Nebius

Executive Summary

Nebius (NASDAQ: NBIS) is the Amsterdam-headquartered neocloud — formerly Yandex N.V., which divested its Russian assets and relisted in 2024 — that has layered a managed inference business, Token Factory, on top of its GPU cloud. Token Factory serves 60+ open-source models (DeepSeek-V4-Pro, Qwen3-235B, Llama, GPT OSS, GLM-5.1, MiniMax-M2.5) through per-token APIs, dedicated endpoints with a 99.9% uptime SLA, and fine-tuning pipelines.^[1] Management called Token Factory the company's fastest-growing segment on the Q1 2026 earnings call, naming Revolut and 1X Technologies as recent customers.^[2][3]

The scale behind it is unusual for an inference vendor: Q1 2026 revenue of $399M grew 684% year-over-year, AI ARR reached $1.9B (up from $1.25B at end-2025), and NVIDIA made a $2B strategic equity investment on March 11, 2026, in a partnership targeting more than 5 gigawatts of NVIDIA systems deployed by end of 2030.^[2][4] Nebius then spent ~$643M to acquire inference-optimization firm Eigen AI (May 1, 2026) and folded in Clarifai and Tavily, assembling a full inference and agentic-deployment stack on top of a $27B five-year Meta infrastructure contract.^[5][6][2]

---

Product Overview

Token Factory is the managed-inference layer of the Nebius cloud: pick an open model, hit an API with transparent $/token pricing, and scale to dedicated endpoints or fine-tuned variants without managing GPUs.^[1] It launched as the production-grade successor to Nebius AI Studio, pitched at companies deploying open-source and custom models with enterprise security, and counts Higgsfield AI as an adopter and Hugging Face as a collaborator for open-model access.^[8]

Two serving configurations cover the latency/cost trade: Fast (sub-second responses for interactive agents and chat) and Base (cost-efficient throughput for large-scale or background processing), switchable on the same API and endpoints.^[1]

Key Capabilities

Product Surfaces

---

Technical Architecture

Unlike inference-first startups, Token Factory sits on infrastructure Nebius owns end to end — the NVIDIA partnership spans "AI factory architecture to production software" and targets 5+ GW of NVIDIA systems by end of 2030, with the $2B investment cited by management as supply-chain certainty for GPU allocation.^[4][2] The serving layer uses autoscaling, speculative decoding, and multi-region routing to hold sub-second latency targets at scale.^[8]

The Eigen AI acquisition is the architecture bet: a 20-person team of MIT HAN Lab alumni whose quantization, KV-cache, and custom CUDA-kernel stack reduces compute and memory per token, integrated directly into Token Factory's inference and post-training layers, with the founding team establishing a Nebius engineering presence in the San Francisco Bay Area.^[6][5]

Key Technical Details

---

Strengths

• Vertically integrated economics — Nebius owns the data centers, the power contracts, and the serving layer, so per-token margins are not squeezed by a cloud landlord; the $20–25B capex program is the moat competitors must rent.^[2]
• NVIDIA as investor, not just supplier — the $2B strategic investment and full-stack partnership (March 11, 2026) give allocation certainty on next-generation systems through 2030.^[4]
• Hyperscale-grade financial visibility — $399M Q1 revenue (+684%), $1.9B AI ARR, positive $130M adjusted EBITDA, and a $27B Meta contract make this the rare inference vendor with audited, public financials.^[2]
• Acquired optimization depth — paying $643M for 20-person Eigen AI signals inference efficiency is the strategic priority; Clarifai and Tavily round out multimodal inference and agentic search.^[6][2]
• Enterprise compliance posture — SOC 2 Type II, HIPAA, and ISO 27001 with EU data residency, a differentiator for European buyers that US-only rivals lack.^[1]

---

Cautions

• Inference is the side bet, not the business — Token Factory is the fastest-growing segment, but Nebius's revenue is dominated by raw AI infrastructure; per-token serving could be deprioritized if hyperscale leases keep outgrowing it.^[2]
• Customer-concentration risk — a $27B Meta contract anchors the model; the company's fate is tied to a handful of mega-deals, and capex guidance of $20–25B against $3.0–3.4B revenue guidance means heavy ongoing financing needs.^[2]
• No published per-token rate card in marketing pages — pricing is "transparent $/token" with volume discounts, but specific model rates require the self-service console, complicating comparison shopping.^[1]
• Heritage friction in some procurement processes — the Yandex lineage still surfaces in vendor reviews; one HN commenter reported legal counsel vetoing Nebius during GPU-provider due diligence.^[7]
• Quantized serving trade-offs — community reports note FP8-quantized model variants on Nebius's serving stack, which buyers benchmarking against full-precision endpoints should verify per model.^[7]
• Open models only — no proprietary frontier models (Claude, GPT, Gemini) are served, so teams needing closed models still need a second provider.^[1]

