Predicting the future is, by definition, a limited exercise. I consider any forecast about what lies ahead either misleading or, at best, a rough estimate — and that is precisely the point: an estimate, no more reliable than yours, though supported by documented past events you may not be examining. Anyone who can genuinely foresee the future either lived it or has the means to force it into being. Everything else is projection.
On Predictions and What They Are Worth
There are only two cases in which someone can truly predict what is coming: when the person travels from the future and describes what they lived, and when an individual holds the means to bend the future toward their prediction, making it real by consequence. Outside of these cases, any prediction is either misleading or merely an estimate — and that is what this is: an estimate as reliable as yours, but grounded in documented events, which you may not be consulting.
The Fear of Replacement
There is a widespread fear that AI agents could steal our credit, our jobs, our ways of thinking and acting, and — by all that is sacred — even our interpersonal and personal relationships, turning us into the antagonists of our own lives. In other words: not you anymore, but AI as the decisive factor.
The concern is understandable when, looking around, you see thousands of people relying heavily on AI for medication choices, lifestyle decisions, and personal relationships. It is at once comic and tragic when people replace real advisors with automated agents. But is this it, then? Will we be replaced, and does this mark the beginning of the end?
No. And again: no.
The Real Limitations of AI
The Cost No One Sees
First, we need to understand something basic: AI cannot do everything, and it is limited — very limited. This limitation manifests across several dimensions, including processing order, energy, cooling, data quality, and, of course, financial scale.
Ask yourself how much you paid to use GPT, and whether that amount could possibly cover the cost of financing, construction, and the thousands of developers who have spent years building and refining the model. You will likely begin to grasp the size of the barrier.
The Model Wars
AI systems are expensive — far more expensive than the general public imagines — and we demand more from them at an accelerating pace, because there is a war between models not unlike the Browser Wars, but at a significantly larger scale. Most people cannot fathom how much energy each small request to an AI chat costs: the analysis of an image, the generation of visual content, the answer to a complex question. Most simply accept this service as “free,” because they have never received a bill like the ones OpenAI almost certainly does.
Just to satisfy curiosity: the company’s infrastructure through 2035 will cost one trillion dollars — an astronomical figure, though not representative of real value today. Estimated at $111 billion through 2027 [1], OpenAI’s annualized revenue reached $20 billion in 2025 (+233% vs. 2024), yet projected infrastructure spending stands at $600 billion through 2030 and $1 trillion through 2035. This means a promise of future value, even after receiving one hundred billion dollars in investment from NVIDIA — which, curiously, will serve to finance OpenAI’s rental of NVIDIA’s own GPUs [2]. Consider further that Microsoft made a similar move by investing approximately $250 billion, on the condition that OpenAI consume its Microsoft Azure services [3].
The Cost Structure
The primary costs of AI services can be organized — even if roughly — as follows: hardware and energy infrastructure, cloud services, and network services — alongside other slices that correspond to development itself. The investments from NVIDIA, Broadcom, and Microsoft exist more to soften the blow of system costs than to finance real growth. There is still no clear metric for the actual gains and real costs of such a system.
As for the values generated and the investments poured in, everything seems to rest on the assumption that people will use AI more than they already do and pay more than they already do for it. Whether OpenAI is a $500 billion company or not, it is generating $20 billion in cash in 2025 alone; if it maintains that extraordinary growth — without losing market share to Anthropic or Grok — it will take several decades to pay back Microsoft’s investment.
This is where many people assume the whole system is paying for itself and that the cash generated is flowing to shareholders. But it is not: all cash flow is consumed simply to keep the system running and competitive. Nothing is left over for shareholders, nor for any margin of comfort.
When the System Does Not Pay for Itself
The SORA Case
Generating text with AI is expensive. Generating an image is more expensive still, consuming exponentially more processing power and energy. A video in full HD requires magnitudes more — which is precisely why images and video are restricted in most AI services.
If a system does not pay for itself, it dies. That is exactly what happened to OpenAI’s SORA [5]: it did not die as a standalone product from lack of fame, but because too few people are willing to pay to generate video memes on the internet. The system consumed far more processing per response due to its graphical demands — and this is also why Claude does not generate images directly in chat. Not because of lack of capability — they genuinely have it — but because the system does not pay for itself.
Having buried its AI capable of generating videos and animations “similar to Disney’s,” OpenAI also terminated its one-billion-dollar deal with the animation studio [6]. It is not as though the company wanted to walk away from a Disney deal — it is simply that it is not profitable enough. Forrester analyst Thomas Husson told BBC News it was “a black hole of resources” with “limited monetization” [5].
The Problem with the Free Version
The blunt truth is that AI systems, although genuinely useful and impressive, are not paying for themselves. They consume enormous amounts of energy and processing; with every new chip contract, competition makes them more expensive — and with few semiconductor companies in the market capable of meeting the challenge, scarcity grows more brutal, driving up RAM and processor prices in a cascade felt worldwide.
This is a system that currently survives on hopeful investment — a bet that AI, in ten to fifteen years, will generate enough revenue to sustain itself.
How can they afford to wait? Use cases. We are using AI more and more in our daily lives — not just ChatGPT, but Claude and dozens of others that accompany us from YouTube recommendations to Google search simplifications. Yet how many of those people actually pay, or are willing to pay, for it? Setting aside commercial use cases or paid add-on features, most are content with the free version — which is deeply problematic for AI companies.
Each company has its own strategy, but with one thing in common: restrict the free version until not paying becomes uncomfortable. Notice how Anthropic handles this: you get access to the free version, but your tokens run out so quickly once conversation complexity scales that you almost feel the need to pay — if there were no other free AIs available. Or Grok, which appears to throttle free-tier access when its servers are under heavy load, reserving full access for the paid version. It will not be long before ChatGPT adopts the same stance.
