Chips, Compute, Data & Energy

Evidence statusPrimary source · Research

The review links AI's carbon, water, hardware and grid burdens and organises mitigation options across system design, computing infrastructure, energy supply and governance.

Evidence statusPrimary source · Technical preview

OpenAI and Broadcom unveiled a purpose-built inference accelerator developed around large-language-model serving. The companies described it as the first step in a multi-generation platform intended to improve performance per watt and expand OpenAI's full-stack compute strategy.

Evidence boundaryEarly performance statements are company-reported and the source does not provide broad independent silicon benchmarks or fleet-scale production results.

Evidence statusPrimary source · Product release

NVIDIA reported that Vera Rubin had entered production across a large manufacturing and server ecosystem. The platform targets pod-scale agent throughput and introduces optical networking intended to support future very-large accelerator fabrics.

Evidence boundaryProduction status and throughput comparisons are vendor-reported; they do not by themselves establish delivered customer capacity or application-level efficiency.

Evidence statusPrimary source · Product release

NVIDIA introduced Vera as a CPU designed for agentic workloads, reinforcement learning and data processing, both as a standalone server processor and within Vera Rubin and storage systems. Major labs, cloud providers and system vendors were named as planned adopters.

Evidence boundaryAdoption statements describe partner plans, while the quoted x86 comparison is based on NVIDIA's specified benchmark configuration.

Evidence statusPrimary source · Research

The report examines AI applications in East Asian energy systems alongside rising data-centre electricity demand and the resulting implications for grids and public policy.

Evidence statusPrimary source · Research

Google published implementation guidance for training large models on seventh-generation Ironwood TPUs. The system scales through chip pods, optical circuit switching, data-center networking, high-bandwidth memory and a co-designed XLA and Pallas software stack.

Evidence boundaryThe guide explains Google's platform architecture and recommended practices; it is not an independent cross-accelerator benchmark.

Evidence statusPrimary source · Product release

OpenAI announced a partnership to add 750 megawatts of Cerebras low-latency inference capacity to its compute portfolio in staged deployments through 2028. The strategy treats real-time response speed as a distinct infrastructure requirement alongside large-scale training.

Evidence boundaryThis is a capacity agreement with phased future delivery, not evidence that the full 750 MW was online in 2026.

Evidence statusPrimary source · Product release

NVIDIA detailed the Rubin rack-scale platform as a co-designed system of CPU, GPU, switching, networking and data-processing components for training, inference and agentic AI. Partner systems were scheduled for the second half of 2026.

Evidence boundaryPerformance and cost reductions are NVIDIA projections against selected prior-generation configurations; broad customer results depend on full-system deployment.

Evidence statusPrimary source · Technical preview

AMD outlined the Helios rack-scale architecture and expanded Instinct MI400 portfolio as an open, modular foundation for very large AI systems. The roadmap combines accelerators, EPYC CPUs, Pensando networking and ROCm, while previewing later MI500 hardware.

Evidence boundarySeveral figures are roadmap targets or vendor projections; MI500 was a future preview and should not be represented as a 2026 available product.