Compute & Infrastructure

A data analytics group tracking AI data centre construction via satellite imagery has concluded that at least 40% of projects scheduled for 2026 completion will be delayed, driven by shortages in both skilled labour and critical materials including electrical switchgear, transformers, and cooling infrastructure. The hyperscalers themselves deny schedule disruptions, creating a direct conflict between corporate communications and third-party physical evidence. Tom's Hardware notes the disparity is visible in construction site activity levels compared to projected milestones.

Compounding this, a separate analysis documents that local community opposition — through courts, planning bodies, and direct political lobbying — is forcing outright cancellations and significant delays at projects across the US and internationally. Tom's Hardware reports this is already costing hyperscalers billions in sunk costs and deferred capacity. The convergence of supply-side construction bottlenecks with demand-side community resistance creates a structural capacity gap that announced investment figures do not capture.

TSMC has increased both its revenue guidance and capital expenditure plans, explicitly attributing the revision to sustained AI-driven demand and describing it as a 'multiyear megatrend.' The company is accelerating expansion of 3nm-capable capacity. Tom's Hardware reports TSMC simultaneously warned that escalating Middle East conflict poses a profitability risk through higher input costs, likely referencing energy price exposure and logistics disruptions. TSMC's dominant position in advanced node manufacturing — particularly for NVIDIA's GPU supply chain and Apple's silicon — means its cost structure directly sets the floor for AI compute economics globally.

This guidance revision is a confirmed forward-looking signal from the single most critical chokepoint in AI hardware supply. The CapEx increase is an announced plan rather than completed capacity, but TSMC's execution track record on facility buildouts is substantially stronger than most peers. The Middle East caveat is a material risk disclosure that investors and procurement strategists should weight seriously given ongoing regional instability.

xAI staff are actively contacting semiconductor fabrication equipment and materials suppliers to obtain pricing and delivery timelines for the Terafab project, and are reportedly prepared to pay above-market rates to secure priority positioning. Tom's Hardware characterises the pace as 'light speed' by Musk's own framing. This places Terafab firmly at the feasibility and procurement stage — it is not under construction and no site has been confirmed as operational. The premium pricing signal indicates xAI anticipates competing for constrained supply against established customers including TSMC's existing foundry clients.

The strategic intent is plausible — vertical integration into chip fabrication would reduce xAI's dependence on TSMC and NVIDIA supply allocation — but the execution risk is enormous. Building a competitive leading-edge fab from scratch takes five to seven years minimum and tens of billions in capital even under ideal conditions. The supplier outreach phase is a necessary precursor to any serious planning, but it is many steps removed from a functioning facility.

Google is negotiating with the US Department of Defense to deploy its Tensor Processing Units and Gemini AI models within classified computing environments, using the Google Distributed Cloud architecture. Tom's Hardware reports Google is seeking contractual controls governing TPU use for mass surveillance and autonomous weapons applications — a governance condition that reflects both reputational management and internal employee relations constraints following past controversies over Project Maven.

The infrastructure implication is significant: deploying TPUs in air-gapped or classified environments requires purpose-built hardware configurations and supply chain controls distinct from commercial cloud deployments. If this contract closes, it establishes a template for custom silicon — rather than commodity GPU racks — as the preferred compute substrate for sovereign AI deployments, with implications for how defence agencies worldwide approach hardware procurement.

Volt has announced plans for an 800MW AI data centre campus in Rotterdam, Netherlands, branded as an 'AI Gigafactory' and potentially powered by North Sea offshore wind. Data Centre Dynamics reports the project is at the announcement stage with no confirmed construction timeline or power purchase agreements disclosed. Rotterdam's port infrastructure and proximity to subsea cable landings make it a strategically credible location, and North Sea wind capacity is a genuine power source — but 800MW of renewable-aligned capacity would require substantial grid interconnection agreements with Dutch grid operator TenneT, which is itself managing significant congestion challenges.

Separately, Eos Energy and Turbine-X have announced a partnership to develop power infrastructure for US AI data centres combining zinc battery storage with natural gas generation — a hybrid approach that signals recognition that renewables alone cannot meet the reliability requirements of large-scale AI compute without dispatchable backup. Data Centre Dynamics presents this as a commercial partnership at the development stage, not an operational deployment.