’t “the biggest model,” but the best ratio of value / watt / dollar.
- Human-in-the-loop as a design principle
Rather than seeking total autonomy, they design workflows where:- AI handles the heavy lifting (classification, drafting, proposing, planning).
- Experts validate, correct, approve, and feed data back into the system.
This reduces hallucination risks, improves quality, and paradoxically accelerates adoption by building user trust.
- KPI aligned with business, not just technological vanity
Success isn’t measured by “number of prompts” or “hours saved,” but by:- New revenue generated.
- Margin improvement on specific products.
- Reduction in churn, resolution times, critical errors, etc.
6. How should the tech ecosystem respond?
For a reader of a tech publication — CIO, CTO, platform lead, architect — the message from IBM, Hinton, and McKinsey isn’t “stop the machines,” but “adjust your plan before the wave overtakes you.” Some practical ideas:
- Avoid overdimensioning infrastructure layers
Before thinking about clusters of hundreds of GPUs, it makes more sense to:- Start with managed models and measure actual usage.
- Identify the few cases where on-prem or bare-metal makes clear sense (latency, sovereignty, long-term cost).
- Design a flexible hybrid architecture, not commit to a single strategy.
- Redesign processes with AI, not just plug AI into old workflows
Whenever a new use case appears, ask yourself: “If we had this capability 10 years ago, would we have designed this process the same way?”
If the answer is no, it’s an opportunity for reengineering, not just automation. - Safeguard the human factor as an asset, not residual cost
Talent that combines business insight with AI understanding (platform engineers, MLOps, technical product managers, data-savvy analysts) is the only way to ensure these investments don’t turn into white elephants. - Prepare for regulation and energy costs
As AI data centers depend on the electrical grid and leave a carbon footprint, regulators will tighten conditions. Ignoring this variable today may result in restricted or penalized infrastructures tomorrow.
A future with AI… that has to add up in Excel
The uncomfortable conclusion is that the sector is building an AI infrastructure on a historic scale without a proven economic model yet justifying figures of around $8 trillion in global CAPEX.
Between the dream of superintelligence and the nightmare of mass layoffs, there’s an intermediate point: using AI to redesign how we work, not just to cut payrolls. This demands less hype and more technical rigor, less data-center “hype,” and more process engineering.
Because, ultimately, the true frontier of AI isn’t just in the latest chips but in something much less glamorous: making the numbers add up.
Source: Noticias inteligencia artificial