Compute & Infrastructure

SK Hynix has broken ground on an advanced memory packaging facility in West Lafayette, Indiana, targeting production readiness in time to supply HBM for NVIDIA's Rubin-Ultra platform, expected around 2028. This is a confirmed construction commencement, not a planning announcement. The facility addresses one of the most acute geographic concentrations in the AI hardware supply chain: HBM packaging has been almost entirely located in South Korea, creating a single-region dependency for a component that is non-substitutable in high-end GPU clusters. The Register reports the site is specifically designed for advanced packaging and testing, the most technically complex and yield-sensitive stages of HBM production.

Simultaneously, SK Hynix reported a five-fold jump in quarterly profit driven by surging HBM pricing, and confirmed plans to increase capital expenditure 'significantly' in 2026. Bloomberg notes investor concern about the sustainability of the HBM supercycle, with questions centring on whether Samsung's ramp of competitive HBM and potential demand softening post-2026 training buildout could compress margins. The Indiana investment therefore carries dual logic: it satisfies US domestic content requirements relevant to CHIPS Act incentives while locking in long-term supply relationships with US-based hyperscalers and NVIDIA.

Google's announcement of two distinct TPU 8 variants — TPU 8t optimised for training and TPU 8i for inference — alongside the replacement of x86 with its Arm-based Axion CPU in the same systems, represents a strategic architectural commitment rather than an incremental product refresh. The Register and ServeTheHome both confirm the dual-track design. The inference-specific variant directly targets the cost structure problem: inference at scale is becoming the dominant AI compute workload by volume, and purpose-built silicon can achieve substantially better tokens-per-watt and cost-per-query than training-optimised accelerators used in a dual-purpose configuration.

The elimination of x86 from Google's TPU host architecture in favour of in-house Axion Arm cores completes a vertical integration move years in the making. This removes Intel from Google's AI infrastructure stack at the server level, following the earlier displacement of NVIDIA GPUs from Google's internal training workloads. For the broader market, this is a proof-of-concept that vertically integrated hyperscalers can operate a fully proprietary compute stack at scale — a model that Amazon (Trainium/Inferentia + Graviton) and Microsoft (Maia + Cobalt) are pursuing with varying degrees of commitment.

Microsoft's confirmed A$25 billion investment in Australian AI infrastructure by end of 2029 is the largest hyperscaler country commitment in the Asia-Pacific to date and explicitly frames data sovereignty and regional AI capacity as the rationale. Bloomberg describes it as Microsoft's biggest-ever single-country investment. The investment will fund data centre construction and cloud expansion, though the split between new builds versus capacity upgrades at existing facilities is not yet confirmed in available reporting. At roughly $4.5B per year over the commitment period, this represents a sustained capital allocation that will require significant land, power, and cooling infrastructure to be secured in Australian markets not previously scaled for hyperscaler-grade density.

A survey reported by The Register finds that one in five UK firms have already migrated AI workloads abroad due to domestic energy costs — a finding with direct policy implications for a government that has positioned AI as a growth driver. The UK's industrial electricity prices are among the highest in the OECD, a structural disadvantage for compute-intensive workloads where energy is the dominant operating cost. Unlike call centre offshoring, which represented labour arbitrage, AI workload migration represents energy arbitrage — and the destinations are typically data centre markets with cheaper renewable power or lower grid charges, including Ireland, the Nordics, and increasingly the Middle East.

This dynamic creates a direct contradiction in UK AI policy: investment in sovereign AI capability and domestic model development is undermined if the compute infrastructure to run those workloads is price-uncompetitive. The government's industrial strategy has emphasised AI adoption, but without addressing the structural energy cost gap, domestic compute capacity will struggle to attract the scale deployments needed to anchor UK AI infrastructure leadership.

Two developments this week extend the semiconductor geopolitical contest into new territory. The bipartisan MATCH Act, introduced in early April and now advancing in Congress, would remove the Department of Commerce's discretionary authority over chip-export licensing and specifically extend controls to DUV lithography equipment — tools that ASML and others have continued exporting to China under existing waivers. Tom's Hardware notes the Act targets chipmaking tools, not just finished chips, which would tighten the loop on China's ability to build out domestic advanced node capacity using second-tier equipment. If enacted as written, this would represent a significant escalation beyond current BIS export controls and could create diplomatic friction with ASML's home government in the Netherlands.

Separately, Elon Musk disclosed that his TeraFab initiative will use Intel's 14A process technology, with Tesla constructing a pilot production line and SpaceX responsible for high-volume manufacturing. Tom's Hardware characterises this as likely a technology licensing arrangement rather than Intel operating the facility. This remains a disclosed plan with no confirmed construction timeline or regulatory approval — it should be treated as speculative at this stage. If realised, it would constitute a novel industrial structure: a private defence and space company operating a leading-edge semiconductor fab under a tech licensing model, outside the traditional foundry-customer framework.