The Energy and Structural Impact
Beyond the costs already mentioned, it is worth intensifying the energy metric. Some research suggests that a single query to ChatGPT can cost up to ten times more energy than the same query on Google — not because Google is more efficient, but because it does not need to over-filter and applies a delimited context to your query. It is, in other words, a search across its data centers based on PageRank, followed by display. ChatGPT, by contrast, is considerably more complex in terms of data processing, which drives up the cost.
With an estimated 9 billion daily searches, this would translate to an additional demand of nearly 10 TWh of electricity per year [7]. In other words: if demand keeps growing, you may not need to share your job — but you will certainly share your electricity, and likely other resources, now that you also compete for processor power and RAM [8].
“Computer costs have fallen steadily over the last 40 years until the AI boom reversed that trend.” — Oxford Economics
It is worth noting, however, that chips have achieved meaningful efficiency gains specifically for AI workloads — at least +40% efficiency improvement with the Nvidia Blackwell architecture (2026).
Is It a Bubble?
Investors expect the AI market to reach its peak of $1 trillion in infrastructure — or even $5 trillion [9] — pouring ever more money into companies that currently generate no real wealth, for users content with the free version, while the price of microchips and memory ripples through the personal computing market in what I call unfair competition, also driving up electricity costs. It is the bet of all bets.
“Notably, this excess investment is being funded by the public…” — TIME [9]
The answer on whether this is a bubble depends. A bubble is generally defined by a market-moving event; the “burst” occurs when one actor begins a withdrawal that incentivizes another, feeding a spiral that tends toward zero. No one wants to be the last one on the boat — but this only makes sense when the “boat” has no intrinsic value. And the truth is, AI does have value — substantial value. That is not the question.
The point is not that AI is useless or valueless, but that perhaps this value does not justify — or equal — an investment of roughly $1 trillion in a matter of years. Even being the right technology at the right moment, things can still go very wrong if investments are absurd and real revenue remains distant.
The Future of Programmers and Professions
It is therefore my estimate that AI cannot — at least not now — replace all mathematicians and programmers, let alone every other profession, because there is a shortage of energy, processing, memory, and, above all, resources. Even while riding the current wave of hype, the system demands ever-larger investments in the six-digit range, beyond its cognitive and structural limits. And it still lacks the capacity to replace senior developers.
To imagine that the advance of AI, combined with trivialities and its operational cost, will be sufficient — within five to ten years — to upend the market and reduce it to a NoCode/AICode audience is an excessively optimistic and, to some degree, lazy estimate.
Without substantial and sustained investment, AI is, definitively, an expensive and fallible experiment. The real question is when — and whether — someone will pull off the first major withdrawal. I doubt the system reaches zero, because large corporations, military sectors, and state entities maintain a vested interest, and they can mobilize public resources to scale equivalently powerful systems, even if at a significantly smaller scale.
There is also the possibility that some companies will not fully replace their workforces with AI — particularly because tokens are expensive and run out at an extraordinary rate. However, they may maintain less efficient, smaller-scale systems dedicated exclusively to internal coding. This is not impossible and could become a trend if decentralized AI solutions become more widespread and accessible. Even so, at least one human professional will still be needed to operate or validate AI output — which does not eliminate human presence, but raises the level of specialization required, possibly increasing costs in the process.
Conclusion
Will AI replace all professions? No, and the answer is not merely political — it is structural. The system does not have sufficient resources for that now and will likely not have them within ten years. AI can displace functions, compress layers of intermediate work, and automate repetitive tasks, but fully replacing complex professions — those requiring judgment, context, accountability, and continuous adaptation — is an entirely different order of problem. AI cannot pay for itself by doing that at full scale.
Is it generating profit now? Not in any real, sustainable sense. What exists today is subscription revenue and corporate contracts that barely cover operational costs. The rest is financed by investors betting on a future that has not yet arrived. All generated cash flow returns immediately to keep the system running and competitive — nothing is left for shareholders, nor for any breathing room.
Is there an estimate for generating profit in the next ten years, or only structural cost? There are optimistic estimates — and that is precisely what they are: optimism. The premise is that adoption will grow, that more people will pay, and that infrastructure costs will fall as the technology matures. It may happen. But it depends on fragile variables: a drop in semiconductor costs, real expansion of the paying user base, and the absence of a major confidence collapse in the sector. The scenario of dominant structural cost persisting for another decade is perfectly plausible.
The accumulation of armed conflicts around the world — among them the Russo-Ukrainian war, American-Iranian tensions, and growing instability across the African continent — is already exerting pressure on the semiconductor market, with impacts that will likely intensify over the coming years on AI infrastructure.
In a scenario of escalation toward a broader conflict, the relentless demand for chips and memory from AI companies would compound military and industrial needs, exhausting reserves, creating lengthy supply chain delays, and driving hardware prices to prohibitive levels. In that context, the first models to become unviable would be precisely those financed by investors and offered free of charge to users — the very ones that today already operate on the edge between cost and revenue. There is no structural reason for investors to sustain free services indefinitely under conditions of scarcity.
This risk already has expression in the markets. TSMC, the world’s leading manufacturer of advanced chips, recorded significant stock appreciation, rising from approximately $28 to $69 over the past year — growth exceeding 100% [10]. SK Hynix, one of the three largest RAM manufacturers in the world, followed an even steeper trajectory, climbing from roughly $130 to approximately $960 over the same period, representing appreciation exceeding 500% [11]. [Data from 2026 with rough currency conversion.]