Compute & Infrastructure

Alibaba has announced it will expand data centre footprint in Europe and the Middle East while simultaneously targeting 20 GW of global cloud capacity by 2032 — a figure that would make it one of the largest compute landlords on earth. Alongside this, the company unveiled the Zhenwu V900 AI chip, slated for commercial release in Q1 2027, per Data Center Dynamics. The geographic expansion into Europe and the Middle East, reported by Bloomberg, is strategically timed to capture AI workloads from markets where US hyperscaler dominance is under regulatory scrutiny.

The 20 GW figure is a projection confirmed as an official target, not a built capacity number — the gap between now and 2032 represents an enormous construction and capital execution challenge. More immediately significant is the Zhenwu V900: if Alibaba can deliver a competitive AI accelerator at scale in early 2027, it reduces China's hyperscaler exposure to US export controls on NVIDIA and AMD hardware. This is the clearest example yet of a Chinese cloud operator treating chip self-sufficiency and data centre footprint as a single integrated sovereignty strategy.

OpenAI and Anthropic are reported to be actively pursuing 20–30 MW data centre deals to plug near-term capacity shortfalls, even as both companies have committed billions to gigawatt-scale facilities that remain under construction, per Tom's Hardware. This bifurcated procurement strategy — small deals now, massive campuses later — reflects a structural mismatch between the multi-year lead times for gigawatt-scale power procurement and construction, and the quarterly cadence of model release and inference demand growth.

This dynamic has direct implications for operators of mid-tier colocation facilities, which had been under margin pressure from hyperscaler self-build trends. The scramble for immediately available, pre-energised capacity is creating a seller's market in the 20–100 MW range. It also validates the positioning of regional players like BDx, which broke ground on a 120 MW facility in Jatiluhur, Indonesia with a stated go-live of early 2027 per Data Center Dynamics — the timeline aligns precisely with the window when frontier labs need interim capacity most.

NVIDIA has launched the DSX Ready qualification program, a vendor certification framework for power and cooling hardware used in data centres built on its DSX AI factory blueprint, per ServeTheHome. On its surface this is a compatibility assurance program. Strategically, it is NVIDIA codifying its reference architecture as the industry standard and creating a certified supplier ecosystem that reinforces customer dependency on NVIDIA's design choices at the facility level.

This move comes precisely as the cooling technology landscape is fracturing. Full liquid cooling at densities above 250 kW per rack is moving from pilot to production deployment, as detailed by IEEE Spectrum in coverage of CoolIT's fanless direct liquid cooling systems. By establishing a qualification program now, NVIDIA positions itself to influence which cooling vendors become the default choices for the next generation of AI factories — a market worth tens of billions annually. Operators who build to DSX specifications become structurally anchored to NVIDIA's hardware roadmap.

SoftBank is executing what Bloomberg describes as one of the largest corporate high-yield bond sales in history, offering record yields to attract investor capital for its AI infrastructure push, per Bloomberg. The record yields are not incidental — they reflect investor pricing of SoftBank's sub-investment-grade credit against the execution risk of deploying capital into AI infrastructure at the pace Son has committed to, including the $100 billion Stargate commitment.

The fact that this deal is proceeding confirms that high-yield capital markets remain accessible for AI infrastructure at scale, but the yield premium required represents a meaningful cost of capital disadvantage relative to investment-grade hyperscalers financing equivalent buildout at far lower rates. If AI infrastructure returns prove slower to materialise than Son's timelines suggest — a risk made concrete by OpenAI and Anthropic's own interim capacity scramble — the debt service burden on these bonds becomes a strategic liability.