Nebius CTO Danila Shtan says the AI cloud market grew from 5-7 to 300-500 companies, yet Nebius targets AWS and Google Cloud with revenue near 1% of AWS.
Shtan predicts AI demand will grow tenfold to a hundredfold in three years, leaving room for independent mid-tier providers.
Nebius underestimated storage, with customers hitting 500-700 GB/s on one shared volume.
Nebius bets the hyperscalers will not bother with AI cloud
The AI cloud is a business the hyperscalers have not committed to. For them, renting out GPUs is a $2 to $4bn market set against franchises many times larger, and entering it means rebuilding infrastructure designed for other workloads, an investment they cannot yet justify. Nebius is betting that window stays open long enough for an independent provider to hold the mid-tier, and that the contest is decided by execution and engineering rather than capital.
Takeaways
Hyperscalers leave room for an independent AI cloud. AI cloud is a $2 to $4bn business for the hyperscalers, small beside their core franchises. Google, Meta and Microsoft juggle capacity between internal teams and the market, and OpenAI and Anthropic consume their own. Nebius's revenue is about 1% of AWS's.
Demand grows tenfold while hundreds of clouds fight. Shtan predicts overall AI demand will grow tenfold, maybe a hundredfold, in three years. He counts 300 to 500 neoclouds today, up from five to seven when Nebius started, a term he rejects, and expects consolidation with a long tail surviving.
Storage was the underestimated part of the stack. Nebius customers routinely pull 500 to 700 gigabytes per second from one shared volume. One demanding case needs 700 GB/s from three petabytes spread across 1,000 servers, or 8,000 GPUs. Four years ago network attached storage meant hundreds of megabytes.
Execution is the binding constraint on the build-out. GPUs, CPUs, NVMe, facilities, electricity, workforce and network transceivers are all short, and transformer lead times run to years. Shtan calls experienced project managers the scarcest resource he has seen. Nebius controls its own build-out and has announced an asset-light model.
Nebius is software-first on a Yandex hardware team. Nebius finalized its separation from Yandex in 2024. The hardware team, 40 to 50 people that year, had operated hundreds of thousands to millions of servers, while software was already 200 and the platform was rewritten from scratch. It sells no managed bare metal.
Inference optimization is still mostly talk. Everyone discusses disaggregating prefill and decode and nobody does it properly, so Nebius runs it only for its most demanding customers. Most of its customers, 95%, run modified or fine-tuned models, and pruning and compression remain underexplored. An acquisition doubled the model-optimization team.
What comes next
OpenAI and Anthropic are building forward-deployed organizations and will take on work traditionally done by consultancies like Deloitte and Accenture. Both want to be a single touch point for everything, and keeping users inside their ecosystem lets them steer and upsell.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Shtan expects a couple of breakthroughs in inference optimization for scoped, specialized models, and he sees no novel architecture today he would bet on, though he hopes something is cooking in a stealth lab.