Compute & Infrastructure

One of the core financial signals underpinning the AI infrastructure trade — the price per unit of compute consumed — is softening, according to Bloomberg. This matters structurally: data centre operators and hyperscalers have justified enormous capex on the assumption that compute pricing would hold or rise as demand scaled. Declining inference prices compress the revenue per rack, which in turn tightens the return-on-capital math for facilities still under construction.

Richard Windsor of Radio Free Mobile, speaking to Bloomberg, goes further — arguing markets have fundamentally misread AI compute demand. The Philadelphia Semiconductor Index fell as much as 6% on July 2 after an 88% advance in the prior quarter, a correction Windsor frames not as profit-taking but as a reassessment of demand quality. The distinction matters for infrastructure investors: if demand is real but commoditising, the supply build is rational but margins compress; if demand has been overstated, the buildout itself is at risk of generating stranded assets.

CPP Investments — Canada's largest pension fund — is committing $1.75 billion to the EQT and EdgeConneX AI data centre platform, as reported by Data Centre Dynamics. EdgeConneX claims a 10GW development pipeline across its portfolio. To contextualise that number: 10GW of data centre capacity would represent a significant fraction of current global hyperscale capacity, making this a development pipeline rather than near-term deliverable — the distinction between announced pipeline and capacity actually permitted, contracted, and under construction is critical and not clarified in available reporting.

This investment is structurally notable because pension capital entering at scale signals a belief in long-duration, infrastructure-like returns from AI compute real estate — even as public equity markets are repricing semiconductor names downward. The divergence between private infrastructure capital (still bullish, long horizon) and public equity markets (reassessing near-term demand) represents a meaningful tension that will resolve one way or the other over the next 24 months.

IEEE Spectrum's analysis of AI data centre power dynamics identifies a problem qualitatively different from the headline energy consumption figures: AI workloads generate extreme moment-to-moment volatility in power draw as GPU clusters ramp between idle and full training or inference loads. IEEE Spectrum reports that utilities are adjusting long-term forecasts for volume but have not yet fully adapted their grid stability frameworks to handle the frequency and amplitude of load swings that high-density AI compute clusters produce.

This has direct implications for data centre siting and design. Facilities that cannot buffer their own load variability — through on-site storage, flywheels, or UPS architectures capable of absorbing rapid ramp events — risk becoming sources of local grid instability, which in turn creates regulatory and operational exposure. Data Centre Dynamics notes that the shift to inference workloads, which have spikier and less predictable demand profiles than training runs, is intensifying this dynamic. Power equipment suppliers are bifurcating accordingly: firms positioned for high-density, power-quality-sensitive applications are gaining, while traditional UPS and switchgear vendors designed for more stable enterprise loads are being marginalised, per Bloomberg's analysis of the power equipment market.

SEMI — the industry association representing SK Hynix, Samsung, and Micron — is lobbying against a US legislative proposal that would direct the administration to prioritise American manufacturers in memory chip supply allocation, according to Tom's Hardware. The industry's position is that mandated domestic prioritisation would extend shortages by disrupting the globally integrated procurement and allocation systems through which memory is efficiently distributed. SEMI's counter-proposal — tax deductions on consumer electronics to stimulate demand — suggests the lobby's primary concern is not shortage per se but preserving market-driven allocation.

This is a significant chokepoint signal. High-bandwidth memory — dominated by SK Hynix, with Samsung and Micron as secondary suppliers — is the binding constraint on NVIDIA GPU output and AI accelerator capacity more broadly. Any policy that disrupts the allocation of HBM to AI accelerator manufacturers, even with nominally pro-domestic intent, could directly slow AI infrastructure deployment. The fact that the three largest memory producers are aligned in opposition to this proposal indicates the industry views market mechanisms as more efficient allocators than government mandate — but it also exposes how little slack exists in HBM supply for any policy friction.

South Korea has announced an accelerated national AI semiconductor strategy, and a new power semiconductor fabrication facility is breaking ground in Dresden, Germany, per Semiconductor Engineering's weekly industry review. These are sovereign infrastructure plays with distinct motivations: Korea is defending its leadership in memory and advanced logic against Chinese competition and potential US policy shifts; Germany's Dresden investment reflects the EU's continued effort to rebuild domestic semiconductor manufacturing capacity after the European Chips Act, with power semiconductors serving both automotive and AI infrastructure markets.

On the server hardware side, ASRock Rack demonstrated one of the first production servers built around Arm's new AGI-targeted server CPU at Computex 2026, according to ServeTheHome. The 1U single-socket form factor signals that Arm is positioning its architecture not just for cloud inference (where it has already made inroads via AWS Graviton and Ampere) but for the full AI server stack. This is early-stage — ASRock Rack is a Tier 2 ODM, and the platform has not yet secured major hyperscaler design wins — but it represents a credible architectural alternative to x86 for AI infrastructure workloads that warrants tracking.