Import Substitution Built Solutions, Not Product Companies — and That Decides the Fate of Russian AI Exports

Russia's post-2022 software localization drive produced working systems. It did not produce product businesses. Here is why that matters — for anyone watching Russian industrial AI, and for potential partners in the EAEU, the Middle East, and Southeast Asia.

The false dichotomy

"Is Russia catching up or pulling ahead?" The question only makes sense on a single track. We are running on different ones. Until we admit that, the export debate runs in circles. One camp proves we are hopelessly behind. The other promises a breakthrough any day now. Both are wrong. Both make the same mistake. They treat "the AI market" as one thing. It is not. It is three layers with fundamentally different economics.

The three layers everyone confuses

Layer one: frontier foundation models. This is a game of capital and energy. Industry estimates put the US–Russia compute gap at hundreds of times. Roughly ten thousand accelerators serve AI across all Russian data centers. Parallel-import shipments fell from thousands of servers in 2024 to dozens. And hardware is no longer even the main constraint. Grid power is. A year to get connected — and no spare capacity anyway. Competing head-on here is not pessimism. It is arithmetic. And it is perfectly normal. Germany has no frontier model either. Nobody calls German machine-building backward.

Layer two: applied platforms and tooling. MLOps, vector databases, agent orchestration, fine-tuning tools. Competition here is global. And it is largely lost already — not by us, by everyone. Lost to open source. Building a national advantage here is pointless. You would be funding something that becomes free in six months.

Layer three: vertical domain solutions. Real-time dispatch of mining haul trucks. Gas consumption optimization in chemical plants. Machine-vision defect detection. Predictive maintenance for turbines. A process assistant that knows your machine park and your tooling. Compute does not win here. What wins:

• access to real production data;

• a domain engineering school;

• the right to experiment on a live industrial site.

The big shift: the model is no longer the moat

Few people say this out loud. Open models have erased much of the entry barrier at the base-model level. The advantage has moved. It is no longer in the model. It is in data — in its very existence, built up from industrial automation and sensors. In integration with the production loop. In the ability to prove impact. For us, that is good news. We have the blast furnaces, open pits, cracking units, and power blocks. A Silicon Valley startup does not. Does that mean exports will happen by themselves? No. And here I will be unpleasant.

Three gaps between "it works" and "it sells"

Gap one: turning in-house builds into scalable products. Import substitution — Russia's post-2022 software localization program — was designed to build solutions "for ourselves." Customer, developer, and user were often the same corporate loop. One number explains everything. Through the Industrial Competence Centers — a state-funded industry software program — project spending in 2022–2025 totaled about 187 billion rubles (~$2 billion). Developer revenue from selling and replicating those products: about 1.6 billion rubles (~$17 million). A gap of more than a hundred times. That is not an accusation. It is a diagnosis of the business model. We funded development. We did not fund productization: documentation, localization, implementation methodology, a partner network, second- and third-line support. And export starts exactly there. Not with code.

Gap two: proof of impact. Foreign buyers do not care about your unique neural network. They care about four questions. Where has this run in production for over two years? What is the measured effect against a baseline? Who is liable when it fails? Is there support in the local language, in the local jurisdiction? And here we hit what I call one of the deadly sins of digitalization: no project management discipline, including no measurement of effects. The "before" baseline usually does not exist. So there is nothing to show. This gap hurts exports more than any sanctions. Because it is self-inflicted.

Gap three: channels. Exporting industrial software is not a shipment. It is presence: a local partner, an integrator, a trained local team, a local reference. It is a three-to-five-year horizon and money that does not pay back on the first contract. Not one Russian corporate program budgets this as a norm today. At best it runs on one vice president's enthusiasm.

Where the real window is

It exists. And it is specific. Not "Russian AI in general." Verticals where we have our own engineering school and live industrial sites: mining, metallurgy, oil and gas, petrochemicals, energy, transport. And markets where the vendor's independence from geopolitics matters more than price: the EAEU, the Middle East, Southeast Asia, Africa, Latin America. Our export pitch is not "we are cheaper." Chinese developers will always be cheaper. Our pitch is this: "We will deploy inside your closed network, hand over full access, train your team — and we will not switch you off tomorrow by political decision." Think about it. That is a strong position. Precisely the one a Western vendor cannot take, by definition. And it reads especially well in markets that have spent recent years watching licenses get revoked unilaterally.

What has to happen: three pillars

I always break a solution into three pillars — management, technology, people. A failure in any one of them zeroes out the other two.

Management

1. Change the success metric. Stop measuring import substitution by the fact of replacement. Start measuring the share of revenue from replication and external sales. These are different metrics. They produce fundamentally different team behavior.

2. Separate the "internal corporate IT project" from the "product." Different economics. Different competencies. Different people.

3. Put export into the budget as a three-to-five-year program. An export effort without a budget line does not exist.

Technology

4. Design the architecture for replication, not for yourself. That means:

• configurability instead of custom builds for every client;

• multi-tenancy;

• a separate localization layer — language, units of measure, industry reference data, local standards;

• documentation and APIs clear enough for a partner to implement, not just the original team.

A solution only its authors can deploy cannot be exported. It does not even scale at home.

5. Build the evidence base into the product itself. Measuring impact must be a system function, not an analyst's act of heroism. Otherwise, two years in, you cannot answer the buyer's main question: "show me the effect."

People

6. A product role and a transferable implementation methodology. A plant does not need a product manager. A product cannot live without one. Same goes for an implementation methodology you can hand to someone else's team.

7. Local presence in the target market. A partner, a trained local team, a first reference. Selling industrial software "from Moscow, by email" works in no market on Earth.

"Build or buy" has an expiration date

Import substitution is, at its core, a "build" decision. Made once, at country level, in 2022. Never revisited since. But the answer to "build or buy" has a shelf life. Open models have since erased half the barriers. Part of what we stubbornly build from scratch is now smarter to adapt. That is not capitulation. It is the normal reappraisal any mature portfolio goes through. Tellingly, the AI development support law passed in July 2026 accepts this logic: a model assembled from openly licensed components, fine-tuned by a Russian legal entity, and deployed in a Russian data center receives "national" status. The legislator saw what the market had understood earlier.

The bottom line

We will not chase the frontier-model race. That is sensible. We can win in industrial AI, where technology is 20% and data, processes, and engineering are 80%. But while the ratio of spending to replication revenue stays where it is, talk of export is premature. Not because our technology is bad. Because our business model is not an export model yet. The good news: a business model is exactly the thing you can change by decision. Unlike gigawatts and accelerators.

P.S. For the financial sector. Finance is a rare exception: there, in-house development really did become a product. But that was achieved by the balance-sheet scale of a few players, not by a system. Which is exactly why banks should study the industrial case closely. The same fork lies ahead: stay an internal build — or become a product.