Expert Interview | Jimsher Chelidze: "We Will Not Catch Up in the Race for Large Models — and We Should Not. Industrial AI Is Won Not in the Data Center, but on the Shop Floor"

The topic of industrial AI is currently one of the most discussed, yet complex to grasp. In January 2026, Vladimir Putin approved a list of directives on AI development, which included forming a comprehensive suite of Russian technological solutions, integrating AI across economic sectors and public administration, and developing the export of AI technologies. By April, he directed that AI tools be integrated into all economic sectors—from industry to medicine—by 2030. In July 2026, the State Duma passed a foundational law supporting AI development, introducing the concepts of "sovereign" and "national" models and establishing the legal framework for priority support of domestic developments.

Over the past few years, Russia has launched hundreds of projects and spent tens of billions of rubles, yet there is still no systemic answer to what actually works. Companies allocate resources to pilots that fail to scale. The state subsidizes developments that never become products. Engineers fear being replaced by algorithms—yet algorithms still cannot handle the tasks that humans solve.

We asked Jimsher Chelidze*, an expert in industrial technologies, to address the pressing questions currently on the minds of owners and directors of industrial companies. In this extensive expert interview, "Industrial AI: Export, Engineers, Barriers, and the Role of the State," which we are publishing as a six-part series, he explains why Russian AI has not yet become an export product, how the role of the engineer is changing, why most projects fail, what to do with data, people, and security, and where the state helps versus where it creates barriers.

The answers are published in full, without abridgment.

*Jimsher Chelidze is the CEO of Chelidze & Partners LLC, a business partner for digital development at the Horizontal Drilling Center LLC, and a member of the Industrial Committee of the EAEU Business Association. He is the author of articles and publications on the philosophy of technology, five books on digital transformation, and two AI products. He has practical experience working with Gazprom Neft, LUKOIL, the Russian Ministry of Energy, Gazprom Bureniye, and other industrial companies in Russia, Kazakhstan, and China.

Many industrial AI solutions are being developed with a focus on import substitution. Can Russian industrial AI become an export product and compete with global leaders—or are we destined to remain in a catch-up development paradigm?

This is a false dichotomy, and it is costing us dearly. "Catching up" or "overtaking" only makes sense if you are running on the same track. But we are running on different tracks.

Let us divide the AI market into three layers, as they are constantly confused, and this confusion breeds both inflated expectations and unwarranted pessimism.

The first layer is cutting-edge foundation models. This is a game of capital and energy. Current industry estimates indicate a gap of hundreds of times between the US and Russia in available computing power, with only about ten thousand GPUs across all Russian data centers dedicated to AI. Moreover, the main bottleneck is no longer even the hardware, but connected electrical capacity: it takes a year for grid connection, and there is simply no spare capacity left. Competing "head-on" here is not pessimism or alarmism; it is arithmetic. And this is normal: Germany does not have its own frontier model either, yet no one considers German mechanical engineering to be backward.

The second layer comprises applied platforms and toolsets: MLOps, vector databases, agent orchestration, and fine-tuning tools. Competition here is global, and it has largely already been won by open-source software—by everyone, all at once. Building a national advantage here is pointless.

The third layer is domain-specific vertical solutions. Real-time management of mining haul trucks. Optimization of gas consumption in chemical production. Defect detection via machine vision. Predictive maintenance of turbines. A technologist's assistant that knows your specific machine park and tooling. Here, the winner is not the one with the most GPUs. The winner is the one who has access to real production data, an industry-specific engineering school, and the right to experiment on an operating facility.

And here is the major shift of the last two years, which few openly discuss: open models have nullified a significant part of the barrier to entry for the base model. The model is no longer a competitive advantage. The advantage has shifted to data (its availability, which stems from industrial automation and sensors), integration with the production loop, and the ability to prove the effect. This is good news for us. Because we have blast furnaces, open-pit mines, cracking units, and power blocks, while a startup from Silicon Valley does not.

Does this mean export will "just happen"? No. And here I will be unpleasant.

Between "the solution works" and "the solution is sold abroad" lie three gaps.

