Billion-dollar data centres, or autonomous systems that hit the target precisely? As large language models become a game played not by countries but by tech giants, the real opportunity for start-ups and institutions lies in the “agentic AI” layer. The future will belong not to those who build giant models, but to those who put them to the right work.
Dear readers of Ekonomi Dünya, standing at the edge of the digital economy and looking out, it is no longer possible to deny that the technology world has entered a new era. On one side are companies such as Google, Microsoft, Amazon and OpenAI, investing hundreds of billions of dollars in data centres; on the other, countries and institutions trying to keep up with these giants. France, the United Arab Emirates, even China feel the weight of this race deeply. The picture tells us one thing: the race to develop large language models has become a game that only trillion-dollar companies can play, not countries.
So how should we read this picture, this arena where giants devour one another? More importantly: where are the real opportunities hidden for start-ups, institutions and countries that move with strategic intelligence through this chaos? In this article, I go after exactly these questions.
Between 2023 and 2025, the goal many countries and institutions set themselves was clear: “We will build our own large language model.” The intention is understandable. It springs from a perfectly legitimate concern for technological independence, national security and data sovereignty. In practice, however, the picture looks very different.
The compute cost of training a GPT-4-class model from scratch is estimated at between one hundred million and one billion dollars. Add data infrastructure, hardware supply chains, research teams and ongoing fine-tuning costs, and the scale becomes striking. France's Mistral AI is one of the strongest European responses in this field; yet even Mistral clearly shows how hard it is to compete with global leaders in foundation model development.
The key question is this: is entering this race really “strategic” for a country or institution that could channel its resources elsewhere? The data suggest otherwise. Foundation model development is a low-margin, high-cost, long-term activity full of uncertainty. Applications built on top of these models, by contrast, can create value much faster and reach profitability far sooner.
The real war in the AI ecosystem will be fought not between foundation models but between the “agents” that turn those models into meaningful functions. Agentic AI, systems that can make autonomous decisions and carry out sequences of tasks towards specific goals, is currently the fastest-maturing layer of technology. Its defining feature is that it is not a single AI call but a process. Imagine an integrated system that analyses customer requests, routes them to the right department, fills in the necessary forms and reports the solution. Doing this requires neither Google's data centres nor billions of dollars. What it requires is the right architecture, sector expertise and strategic focus.
The Power of Vertical Expertise: The most strategic use of agentic AI comes not from general “know-it-all” models but from vertical applications that go deep in one field and have internalised its data structures, terminology and workflows. In sectors such as law, health, finance, logistics and agriculture, building “sector-specific agents” offers a serious competitive advantage and a sustainable business model.
We can assess the concrete returns of this approach under these headings:
Historically, technology companies needed large engineering teams to develop products. That picture is changing radically. Agentic AI tools allow small teams to run far larger operations. A three-person software team, with a product built on the right agent infrastructure, can now handle work that once took a fifty-person department.
This transformation opens a critical window for start-ups with limited capital and for mid-sized companies. The sluggishness of large organisations, the slow decision-making of big teams and corporate bureaucracy leave a serious opportunity gap in front of agile structures. At this point, where competition is being redefined, the advantage lies not in scale but in speed and, above all, expertise.
So what should practical strategy look like against this big picture? GPT-4o, Claude 3.5, Llama 3 or Gemini Ultra... These engines already exist and are accessible. Rather than building them from scratch, developing smart tools on top of them that solve your own sector's problems is both more efficient and more sustainable. The real competitive advantage lies not in how big the model is, but in which problem it solves, how, and how fast.
General-purpose solutions that “do everything” are quickly becoming commodities in the AI market, and margins are being squeezed. Vertical solutions that understand a specific sector's processes in depth and are fine-tuned with sector-specific data, by contrast, remain a field with relatively little competition. A contract analysis agent for the legal sector, a risk assessment system for finance or an income optimisation tool for agriculture show where the real opportunities lie.
At this point we need to touch on an issue the AI ecosystem leaves in the shadows: a large share of the datasets used to train foundation models is fed by copyrighted content. Journalists, writers, visual artists and musicians are raising their voices about this ever more loudly. Ongoing lawsuits and legislative debates in various countries suggest the field will enter an even more complex legal environment in the near term.
The phrase “it steals, but it works” captures the essence of the dilemma: the current AI infrastructure clearly creates enormous efficiency and value. But the ethical and legal legitimacy of the data foundation on which that value is built is still disputed. Transparency of data sources and the use of licensed content will become both a serious competitive advantage and a legal requirement in the period ahead.
In the face of this historic transformation in the AI ecosystem, two basic stances are possible: enter the giants' arena and play their game, or build agile, value-adding solutions outside that arena, in your own field of expertise.
The data point to the second path. The foundation model race is now played in a league whose rules are set by a few giant companies. But the field of building autonomous systems that use these models to solve real-world problems, backed by sector expertise, is still open, still uncontested and largely unexplored. The question is: which side will you stand on?
Stay in light and love.
