Compute & Infrastructure

SoftBank's SB Energy is redeveloping Department of Energy land in Ohio — a former nuclear weapons site — for a massive data centre campus that will include 10GW of server capacity, 10GW of new generation facilities, and a $4.2bn grid upgrade, according to The Register. The project represents a shift in hyperscale buildout strategy: rather than relying on utility commitments or power purchase agreements, SoftBank is co-locating generation capacity at facility scale. This model addresses grid connection delays that have constrained AI infrastructure deployment, though it raises questions about capital efficiency and stranded asset risk if AI demand proves less durable than projected.

The use of DoE land previously contaminated by uranium processing adds regulatory complexity, though the site selection suggests government willingness to repurpose legacy defence infrastructure for AI competitiveness. The 10GW generation component — likely natural gas or nuclear given baseload requirements — will test whether hyperscalers can secure environmental permits faster than utilities, or whether community opposition simply shifts from data centres to power plants.

Elon Musk announced plans to create 'Terafab', a chip fabrication venture spanning Tesla, SpaceX, and xAI that would produce a terawatt of computing power annually — approximately 50 times current global chip production — with most capacity destined for space-based data centres, according to The Register. Musk claims the effort will use 'new physics' to achieve this scale. The plan lacks critical details: no manufacturing partnerships with TSMC or Samsung have been announced, no capital expenditure estimates provided (comparable fabs cost $15-30bn each), and no timeline specified. Industry analysts note that even if technically feasible, ramping a single advanced node fab to volume production typically takes 3-5 years — scaling to 50x current production would require a manufacturing ecosystem that doesn't exist.

The space data centre component depends on SpaceX perfecting Starship, which remains in development. Bloomberg reports that Fidelity, a SpaceX investor, sees a path forward but acknowledges significant technical hurdles. The economics of launching and maintaining orbital data centres — including radiation hardening, thermal management in vacuum, and latency constraints — remain unproven. Unlike Musk's automotive or launch ventures where prototype demonstration preceded scaling, Terafab appears to be announced without proof of concept.

Heat flux in next-generation AI accelerators is projected to exceed 1,000 W/cm², shifting the primary bottleneck in semiconductor design from lithographic resolution to thermal management, according to analysis presented at a Wiley Knowledge Hub event. Legacy thermal measurement techniques cannot characterise hotspots in 3D-stacked chips or heterogeneous integration architectures, creating a metrology gap that constrains performance validation. This is confirmed by Nvidia's hardware roadmap: the company's LPU-based LPX rack will consume up to 160kW and require full liquid cooling, matching the power density of its Vera Rubin systems, Data Center Dynamics reports. Rack-level power density has increased 4x in three years, and air cooling is no longer viable for frontier AI systems.

The shift has supply chain implications: cooling infrastructure vendors (liquid cooling loops, cold plates, facility heat rejection systems) become critical path dependencies, while demand for advanced thermal metrology tools — including infrared thermography, scanning thermal microscopy, and time-domain thermoreflectance — is accelerating. Facilities designed for 30-50kW racks face costly retrofits or obsolescence.

JPMorgan launched a credit default swap basket covering the debt of five hyperscalers, providing institutional investors with a liquid instrument to hedge exposure to AI infrastructure borrowing, Bloomberg reports. The product's creation reflects recognition that AI capital expenditure represents a distinct credit risk profile — hyperscalers are issuing unprecedented volumes of debt to finance GPU purchases and data centre buildout, creating concentrated exposure in fixed income portfolios. Unlike traditional tech capex, AI infrastructure investments face uncertain monetisation timelines and binary outcome risk: if AI model performance plateaus or enterprise adoption disappoints, operators could face stranded assets measured in tens of billions of dollars.

The timing is significant: it arrives as total AI infrastructure debt issuance accelerates but before any major defaults or writedowns have occurred. This suggests credit markets are pricing in tail risk that equity markets have not fully discounted. For infrastructure planners, the availability of hedging tools may paradoxically increase capital availability — lenders more willing to finance buildouts when they can hedge — but also signals that sophisticated investors see material downside scenarios.

South Korea's Upstage is negotiating a deal with AMD to deploy 10,000 MI355 GPUs across the country following a visit by AMD CEO Lisa Su, Data Center Dynamics reports. Meanwhile, Eleveight AI has deployed 512 Nvidia B300s at a 2MW facility in Gagarin, Armenia, Data Center Dynamics reports, with construction slated for completion this month. Australia's government published formal expectations for data centre and AI infrastructure projects, including job creation requirements and protection of power and water resources, Data Center Dynamics reports.

These moves reflect a pattern: mid-tier economies building domestic AI compute capacity to reduce dependence on US or Chinese cloud providers, while creating regulatory frameworks to ensure projects deliver local economic benefits rather than simply consuming resources. Armenia's buildout is particularly notable — a small economy with limited power infrastructure investing in frontier AI hardware suggests compute capacity is becoming a geopolitical asset independent of immediate commercial returns. Australia's regulatory stance signals that governments will increasingly condition data centre permits on tangible local value creation, adding non-technical constraints to buildout timelines.