---

What Developers Say

Community discussion of Nebius on Hacker News through June 2026 is thinner than its financial profile would suggest — there is no major dedicated Token Factory thread, and most mentions are stock chatter or passing infrastructure references; that gap between Wall Street attention and developer mindshare is itself a data point.^[7]

"We do both: managed Kubernetes when it's available (AWS, Nebius, others)" — reissbaker on HN, building a GPU cloud, April 2026^[7]

"Nebius AI Studio" runs FP8 quantized model versions; "my experience with FP8 output has been pretty decent" — KronisLV on HN, using GLM with Claude Code^[7]

"When we were doing a deep-dive into cloud GPU providers, legal counsel veto'd them for this reason." — chias on HN, on Nebius's Yandex heritage during a European-infrastructure evaluation^[7]

The pattern: practitioners who use Nebius treat it as a competent GPU/inference utility, while the criticism that does surface is about provenance and procurement comfort rather than product quality.^[7]

---

Pricing & Licensing

Per-model $/token rates are published in the self-service console rather than the marketing site; volume discounts apply as usage scales, with no infrastructure or idle-GPU charges on shared endpoints.^[1]

Licensing model: Proprietary managed platform serving open-weight models; fine-tuned custom models deploy under the same per-token billing.^[1]

Hidden costs: Dedicated endpoints carry reserved-capacity commitments; quantized (e.g., FP8) variants may differ from reference model quality, which is a cost-quality trade rather than a line item.^[1][7]

---

Competitive Positioning

Direct Competitors

When to Choose Nebius Over Alternatives

• Choose Nebius Token Factory when: you serve open models at volume, want per-token pricing backed by an owner-operator with a 99.9% SLA, need EU data residency or SOC 2/HIPAA/ISO 27001, or want one vendor spanning GPU clusters to managed inference.
• Choose Baseten when: custom-model deployment workflows, compliance plus hands-on engineering support, and broad GPU selection matter more than owning-the-stack economics.
• Choose Together AI when: you want research-driven serving optimizations and a training-to-inference platform from one venture-backed vendor.
• Choose Groq/Cerebras when: custom-silicon latency is the deciding factor.

---

Ideal Customer Profile

Best fit:

• Enterprises standardizing on open-weight models (DeepSeek, Qwen, Llama) that want a contractual SLA and audited public-company counterparty rather than a startup
• European organizations needing EU data residency with SOC 2 Type II/HIPAA/ISO 27001 inference^[1]
• Teams scaling from per-token APIs to dedicated endpoints to GPU clusters without switching vendors
• High-volume batch workloads that can exploit the Base tier's throughput pricing^[1]

Poor fit:

• Teams that need proprietary frontier models (Claude, GPT, Gemini) from the same endpoint
• Buyers whose procurement or legal teams balk at the Yandex heritage^[7]
• Small projects wanting the absolute lowest per-token price over SLAs and compliance

---

Viability Assessment

The viability question inverts the usual startup calculus: Nebius will not run out of money quietly — it is a public company with NVIDIA on the cap table and $27B of contracted Meta demand — but Token Factory's fate depends on whether management keeps investing in the high-margin serving layer or lets the infrastructure mega-deals consume all strategic attention.^[2][4]

---

Bottom Line

Nebius is the public-market heavyweight entering managed inference from below: it owns the GPUs, the data centers, and now (via Eigen AI) a credible optimization stack, and it sells all of that through Token Factory at per-token prices with enterprise SLAs and EU residency.^[1][6] For buyers, the trade is financial durability and integration depth in exchange for a thinner developer ecosystem and an open-models-only catalog.

Recommended for: Enterprises serving open-weight models at volume that value an audited, NVIDIA-backed counterparty, compliance certifications, EU residency, and a path from per-token APIs to dedicated capacity.^[1][4]

Not recommended for: Teams needing closed frontier models, developer-experience-first startups better served by Baseten or Together's tooling and community, or procurement environments where the Yandex heritage is disqualifying.^[7]

Outlook: Watch whether Token Factory's "fastest-growing segment" claim gets quantified in coming quarters, how fast Eigen AI's optimizations land in production endpoints, and whether $7–9B ARR guidance survives 2026 — execution there would make Nebius the first inference platform with hyperscaler economics.^[2][5]

---

Research by Ry Walker Research • methodology