Compute & Infrastructure

Amazon, Microsoft, and Alphabet all signalled sustained AI infrastructure spending in their latest earnings cycles, with no meaningful pullback in forward capex guidance. This collective reaffirmation is analytically significant not just for NVIDIA but for the broader supply chain — memory, power infrastructure, networking, and cooling suppliers all depend on hyperscaler build cycles remaining intact. Bloomberg characterised the reports as providing 'relief to a sector battered in recent days.'

Vertiv, a bellwether for data centre power and cooling infrastructure, reported earnings and revenue growth yet saw its shares fall — mirroring the pattern seen with Samsung and SK Hynix. Data Center Dynamics noted the dynamic as a function of volatile AI data centre investment sentiment rather than any deterioration in underlying demand. The gap between reported performance and investor reaction reflects elevated expectations priced into these stocks, not a supply-demand reversal.

Fubon Research has published a projection — not confirmed by Google — that Google could produce more AI accelerators via its TPU programme in 2028 than NVIDIA sells in that year. Tom's Hardware reports the analysis further suggests Google may need to engage Intel Foundry to hit those volumes, given TSMC's constrained advanced-node capacity. This is a speculative projection from a single analyst house and should be treated as a directional signal, not a confirmed build plan.

The strategic implication, if even partially correct, is significant: it would represent the most direct challenge yet to NVIDIA's dominance in AI accelerator unit volumes, and would validate Intel Foundry's viability as a high-volume advanced AI chip manufacturer — a thesis Intel has been unable to prove commercially at scale. The scenario also underscores a broader trend of hyperscalers vertically integrating silicon to reduce dependence on NVIDIA's supply-constrained H-series and B-series products. ARM's record royalty and licensing results, including over $2 billion in AGI CPU demand, further confirm that non-NVIDIA silicon architectures are gaining traction in the data centre. Data Center Dynamics

Samsung and SK Hynix both reported record-breaking profits in the same week, driven by HBM and high-density DRAM demand tied to AI accelerator production. Bloomberg reported that investors continued selling for days after the announcements — a telling sign that financial markets are treating AI hardware demand as potentially cyclical rather than structurally durable. The memory market has experienced severe boom-bust cycles historically, and investors are discounting current earnings against the risk of a demand cliff.

From a supply chain standpoint, HBM remains one of the tightest chokepoints in AI accelerator production. SK Hynix holds a commanding lead in HBM3E supply, with Samsung still qualifying its HBM3E product with NVIDIA as of mid-2026. Any softening in hyperscaler GPU orders would cascade quickly into HBM order books, given the close coupling between accelerator production schedules and memory stack procurement. The record earnings confirm current demand is real; they do not resolve the question of whether 2027 and 2028 demand will sustain current pricing.

South Korea has confirmed a 20 trillion won ($13.9 billion) injection into its sovereign wealth fund, explicitly mandated to invest in AI, data centres, and domestic infrastructure. Bloomberg notes this marks the first time the fund's mandate has been expanded to include domestic assets — a structural policy shift driven in part by the recent rout in Korean technology stocks, which has depressed valuations in Samsung, SK Hynix, and related firms. This is a confirmed government decision; specific investment allocations and timelines are not yet public.

In Southeast Asia, Thailand's True Internet Data Center is seeking approximately $2 billion in loan financing to build a new facility, positioning Thailand as a regional AI and cloud hub. Bloomberg reports the financing is being arranged with external lenders, though the facility itself is in early-stage development. Separately, a 400MW behind-the-meter gas-powered data centre is being planned in Abilene, Texas by PowerPlay AI and Sharon AI as a joint venture — a confirmed project announcement but with no disclosed construction start date. Data Center Dynamics

Crusoe and Aalo Atomics have confirmed a partnership to deploy a Spark small modular reactor unit at Idaho National Laboratory to power an AI data centre. Data Center Dynamics reports this as a confirmed deployment partnership, though SMR construction and commissioning timelines remain subject to regulatory and engineering milestones. This is one of the first confirmed operational integrations of an SMR with commercial AI data centre infrastructure at a federal research site, moving the nuclear-for-AI-compute thesis from concept to a live project.

The Idaho National Laboratory context is significant: it provides both a regulatory-friendly environment for experimental reactor deployment and a credibility anchor for the technology. The broader data centre industry is watching SMR and behind-the-meter gas projects closely as grid interconnection queues lengthen — in some US markets, wait times for utility-scale grid connections now exceed five years, making behind-the-meter generation increasingly attractive for large-scale operators despite higher unit energy costs.