Compute & Infrastructure

Seoul is executing a two-track sovereign compute strategy. First, Samsung Electronics and SK Hynix have been publicly co-opted into a national investment programme under President Lee Jae Myung, with both CEOs present at a government briefing to signal coordinated large-scale commitments in memory chips, data centres, and robotics. The specific investment figures have not yet been publicly disclosed — this remains an announced plan, not confirmed capital deployment — but the political staging is deliberate, designed to project continuity of supply for global AI customers dependent on HBM and DRAM. Bloomberg

Second, a senior government official has confirmed that Seoul is exploring a new investment fund capitalised by excess tax revenue generated by semiconductor sector profits. Bloomberg This is still at the planning stage, but the mechanism — recycling chip-sector windfalls into long-term infrastructure and technology investment — mirrors the sovereign wealth logic applied by Gulf states to energy revenue. SK Hynix's planned US ADR debut adds a parallel capital markets dimension, broadening the investor base for Korean AI memory assets at a moment when HBM supply remains a critical bottleneck in AI accelerator stacks.

Google has capped Meta's access to Gemini AI models because it lacks sufficient compute capacity to fulfil Meta's demand, according to reporting by the Financial Times cited by Bloomberg. The strategic implication extends well beyond the bilateral relationship: if Google — which operates one of the three largest global data centre fleets and has deployed custom TPU infrastructure at scale — cannot satisfy a single large enterprise customer's inference load, it is a direct indicator that the industry-wide gap between AI compute demand and available capacity is not a projection but a present operational condition.

This episode also illuminates a market structure risk. Meta's interest in Gemini suggests that even the largest AI developers are sourcing model capacity externally rather than relying solely on in-house infrastructure, which concentrates demand on a small number of hyperscale providers. Rationing by Google effectively forces customers to queue, self-provision at enormous capital cost, or accept service constraints — all of which have downstream implications for the pace of AI application deployment.

Foxconn parent Hon Hai Precision Industry reported a 40% year-on-year quarterly sales increase, beating expectations, with management attributing continued growth to AI demand. Bloomberg As the primary assembler of Nvidia's high-density AI server racks, Hon Hai's results function as a real-time proxy for the state of Nvidia's order fulfilment pipeline. A 40% jump is not a signal of demand cooling — it is a signal that the supply chain is still being pushed at maximum throughput.

However, semiconductor equities came under pressure this week as investors began questioning whether the pace of AI infrastructure spending can be sustained beyond 2026. Bloomberg This is a divergence between current operational data — which shows demand running hot — and capital market forward pricing, which is discounting a potential deceleration. Infrastructure planners should treat these signals independently: near-term procurement decisions should be driven by the operational reality confirmed by Hon Hai, while longer-term capacity commitments carry execution risk if hyperscaler capex cycles turn.

Jim Keller's semiconductor tooling startup has rebranded from Atomic Semi to Fab2 and is establishing a factory in Texas focused on producing small-scale semiconductor fabrication equipment — positioning itself explicitly as a 'fab fab,' a facility that manufactures the means of chip production rather than chips themselves. Tom's Hardware The strategic thesis is that high-volume, low-cost small fabs can be deployed at a fraction of the capital expenditure of a leading-edge TSMC node, enabling a broader base of organisations — defence contractors, national labs, sovereign programmes — to fabricate custom chips without accessing the concentrated advanced node supply chain.

This is early-stage and the commercial viability of small-fab economics at meaningful scale is unproven. However, the timing is significant: it intersects with US CHIPS Act-driven interest in domestic fabrication resilience, and Keller's profile will attract both capital and policy attention. The venture is not a near-term threat to TSMC's position in advanced logic, but it represents a serious attempt to create an alternative tier of the supply chain that could serve niche but strategically critical applications.