Marc Boroditsky on Nebius closing a $2B+ deal with Microsoft for AI cloud infrastructure
Sep 11, 2025 with Marc Boroditsky
Key Points
- Nebius signs a five-year Microsoft deal valued at $17.4 billion minimum, with potential upside to $19.4 billion, positioning the AI infrastructure company among the largest cloud deals of its kind.
- Nebius builds a full-stack AI platform spanning cluster management, model access, training, and inference delivery, aiming to become the AWS equivalent for AI customers from solo engineers to enterprise scale.
- The company expects long-term value to shift from current training workloads to production inference, where AI applications generate revenue from paying end customers in a $100 billion-plus addressable market.
Summary
Nebius has signed a five-year AI infrastructure agreement with Microsoft valued at a minimum of $17.4 billion, with potential upside to $19.4 billion. Marc Boroditsky, Nebius's chief revenue officer, who joined the company roughly three months ago, confirmed the deal is one of the largest of its kind in AI infrastructure.
The Microsoft contract is framed internally as a 'super lab deal' — the kind Nebius CEO Arotti had telegraphed as part of the long-term plan. But Boroditsky is clear that the bigger prize is becoming the AWS equivalent for AI: a full-stack cloud provider that serves everyone from a solo engineer signing up via self-service to Cloudflare and Shopify-scale enterprises.
What Nebius actually sells
Nebius positions itself as a full-spectrum AI infrastructure platform. That means cluster management tooling, open-source model access, training and retraining capabilities, and inference delivery through its AI Studio product — all aimed at the AI engineer. SemiAnalysis awarded Nebius a gold ranking in its GPU cloud cluster benchmarking, which Boroditsky cites as third-party validation of the platform's technical depth.
The company's current customer base runs from early-stage AI startups — including Higgsfield, which appeared on the same episode — up through larger enterprises. Boroditsky describes the business model as following the customer lifecycle: support early development and training workloads, then scale with them as applications move into production inference.
Training now, inference next
Boroditsky distinguishes two phases of the market. Right now, significant construction and model training activity is underway, and Nebius is chasing that demand. But the long-term value, he argues, accrues when those projects convert to production inference, because inference revenue signals that AI applications are actually being used by paying end customers. He says that shift is already visible in Nebius's revenue mix.
The long-term commercial target is a $100 billion-plus market opportunity, servicing the hyperscaler-equivalent requirements of the new class of AI customer. Getting there means building out both the physical infrastructure and the software stack simultaneously — Boroditsky says Nebius intends to do both.