The first gap is turning developments into scalable products. Import substitution, by its very design, produced solutions "for internal use": the customer, developer, and consumer were often the same entity. There is a figure that explains it all. Through the industrial competence centers, total project costs for 2022–2025 amounted to roughly 187 billion rubles, while developers' revenue from the implementation and scaling of the created products was about 1.6 billion rubles. The difference is over a hundredfold.

This is not an accusation. It is a diagnosis of the business model. We financed development—but we did not finance productization: documentation, localization, implementation methodology, partner networks, and second- and third-line support. Yet export begins exactly here, not with the code.

The second gap is proving the effect. In the external market, no one cares about your unique neural network. They care about four questions: where has this been working in industrial operation for more than two years; what is the measured effect relative to the baseline; who bears responsibility in case of failure; and is there support in the local language within the local jurisdiction? And here we run into what I call one of the deadly sins of digitalization—the lack of a project management system, including the measurement of effects. The "before" baseline is most often simply non-existent. So there is nothing to show. This gap hurts export more than any sanctions, because it is self-inflicted.

The third gap is channels. Exporting industrial software is not a delivery; it is a presence: a local partner, an integrator, a trained local team, a reference in the local market. This is a three-to-five-year horizon and money that does not pay off on the first contract. We currently do not have a single corporate program where this is budgeted as a norm, rather than the enthusiasm of an individual vice president.

So where is the real window?

It exists, and it is specific. Not "Russian AI in general," but verticals where we have our own engineering school and operating facilities: mining, metallurgy, oil and gas and petrochemicals, energy, and transport. And markets where independence from the supplier's geopolitics matters more than price: the EAEU, the Middle East, Southeast Asia, Africa, and Latin America.

Our export argument is not "we are cheaper." Chinese developers will be cheaper. Our argument goes like this: "We will deploy in a closed loop, grant full access, train your team, and we will not shut you down tomorrow due to a political decision." If you think about it, this is a strong position—exactly the one a Western vendor cannot take by definition. And it resonates particularly well in markets that have watched in recent years as licenses were unilaterally revoked.

What needs to happen to keep the window from closing?

I always break down the solution into three pillars: management, technology, and people. Failure in any one of them nullifies the other two.

Management.

● Change the success metric. Stop measuring import substitution by the mere fact of substitution. Start measuring the share of revenue from scaling and external sales. These are different metrics—and they drive fundamentally different team behaviors.

● Separate "an internal IT project of a corporation" from "a product." These are different economics, different competencies, different people.

● Budget for export as a three-to-five-year program, not as an enthusiast's initiative. Export that does not have a line item in the budget does not exist.

Technology.

● Design the architecture for scaling, not for internal use. Configurability instead of customization for each client. Multi-tenancy. A separated localization layer—language, units of measurement, industry directories, local regulations. Clear documentation and interfaces so that implementation can be done by a partner, not just the authoring team. A solution that can only be implemented by those who wrote it cannot be exported—it does not scale even within the country.

● Build the evidence base into the product itself. Measuring the effect must be a function of the system, not a feat of an analyst. Otherwise, in two years, you will not be able to answer the buyer's main question: "show me the effect."

People.

● Product role and implementation methodology. A plant does not need a product manager. The product does. As does an implementation methodology that can be handed over to other hands.

● Local presence in the target market. A partner, a trained local team, the first reference. Selling industrial software "from Moscow via email" is impossible in any market in the world.

And one more thing that needs to be said directly. Import substitution is essentially a "build" decision made at the country level once, in 2022, and not revised since. But the answer to the "build or buy" question has an expiration date. Over these years, open models have nullified half the barriers, and some of what we stubbornly build from scratch today would be more sensible to adapt. This is not capitulation—it is a normal reassessment that any mature portfolio undergoes.

Bottom line: we will not catch up in the race for frontier models, and that is sensible. We can win in industrial AI, where technology is 20%, and 80% is data, processes, and engineering. But as long as the cost-to-scaling-revenue ratio is what it is, talking about export is premature. Not because our technologies are bad. But because we do not yet have an export-oriented business model.

For the financial sector, there is a specific projection here. The financial sector is a rare exception: there, internal development has indeed become a product. But this is ensured by the scale of the balance sheets of a few players, not by the system. That is exactly why banks should look closely at the industrial case: the same fork in the road lies ahead—remain an internal development or become a product.