Compute & Infrastructure

The US Federal Energy Regulatory Commission has announced it will order grid operators to expedite interconnection applications from AI data centres that either generate their own power on-site or commit to demand curtailment during peak grid stress periods. The 90-day implementation window is unusually tight for a regulatory body of this scale, signalling genuine urgency around grid stability rather than a symbolic gesture. Tom's Hardware reports the order will apply to grid operators under FERC jurisdiction, covering the bulk of the continental US transmission system.

The practical effect is to create a two-tier interconnection queue: projects with co-located generation — natural gas, nuclear, or large-scale renewables — will advance faster than those relying purely on grid draw. This structurally advantages hyperscalers with the capital to build dedicated power plants alongside campuses, and disadvantages mid-tier colocation operators and conversion projects like Digi Power X's cryptomine-to-AI-data-centre effort in North Tonawanda, which simultaneously faces a municipal moratorium. The curtailment commitment mechanism also introduces operational risk for inference workloads that require continuous availability, potentially pushing mission-critical compute toward dedicated on-site power configurations.

Reports covered by both Data Center Dynamics and the Semiconductor Engineering Week in Review indicate Amazon is considering selling Trainium chips directly to external data centre operators. If confirmed, this represents a fundamental shift in Amazon's silicon strategy: Trainium was designed as an internal cost-reduction tool to reduce AWS dependence on NVIDIA, but external sales would position AWS as a merchant silicon vendor competing in the open market.

The strategic calculus is significant. NVIDIA's data centre GPU dominance — with H100 and B200 series commanding the majority of AI training deployments — rests on its position as the only at-scale merchant alternative. A credible Amazon entry into external chip sales would compress NVIDIA's pricing power, particularly for training workloads where Trainium is most competitive. However, the merchant chip market requires ecosystem support — compilers, libraries, and developer toolchains — that AWS has built internally but has not yet demonstrated at the depth NVIDIA's CUDA ecosystem provides. This remains a reported plan, not a confirmed product launch.

Presentations at VLSI 2026 confirmed Intel's 18A platform is progressing from device-level characterisation to full routed designs, incorporating backside power delivery — a capability that directly competes with TSMC's N2P roadmap — and new channel materials. Semiconductor Engineering frames this as platform momentum, and the VLSI venue, which requires peer-reviewed data, gives the results more credibility than marketing disclosures. Separately, Amkor's involvement in advanced packaging for Intel products was noted in the week-in-review as a significant win, consistent with Intel's strategy of outsourcing packaging to third parties while retaining leading-edge logic fabrication.

The distinction between technical validation and volume production readiness remains critical for infrastructure planners. Intel's 18A has passed yield-qualifying milestones in controlled environments, but the foundry services business — where external customers like Broadcom and potential hyperscaler chip programmes would validate the node — has not yet demonstrated the consistent yield economics that would make it a viable TSMC alternative at scale. The announcement that Apple is in discussions about Intel fabrication, referenced in the week-in-review citing Trump administration commentary, adds a high-profile potential customer but remains unconfirmed at the contractual level.

Slovenia has launched FRIDA, a national HPC system designed to enable AI research collaborations between academic institutions and businesses, per Data Center Dynamics. The system is now operational — a confirmed deployment, not an announced plan. While Slovenia is a small market, FRIDA fits a consistent pattern across European nations of investing in domestic HPC and AI compute capacity, driven by EU digital sovereignty directives, concerns about data residency under GDPR, and strategic industrial policy goals.

The broader European sovereign compute picture includes established national HPC programmes in Germany, France, Finland, and Italy, with EU-level coordination through EuroHPC JU. What distinguishes the current wave is the explicit AI workload orientation — earlier HPC investments targeted scientific simulation; newer deployments are being designed with inference and training workloads in mind, and are increasingly positioned as alternatives to US hyperscaler dependency for government and research use cases.