- Amazon’s $200 billion capital expenditure commitment for 2026 is the largest single-year infrastructure spend by any hyperscaler, with Amazon Web Services directing more than $100 billion toward AI cloud infrastructure, custom chips and data centers.
- VMware by Broadcom’s Private Cloud Outlook 2026 report found 56% of surveyed organisations run or plan to run AI inference in private clouds, while public cloud use for inference has declined.
- Supermicro, IBM and NVIDIA’s joint KV cache solution recorded up to 56x faster Time-to-First-Token in their own testing on a 120-billion-parameter LLM at 130,000-token prompts, a gain that narrows sharply outside that specific workload profile.
Amazon’s $200 billion capital expenditure commitment for 2026 is compressing enterprise decision timelines on AI infrastructure placement. Google’s $17 billion investment in its first Wyoming data center, announced November 5, 2026, with the site itself acquired from Crusoe the following day, underscores how fast capacity is being locked up. Where workloads land, public cloud, private cloud, colocation or edge, is now a decision enterprises are being forced to make explicitly rather than by default.
Colocation Fills the Capacity Gap
Northern Virginia’s primary market vacancy reached historically low levels in early 2026 as AI and hyperscale demand pushed high-density colocation deployments.
A separate survey put roughly 70% of CIOs favouring colocation and hybrid environments for AI and machine learning workloads. That figure would make colocation the plurality choice for enterprise AI placement, ahead of either pure public or pure private cloud.
Private Cloud Reclaims Sensitive Workloads
According to VMware by Broadcom’s Private Cloud Outlook 2026 report, 56% of surveyed organisations currently run or plan to run AI inference in private clouds, while public cloud use for inference has declined. The drivers are consistent across sectors: data sovereignty, regulatory compliance and cost predictability. Security-sensitive and business-critical AI workloads are being repatriated from public environments to private ones.
Hyperscalers Package AI for Enterprise
AWS’s $100 billion-plus commitment and Google’s $17 billion Wyoming investment are the supply side. The demand side is enterprises wanting enterprise-grade tooling without building the full stack. Microsoft has responded with Windows hybrid intelligence, an architecture that combines locally running AI models with cloud-based services and shifts workload placement dynamically by demand. An expanded Lumen and Microsoft collaboration, announced around the same period, targets co-selling enterprise AI-ready networks that connect distributed data across hybrid and multi-cloud environments for Azure-based workloads. The pattern across all three hyperscalers is the same: pre-packaged services, specialised hardware and connectivity options designed to abstract infrastructure complexity for the enterprise buyer.
Managed Services Add Governance Tooling
Two product launches in early November illustrate where managed AI services are heading. On November 3, 2026, Compass Group launched AirisOne, an AI-powered managed services platform for facilities managers designed to unify data from building operational systems, identify issues and prevent disruptions. CloudBolt recently added generally available agent-governance capabilities to its cloud management platform, letting organisations tie AI agent actions to named user identities and enforce existing permissions and workflows. Both reflect the same pressure: even well-configured AI agents require governance infrastructure that most in-house IT teams are not resourced to build themselves.
Purpose-Built Hardware and Real Numbers
On November 5, 2026, Supermicro, IBM and NVIDIA announced a KV cache storage solution pairing Supermicro Petascale storage servers with IBM Storage Scale software and NVIDIA Dynamo inference software. In joint testing on a 120-billion-parameter LLM at 130,000-token prompts, the combination produced up to 56x faster Time-to-First-Token compared to a baseline configuration, according to the companies. That is a vendor-originated benchmark on vendor-selected workloads; real-world gains will vary by model size and prompt distribution. The result still identifies where purpose-built infrastructure delivers its clearest advantage: long-context inference at scale, where Time-to-First-Token directly affects application responsiveness. For context on the infrastructure investments driving these configurations, the Nutanix and Supermicro Blackwell Ultra packaging covers the private inference angle in more detail.
Edge AI: Latency as a Hard Requirement
For autonomous equipment and smart manufacturing, local inference latency is a functional requirement, not an optimisation. Industrial robots and autonomous vehicles cannot absorb the round-trip time to a centralised cloud without compromising safety or operational reliability. The volume and velocity of IoT data in these environments compounds the problem: transmitting everything to the cloud for processing is simply not viable at production scale. Edge deployment brings inference to the data source, cutting the perception-action loop to the milliseconds these applications require. Real-time quality control, predictive maintenance and production-line optimisation all depend on that immediacy. Where cloud AI suits batch processing and less time-sensitive workloads, edge AI defines the operational floor for real-time systems.
Hybrid as the Default
A Flexera 2026 survey found that 73% of organisations operate hybrid environments. On November 2, 2026, Lumen and Microsoft announced their expanded collaboration to co-sell enterprise AI-ready networks connecting distributed data across hybrid and multi-cloud environments for Azure-based AI workloads. Windows hybrid intelligence, Microsoft’s disclosed architecture for dynamic workload placement, points in the same direction.
Hybrid gives teams flexibility; it also hands them the full operational surface area of three infrastructure models to run, secure and govern simultaneously. Integration across those tiers, not the choice of any single one, is where most enterprise AI deployments run into trouble.
