The day I realized the emperor had no clothes
Six months ago, a CEO called me in a panic. His team had just burned 1.2 million euros on a "revolutionary" AI chatbot project. The result? An assistant that performed worse than his website's FAQ, with an inference bill that doubled every quarter. "They promised us artificial general intelligence," he told me. "What we got was a parrot on steroids."
This story is not an isolated case. It has become the norm. And it is time to say it plainly: generalist LLMs as we know them have hit a ceiling. The tangible progress of GPT-5 and its clones is measured in small percentages, between +3% and +7% depending on the benchmark, and these gains no longer come from the model itself, but from the human engineering around it: Mixture of Experts, fine-tuning, Chain-of-Thought, prompt orchestration. The engine is stalling. It is the mechanics who are still pushing the car forward.
The law of diminishing returns is hitting LLMs head-on
The data wall
The fundamental problem is brutally simple: generalist LLMs have consumed nearly all usable Web data. The corpus of quality textual data is finite. The internet is estimated to contain around 250 billion pages, but after filtering out spam, duplicates, and low-quality content, the exploitable fraction has already been ingested by current models.
The consequence is mathematical. When you have read every book in the library, rereading the same volumes does not make you smarter. LLMs are churning through exhausted data, and each new iteration produces diminishing marginal gains. The scaling law curves, so often brandished by AGI evangelists, are starting to flatten. This is not just my opinion: the researchers at the major labs themselves are acknowledging it behind closed doors.
AGI: a fantasy at least 5 years away
Let us be direct: AGI is a marketing argument, not a technical reality. When Sam Altman promises artificial general intelligence "in a few years," he is talking to his investors, not to his engineers. Current Transformer architectures, as brilliant as they are, do not produce reasoning. They produce sophisticated statistical prediction. The difference is fundamental.
An LLM does not "understand" your question. It calculates the most probable token sequence based on its training. It is an engineering tour de force, but it is also a structural ceiling. To cross the threshold toward a truly general intelligence, we would need architectural breakthroughs that nobody has any concretely deployable idea of today.
The financial abyss of consumer chatbots
While the media celebrates every new version of ChatGPT, the numbers tell a different story. OpenAI burns 5 billion dollars a year. Consumer chatbots devour colossal inference budgets to generate conversations with virtually zero commercial value. Goldman Sachs dropped the figure: 1 trillion dollars invested in AI for "few tangible results."
Meanwhile, 57 S&P 500 companies openly fear they will never recoup their AI investments. We are no longer in the enthusiasm phase. We are in the phase where executive committees start demanding accountability.
Why this plateau changes everything for your enterprise strategy
If you are a CIO, CTO, or executive, this LLM ceiling has a major strategic implication: the race to the biggest model is over for you. That race belongs to the hyperscalers (Google, OpenAI, Anthropic, Meta) who have the billions needed to eke out a few benchmark points. Your competitive advantage will never be the size of your model. It will be what you do with it using your data.
This is a complete paradigm shift. For two years, enterprises have been chasing the technology. It is time to chase value.
The "plug and play" illusion
Too many organizations still treat LLMs like off-the-shelf software. You plug in ChatGPT Enterprise, deploy Copilot, and wait for the magic to happen. It is like buying a Ferrari and parking it in a muddy field: the problem is not the engine, it is the road.
Your data is the road. If it is fragmented, ungoverned, locked in departmental silos, and polluted by duplicates, no LLM in the world will produce value. Data governance is not a sexy topic. It is the topic that separates the 5% of AI projects that succeed from the 95% that fail.
The pragmatic roadmap: RAG, governance, and ROI from the first iteration
RAG as a value architecture
Rather than waiting for a hypothetical miraculous GPT-6, the enterprises generating ROI today are building RAG (Retrieval-Augmented Generation) architectures anchored in their proprietary data. The principle is simple: instead of asking the LLM to "know" something, you provide it with the relevant context extracted from your internal knowledge bases, and it synthesizes a response grounded in your actual documents.
But beware: naive RAG is also a trap. A simple vector search over PDF documents is not enough. As I have detailed in my analyses on GraphRAG, the real value emerges when you structure knowledge into relationship graphs (components, services, rules, dependencies) enabling reasoning that follows chains of causality, not just lexical similarities.
Three priority workstreams for 2026-2028
1. Domain-specific agents for critical processes. Forget the generalist chatbot. Build specialized agents that automate specific workflows: claims processing, compliance analysis, B2B lead qualification. An agent that does one thing well is worth ten chatbots that do everything badly.
2. Hybrid inference pipelines (cloud and on-premise). Sovereignty is not a European whim. It is an operational necessity. Your sensitive data must remain under your control, and your inference architecture must function even if an American cloud provider decides to change its terms overnight. The hybrid approach (lightweight models on-premise for sensitive data, cloud for the rest) is the only resilient strategy.
3. Proprietary knowledge bases. Your competitive advantage lies in the knowledge your competitors do not have. Capitalize on it. Structure your domain knowledge, formalize the tacit expertise of your employees, and build knowledge bases that become the fuel for your AI systems. It is this informational asset, not the underlying LLM, that will create lasting differentiation.
ROI from the first iteration
Here is a rule I systematically apply with my clients: if your first AI deployment does not generate measurable value within 90 days, you have a framing problem, not a technology problem. ROI does not come from the model. It comes from the precision with which you identified the business process to optimize, measured the current cost of that process, and defined the quantifiable success criterion.
The enterprises that succeed do not start with technology. They start with the question: "Which process costs me the most and could be partially automated?" Only then do they choose the tool.
The hype is over, good riddance
The plateauing of generalist LLMs is not bad news. It is a liberation. It forces us to move past the AGI fantasy and return to the fundamentals of value creation: understanding your business, governing your data, measuring your results.
European investors have a historic opportunity here. While Americans continue to burn billions in the race for the biggest model, we can invest in AI that concretely transforms the back-office and middle-office of our enterprises. Not chatbots. Not impressive demos. Systems that reduce costs, accelerate processes, and improve decisions.
The hype is over. Time for real value. And if you do not know where to start, it is probably with your data.