February 17, 2026

Top news (Global)

1. GPU platform efficiency claims are rising sharply. NVIDIA says its new Blackwell Ultra system offers major throughput-per-megawatt and cost-per-token gains, and it is already being deployed by several large cloud providers: NVIDIA GB300 NVL72 boosts agentic AI performance and efficiency.

2. Storage vendors are pushing “AI-scale” architectures and third-party validation narratives. Dell argues its PowerScale discrete, scale-out design delivers predictable linear scaling and sustained bandwidth, citing NVIDIA Cloud Provider (NCP) validation and Dell internal testing: Dell PowerScale: Discrete Architecture for AI-scale Storage and Efficiency.

3. Liquid cooling is being positioned as an essential enabling layer for AI data centres. Boyd highlights modular liquid cooling building blocks (cold plates, manifolds, CDUs) and simulation tools aimed at scaling deployments from edge to large AI campuses: Scalable liquid cooling essential for modern AI data centers.

Key deals & projects

Europe (Nordics) — Iceland

• atNorth ICE03 (Akureyri, Iceland): design choices tied to climate and materials
  • Feature highlights the site’s use of Iceland’s cool climate and renewable energy to support efficient direct liquid cooling.
  • Construction materials called out include Glulam laminated wood and locally produced Icelandic rockwool, positioned as lower environmental-impact choices while supporting scalable buildout: atNorth’s ICE03 in Akureyri showcases sustainable data centre design.

Global — AI infrastructure platforms (vendor and cloud ecosystem signals)

• NVIDIA GB300 NVL72 (Blackwell Ultra): deployment signals from major cloud providers
  • NVIDIA claims up to 50x higher throughput per megawatt and up to 35x lower cost per token vs Hopper, targeting low-latency “agentic AI” and coding assistants.
  • Named cloud adopters “deploying in production” include Microsoft, CoreWeave, and OCI.
  • NVIDIA also points to the next-generation Rubin/Vera Rubin NVL72 with a stated goal of another ~10x throughput-per-megawatt gain vs Blackwell: NVIDIA GB300 NVL72 boosts agentic AI performance and efficiency.

Power, cooling, and facility efficiency highlights

Cooling (thermal as a primary scaling constraint)

• Boyd: modular liquid cooling + simulation-driven design
  • Positions cold plates, manifolds, and CDUs as modular components to scale installations.
  • Promotes SmartCFD simulation to support design and rollout.
  • Frames demand backdrop as data centre electricity consumption projected at ~448 TWh in 2025 and expected to more than double by 2030.
  • Notes global manufacturing capacity and regional service coverage across North America, Europe, and Asia: Scalable liquid cooling essential for modern AI data centers.

Storage density and power draw (IT layer efficiency messaging)

• Dell: PowerScale discrete scale-out storage positioned for lower footprint/power
  • Dell argues for predictable linear scaling, sustained bandwidth, and stable write performance for AI workloads.
  • Cites NVIDIA Cloud Provider (NCP) validation and a Dell internal analysis of 64 SU configurations (August 2025).
  • Blog claims lower rack space and power draw versus alternatives (also referencing competitor pivots such as “VAST’s E-Box evolution”): Dell PowerScale: Discrete Architecture for AI-scale Storage and Efficiency.

Market structure / platform positioning (vendor lock-in vs openness)

• Dell critique of HPE Private Cloud AI
  • Dell characterises HPE’s “three-click” Private Cloud AI as prescriptive, GreenLake-dependent, and creating vendor lock-in plus Day-2 scaling challenges.
  • Dell positions Dell AI Factory as a more open, validated alternative intended to integrate with existing management tools: The three‑click illusion of HPE’s Private Cloud AI.

Policy and regulation

• No policy, permitting, grid interconnection, or regulatory changes were included in today’s stories.

Two-line wrap-up

Hardware/platform vendors are increasingly competing on measurable “throughput-per-megawatt” and operational scalability narratives.

Material choices, cooling architectures, and validated reference designs continue to be used to differentiate new AI-focused data centre builds and retrofit strategies.