roman chernin

Nebius is building a vertically integrated AI cloud across four layers — bare metal, managed cloud, managed inference (Token Factory), and agentic execution — to capture a diversified customer base from hyperscalers to enterprises, reducing concentration risk and expanding TAM.

Google's Gemini models are part of the frontier model tier that developers use for initial product development before potentially optimizing with open-source alternatives.

External validation from a respected investor (5.3% stake, 15% of portfolio) reinforces Nebius's strategy, but management stays focused on daily execution in a shark-like market where stopping means dying.

European generative AI companies like Black Forest Labs are key to building a self-sustaining regional AI stack by creating compute demand.

Large digital enterprises like Booking.com will follow the same pattern: slow initial AI adoption while building eval/CI-CD foundations, then exponential compute consumption growth.

Cloud-native enterprises like Shopify will see exponential AI compute growth once they build the evaluation and deployment infrastructure to safely iterate on models in production.

European model builders like Mistral generate local AI demand that justifies sovereign infrastructure investment; infrastructure follows builder ecosystems, not the reverse.

Enterprises face a cold-start problem building eval/CI-CD pipelines for AI; once solved, their AI compute consumption grows exponentially on par with AI-native companies.

Cheaper intelligence does not reduce compute demand — it expands the set of economically viable use cases, driving higher inference consumption. DeepSeek's release proved this: Nebius shares dropped 40% but the company recorded its best sales week as customers realized they could run production inference profitably.

Cursor was among the first to successfully fine-tune open-source models for coding, demonstrating the path from frontier models to specialized, cost-efficient inference at scale.