AI inference capacity is spreading beyond the hyperscalers

Training centralized, inference is spreading out

Vultr argues that frontier labs require massive data-center capacity and hyperscale GPU clusters for training large language models, but that the technology debate is shifting from centralized training at scale to local and decentralized inference. The reason, in its account, is that autonomous agents operate in dynamic, multi-step and continuous feedback loops, which change what inference infrastructure has to do. m:…2915

Vultr says local compute is vital to reduce latency and keep sensitive data within the protected enterprise perimeter, and that open AI models enable businesses to fine-tune model weights on internal corporate data. Hosting open models on dedicated infrastructure, it says, provides enterprises with more control over spend. This suggests the inference decision is becoming less about raw model access and more about where work runs, who controls the bill, and what data never leaves the perimeter. m:…2915

Specialist capacity is going global

CoreWeave says it has announced an expansion of its global AI cloud platform with new capacity in Denmark, bringing the majority of 22.5 MW of IT capacity online at atNorth DEN01 data center. The facility uses NVIDIA Blackwell systems connected with NVIDIA Quantum-X800 InfiniBand, and the data center has a heat reuse agreement to supply surplus heat to the district heating network. m:…4788

The company says it is building a Petabit-scale global AI backbone network and extending its Global AI Network into the Middle East and Asia Pacific, with new points of presence in the UAE, India, Indonesia, Singapore, Australia, Korea and Japan. The network expansion begins in Q4 2026 and extends to a full loop in mid-2027, the company says, using 400 Gbps connections at each point of presence and hardware-based MACsec encryption for data in transit. m:…4788

CoreWeave says it operated 51 data centers worldwide as of June 30, 2026, with 1.5 gigawatts of active power at that date and approximately 4.2 gigawatts of contracted power as of August 11, 2026. The scale of that footprint, combined with the new Denmark capacity and the planned network loop, points to a specialist provider building the kind of international backbone that was once associated mainly with the largest cloud operators. m:…4788

A second route: open stacks on alternative clouds

Vultr says avoiding vendor lock-in was the top-cited reason for shifting AI workloads between venues, and that legacy hyperscalers use proprietary APIs and managed AI pipelines that bind enterprises to their ecosystems. It also says 37 percent of respondents partner with alternative cloud providers as their main strategy for building new AI use cases, which suggests the alternative-provider route is a deliberate strategy for a substantial group, not just a fallback. m:…2915

Vultr says it integrates VAST Data AI operating system for data management and the SUSE AI Factory for standardized AI software layers, and that it supports AMD EPYC processors for CPU workloads and AMD Instinct processors for GPU requirements. The company presents this as a full-stack, open approach, in contrast to proprietary pipelines, and the implication for buyers is to compare not just compute prices but how easily a workload can move if the provider changes. m:…2915

Inference as a packaged platform

Hive has rebranded to Deploy, an enterprise AI inference platform, and the company says the platform formerly known as Hive Portal has been refreshed and renamed to Deploy, with existing Hive accounts and logins remaining active under the new brand. Deploy offers three product pillars, the company says: leading LLMs, proprietary specialized AI models and turnkey AI applications. m:…3716

The company says its CSAM Detection API is powered by Thorn technology, that it offers specialized models for IP and publicity detection, and that it maintains its own secure datacenter infrastructure. Deploy says it has raised more than 200 million dollars from investors and has more than 250 employees, and that it supports the NO FAKES Act and the TAKE IT DOWN Act. Taken together, the rebrand suggests inference providers are competing on packaged safety, productivity and specialized models, not only on raw model access. m:…3716

The three moves point in the same direction: inference work is spreading to specialist capacity, decentralized venues and packaged platforms, while the network and data controls around it become part of the product. Executives weighing an inference strategy should therefore ask where each workload runs, how it moves between venues, and what safety and specialized models come with the platform, rather than treating inference as a commodity that only differs by price. m:…2915 m:…4788 m:…3716