Compute & Infrastructure

A new report highlighted by Bloomberg identifies a growing nationwide shortage of high-skilled semiconductor workers as a first-order risk to the US fab revival. Billions of dollars in new fabrication plants are under construction or planned — anchored by TSMC Arizona, Intel Ohio, and Samsung Texas — but the pipeline of trained technicians, process engineers, and equipment specialists is not keeping pace with projected hiring needs. The report calls for both industry resource-pooling on training and sustained federal funding to close the gap.

This is a textbook case of infrastructure investment outrunning human capital. Physical construction can be accelerated with capital; workforce development operates on multi-year educational cycles that cannot be compressed the same way. The risk is not that fabs won't be built, but that they will open below rated capacity or ramp slowly, delaying the point at which domestic US production materially reduces dependence on TSMC Taiwan. For policymakers, this is an argument for treating workforce development as load-bearing infrastructure, not a secondary concern.

South Korea's ₩1,350 trillion ($880 billion) combined semiconductor and AI infrastructure plan — breaking down to roughly $520 billion in chip manufacturing and the remainder in AI data centres and robotics — is facing a structural barrier that capital alone cannot resolve, according to Tom's Hardware. A single proposed megacluster would demand approximately 25% of Seoul's total power consumption. Water availability for cooling is a parallel constraint. The majority of the spending is corporate capex from Samsung and SK Hynix rather than direct government outlays, meaning private firms bear the risk if infrastructure bottlenecks stall returns.

This situation is not unique to South Korea — it mirrors dynamics in the US, EU, and Japan — but the scale is particularly acute given the density of planned investment in a geographically constrained peninsula with an already heavily loaded grid. Grid expansion and new generation capacity (likely nuclear, given South Korea's energy policy trajectory) must be treated as preconditions, not parallel workstreams, or the programme's timeline will slip materially.

The structural pivot of former cryptocurrency mining operators into AI data centre hosting is moving from announcement to delivered capacity. Data Center Dynamics reports that Galaxy Digital has completed the first phase of its facility conversion, with CoreWeave now able to move in. Separately, Blockfusion has signed a letter of intent for a $175 million lease covering up to 300MW in Niagara Falls with an undisclosed AI customer, per Data Center Dynamics. A third former crypto operator, Sato, has signed a preliminary agreement to develop an AI data centre in Bhutan, per Data Center Dynamics, leveraging the country's hydroelectric surplus.

These conversions are strategically significant for two reasons. First, they represent a rapid and relatively capital-efficient route to new AI compute capacity — existing power infrastructure, physical plant, and cooling systems can be repurposed faster than greenfield construction. Second, the 300MW Niagara Falls deal, if executed, would represent a substantial addition to available AI hosting capacity in a region with access to cheap hydroelectric power. The LOI status means this is confirmed interest but not committed capacity; the identity of the AI customer and final lease execution are the key variables.

A newly surfaced Intel patent, reported by Tom's Hardware, proposes an architecture called XBM that stacks DRAM using backend transistor processes and connects via UCIe chiplet links, eliminating the expensive silicon interposer that is central to current HBM implementations. The design incorporates built-in repair logic intended to improve yield on what are inherently complex 3D stacks. This is a patent filing, not a product announcement — Intel has not confirmed a commercial programme.

The strategic significance is substantial regardless of Intel's specific commercial trajectory. HBM supply is a critical chokepoint for AI accelerator production: SK Hynix dominates HBM3E supply to NVIDIA, Samsung is qualifying, and Micron is ramping, but all three depend on advanced packaging processes that are capacity-constrained. An architecture that achieves competitive memory bandwidth without the interposer step would reduce packaging complexity and cost, potentially broadening the supplier base and relieving a supply chain vulnerability that currently gives SK Hynix disproportionate leverage over NVIDIA's Blackwell and Vera Rubin production volumes. UCIe as the interconnect standard also matters — it is an open industry standard, which would reduce proprietary lock-in.

TeraWulf CEO Paul Prager confirmed on Bloomberg that the company has signed a 20-year lease agreement with Anthropic for a purpose-built AI campus at TeraWulf's Kentucky site. A 20-year tenure is exceptional in data centre contracting, where five-to-ten-year terms with renewal options are standard. TeraWulf's Kentucky facility benefits from access to low-cost, predominantly nuclear-sourced power — a factor that aligns directly with Anthropic's stated commitments on clean energy and with the operational economics of sustained AI training workloads. The campus will be purpose-built for Anthropic's requirements rather than adapted from existing infrastructure.

This deal structure reflects a broader shift among frontier AI labs: as training runs grow larger and inference infrastructure becomes a competitive differentiator, the option value of flexible cloud capacity is worth less than the cost certainty and operational control of dedicated, purpose-built facilities with long-term power agreements. It also validates the thesis that power-advantaged locations — particularly those with access to nuclear or hydro baseload — will attract anchor tenants willing to commit at decade-plus timescales.