In the United States, the immediate constraints are the grid, interconnection timelines, generation capacity, grid and power equipment, suitable sites, permitting, copper and the cost of capital. In Russia, the bottleneck is more directly tied to access to the computing base itself – GPUs, AI-ready data centres, high-density racks, maintenance and support, and imports constrained by sanctions. This is why the issue should be understood first and foremost as an industrial one.
For Washington, the task is essentially to accelerate the build-out of infrastructure needed to sustain a private investment cycle. The government is not merely removing barriers; it is also reshaping permitting and grid-connection procedures to bring large new loads online faster, while questions of reliability and electricity prices remain. The beneficiaries are not limited to model developers. They include the entire heavy industrial ecosystem surrounding them – from manufacturers of power equipment to suppliers of construction machinery and defence-technology contractors. Judging by order books and market valuations, the AI boom is already acting as a multiplier for companies such as Caterpillar, Palantir and SpaceX, even though each occupies a very different position in this ecosystem. What we are seeing, in other words, is an industrial mobilisation of capital rather than an immediate political constraint on decision-making.
There is, of course, a counterargument. Such a sharp increase in electricity demand could eventually become a political issue through higher tariffs, environmental constraints or disputes over where new capacity should be located. For now, however, that does not appear to be the dominant dynamic. Even the rise in hydrocarbon prices in some jurisdictions reported by Bloomberg primarily reflects the market adapting to a new demand profile rather than a systemic failure. Moreover, federal and state regulators in the United States appear, at least for the time being, to regard AI infrastructure as desirable industrial demand that needs to be integrated into the power system rather than as a threat requiring political restraint. The distinction matters because it separates structural overload from an investment cycle that remains broadly manageable.
The problems are nevertheless real. Copper is perhaps the clearest example: without it, the expansion of the largest data-centre segment will proceed much more slowly. Industry estimates suggest that a major AI data centre can require as much as 50,000 tonnes of copper, while new mining projects in the United States take years to bring into production. Delays stem not only from raw-material constraints, but also from grid infrastructure, permitting and the cost of capital. Even here, however, the issue is primarily the pace and cost of expansion rather than political risk in the narrow sense. For markets, that distinction is consequential: the sectors most likely to be repriced are utilities, construction and raw-material supply chains, not the partisan balance in Washington.
The energy and infrastructure burden that the United States is addressing through grid expansion, investment and copper supply runs into a constraint of a different nature in Russia – access to computing capacity itself. If Washington's bottlenecks are electricity prices, permitting and an ageing grid, Moscow's are embedded in the sanctions regime: Russian data centres and supercomputers face restrictions on access to advanced Nvidia and AMD chips. The underlying disease is the same – AI's hunger for energy and computing power – but it affects different organs of the system.
The scale of the gap is sobering. Across Russia's commercial data centres, more than 10,000 GPUs are currently deployed in A100-equivalent terms, with roughly another 8,000 in companies' on-premises infrastructure. Industry estimates put the combined 2026 hardware budgets of Russia's principal players – Sber, Yandex, VK, MTS AI and T-Technologies – at around $1.5–2.5 billion. That compares with hundreds of billions of dollars in the United States and close to $100 billion in China, where most of the funding is state-backed.
More troubling still, the workaround is drying up. In a downside scenario, parallel imports could fall from thousands of servers in 2024 to only dozens in 2026, as China absorbs the same equipment and is prepared to pay several times the original price. Sber's supercomputers have meanwhile dropped out of the global top 100, while the bank could invest at least RUB 500 billion in a third machine. Yandex's Chervonenkis still ranked 83rd as of November 2025. In March, the Russian government approved a roadmap covering high-performance computing, AI algorithms, grid computing technologies and supercomputing infrastructure. In short, Russia is building actively – but not always where it matters most, and not fast enough.
Scenarios
Managed catch-up – the most likely scenario. Russia continues to expand capacity through state programmes, Sber's investment and the development of domestic chips, but does so through a narrowing import channel and while remaining a generation behind. Applied AI and Russian-language models remain viable, but the next major qualitative leap is pushed further out. The key indicators would be the commissioning of new supercomputers and the first domestically produced accelerators entering volume production, alongside a continued decline in parallel imports. The result would be that Russia remains competitive in language models and specialised industry applications, but not in frontier-scale training.
A widening gap – moderate probability. If access to advanced chips closes faster than Russia can develop its own manufacturing base, computing capacity will become an increasingly binding constraint on the entire sector. The clearest warning sign would be delays in the domestic component base combined with continued reliance on an ageing stock of accelerators acquired through indirect import channels. The result would be that infrastructure moves to the centre of the AI agenda, displacing the current emphasis on integrating existing models.
This leads to a shift in business priorities that is worth stating explicitly. For Russian companies, the central issue is no longer simply how to integrate existing models into production and technology processes. Increasingly, it is access to the underlying infrastructure itself – computing capacity, data centres and power. Integrating someone else's model can deliver a tactical gain. A genuine step change in both capability and economics, however, ultimately requires control over the computing base itself.