Operating AI Infrastructure at Multi-Model, Multi-Cloud Scale | RAISE Summit 2026

Strategic Shift to Inference-First Models:

• Nebius is pivoting from traditional IaaS (managed Kubernetes, VMs) to a managed "Token Factory" service, recognizing that while training is an investment, inference is the primary driver of business growth and AI application.
• Customers increasingly prefer abstraction layers where they specify traffic patterns (RPS, tokens) rather than raw infrastructure metrics (GPU counts, megawatts, data center specs).
• Base10 describes its market positioning as "inference is everything," addressing the gap between physical GPU infrastructure and the need for reliable, economic, scaled model serving.

Infrastructure & Deployment Strategies:

• Base10 (Aggregation Model):
  • Operates across 18–20 cloud providers spanning 90+ regions to aggregate demand and secure GPU supply, including H100 clusters.
  • Provides multi-cloud orchestration to satisfy data residency, latency, and redundancy requirements, effectively creating a global compute mesh.
  • Solves "multi-model performance" and "cohesive developer experience" as key value pillars beyond raw hardware access.
• Nebius (Vertical Integration Model):
  • Leverages deep integration into its own data centers and hardware design (e.g., custom server RAM, PCB debugging, specific cooler optimization) to optimize cost-to-performance ratios.
  • Utilizes in-house expertise to debug issues at the physical layer (hardware, network) rather than just application code, enabling faster incident resolution.
  • Currently serves as a provider for Base10, illustrating a potential for industry collaboration despite competitive positioning.

Market Trends & Model Dynamics:

• Open Source Acceleration:
  • Adoption of open-source models is driven by their ability to cover frontier-level capabilities in coding and planning, reducing the need for proprietary training on standard tasks.
  • Companies utilize a "hybrid chef" strategy: using open-source models as a base, fine-tuning them with proprietary data to outperform frontier models in niche domains, or integrating specialized external models (e.g., Pionote for diarization).
  • A maturity curve exists where customers start with closed-source APIs (OpenAI, Anthropic) but transition to fine-tuned open-source models as costs scale and proprietary data advantages become critical.
• Token Efficiency:
  • Benchmark performance is increasingly tied to "reasoning budget" (tokens processed); a cheaper per-token model (e.g., GLM 5.2) may be less efficient per task than a more expensive frontier model (e.g., Opus) if it requires more tokens to solve.
  • Specialization allows for novel experiences where models answer questions faster and with higher token efficiency than off-the-shelf alternatives.

Commercial Models & Pricing:

• Nebius Token Factory:
  • Pricing is flexible and customer-centric; while the underlying infrastructure uses GPU hours, the service communicates value in "tokens" to match customer traffic expectations.
  • For large enterprise clients, pricing remains transparently based on GPU hours to allow clear economic modeling, though the platform supports token-based abstraction.
  • The value proposition focuses on transparency, allowing customers to trade cost for redundancy or varied hardware configurations.
• Base10:
  • Primarily charges per token where natural, but retains flexibility to translate usage into other units (e.g., GPU hours) based on customer needs.
  • Heavy investment in auto-scaling infrastructure ensures customers pay only for active usage rather than reserved capacity, mitigating "token maxing" waste.

Geopolitics & Sovereignty:

• The market is projected to remain fragmented regarding model sovereignty, with ongoing shifts between US, Chinese, and European (e.g., French open-source) models.
• Open-source models are expected to persist as a permanent fixture to ensure availability and mitigate geopolitical risks associated with closed-source providers.
• Base10 and Nebius anticipate continued demand for localized or sovereign model hosting as data residency laws and political controversies (e.g., US government restrictions on Anthropic) evolve.