In the competitive GPU cloud infrastructure space, one name is making waves with a model that’s impossible not to notice: Nebius AI. U.S. providers such as AWS, Azure, and upstart CoreWeave get most of the attention, but Nebius is making its way by delivering equal compute for a fraction of the price. New figures show Nebius underpriced its competitor, CoreWeave, by up to 34% on extended Nvidia H100 GPU rentals. It’s not a marketing trick, but a genuine structural benefit making people rethink the cost-per-FLOP.
In perspective, CoreWeave will charge up to $49 per hour for an H100 HGX instance, based on commitment. Nebius, by comparison, will charge the same computer for just $3.15 per hour for a 12-month commitment. It works out over hundreds of thousands of GPU-hours of compute to tens of millions of dollars in savings for businesses training machine learning models.
Suppose you’re a startup looking to maximize your Series A runway or a research organization with limited fixed grant funds. Getting one-third off the same infrastructure is critical. That’s the difference that matters most in machine learning: cost-per-FLOP, and Nebius seems to have attacked each layer of its stack with ruthless cuts.
The outcome is easy enough to understand: with Nebius, you get the same AI work that you do on CoreWeave but for 30%-40% less in real money over time. That isn’t competitive; that’s disruptive.
Nebius Provides Hyperscaler Efficiency Without Hyperscaler Overhead
What enables Nebius to achieve such a cost leadership? It boils down to a unique combination of deep vertical integration, regional arbitrage, and a hyperscaler-grade infrastructure playbook with none of the bloated operational overhead.
While most competitors rely on third-party data centers, off-the-shelf OEM gear, and public GPU markets, Nebius has taken the opposite approach, prioritizing control, customization, and scale. Its infrastructure remains in-house and includes Eastern European data centers, custom liquid-cooled server racks, and custom firmware stacks optimized to wring out every last ounce of Nvidia silicon.
This is not some bare-bones lean startup. In fact, Nebius has invested nearly $2 billion in capital spending over the past three years building out a full-stack platform from scratch. But while Big Tech’s global cloud arms optimize everything for maximal generality, Nebius optimizes tightly for particular markets, applications, and power-to-performance ratios.
Using third-party analysis from PoolCompute, the infrastructure cost of running one H100 on Nebius infrastructure, including power, cooling, and support, can fall as low as $0.025 per GPU-hour. Compare that with the much greater infrastructure overhead in North America, and the competitive difference becomes apparent.
Most importantly, Nebius sidesteps distorted U.S. GPU rental market pricing. CoreWeave, Crusoe, and Lambda typically source their GPUs from inflated secondary markets or broker chain super spans that maximize CapEx and leasing expenses. Nebius, on the other hand, employs a direct ODM-style procurement model that enables them to skip over retailer markups, buy hardware for wholesale- or gray-market-like costs, and achieve efficient rack density planning with high utilization.
Nebius is not selling at a loss; rather, its more efficient structural model naturally allows it to offer lower prices while maintaining healthy margins.
How Nebius Runs H100s for Pennies
While Nebius makes available long-term H100 GPU rentals for $3.15/hr, vs. $4.76/hr from CoreWeave, their GPU-hour operational cost is incredibly low, as little as $0.025/hr. This is due to vertically integrated data centers, liquid cooling, and the ability to access renewable energy that costs $0.04/kWh or less.
Nebius also leverages in-house server designs and ODM production, avoiding inflated OEM and distributor markups prevalent in the U.S. supply chain.
In comparison, CoreWeave’s operations base themselves on leased colocation space in expensive markets such as the U.S., where between $0.12–$0.15/kWh for electricity isn’t uncommon. Including real estate, personnel, electricity, and facilities support, CoreWeave’s OPEX per H100 GPU-hour appears on the order of $0.45–$0.65/hr, approximately 18x–26x higher than Nebius.
This difference in costs matters: despite 34% cheaper top-of-the-funnel pricing, Nebius manages a gross margin of 90%+ on inference workloads, but the tighter spreads for CoreWeave lead them to have to do so just to be competitive. The upshot is a structural lead that allows Nebius to scale profitably in downmarket environments before factoring in anticipated CapEx leverage and utilization gains.
Sovereignty as Strategy: Nebius’ Geopolitical Masterstroke
Another reason why Nebius is able to provide such competitive pricing is its unorthodox geopolitical strategy. The majority of GPU cloud providers are based in Western Europe and North America, where strict regulations, energy costs, and compliance challenges add significant overhead.
Nebius, on the other hand, operates in low cost regions and benefits from lower energy costs, tax breaks, and data-sovereignty incentives in order to achieve cost benefits and attract customers outside U.S. and EU jurisdictions.
This gets into pure market strategy. As the global environment becomes increasingly fragmented along geopolitical lines, governments and businesses alike are looking for sovereign AI infrastructure; systems that aren’t bound by U.S. laws such as the CLOUD Act, or held hostage by possible geopolitical blowback. Nebius capitalizes on this demand by providing services that are sovereign by nature, based in non-aligned jurisdictions with minimal regulatory overhead.
What makes Nebius particularly interesting is that its offerings are localizable. Its parent organization also runs Toloka, a human-in-the-loop platform for annotating data, and TripleTen, an artificial intelligence workforce platform, among others.
These partnerships help Nebius build an ecosystem that not only serves infrastructure buyers, but also developers and researchers looking to build and deploy models rapidly. By doing so, Nebius isn’t merely competing on price, it’s investing in strategic moats of integration, sovereignty, and developer experience.
Forward-looking perspective
No model is infallible, and Nebius isn’t exempt from real challenges. The company operates in an industry that requires capital expenditure and has probably spent more than any other non-hyperscaler on infrastructure outside of the U.S. That level of expenditure requires customers with deep, long-term commitment, consistent demand growth, and the capacity for mitigating geopolitical risks.
In the event that supply chains are disrupted, if GPU export restrictions tighten, or if political unrest interrupts operations in priority regions, Nebius’ model of aggression could turn on the company.
There’s also the question of margins. PoolCompute estimates that Nebius has only ~2% gross margins on hardware, a calculated decision in the name of winning market share, but one that leaves little buffer in the event that conditions change.
While AWS and Google Cloud have margin flexibility from a diversification of services, Nebius is forecasting that volume and operational scale will be enough to transform slim margins into lasting profit. It’s a risky gamble, but one that’s already beginning to pay off.
With that said, Nebius doesn’t need to become the next AWS. It simply needs to secure the alternate route, the one that provides developers, startups, and institutions outside of the West with a less expensive, sovereign, and high-performance GPU cloud that Big Tech cannot replicate without cannibalizing its own economics.
With an increasingly bifurcated global economy, where AI infrastructure is as much about nationalism as engineering specifications, Nebius is quietly positioning itself as the platform for the rest of the world.
Conclusion
In an industry that’s racing towards supercomputing dominance through artificial intelligence, Nebius isn’t competing with the giants on their terms. Instead, it’s updating the playbook by meshing hyperscaler economies of scale with startup nimbleness, geopolitical acuteness, and a business model that can’t be matched.
The bottom line? Nebius won’t be the loudest mouth in the room, but one of the most vital voices.
