Compute & Infrastructure
Top Line
The four largest hyperscalers — Amazon, Google, Meta, and Microsoft — have collectively committed approximately $2.4 trillion in forward AI infrastructure spending, with over $745 billion in new capex expected in 2026 alone, representing a sustained demand signal that will stress semiconductor and energy supply chains for years.
Amazon's annual capex is projected to reach $220 billion, driven significantly by AI memory costs, as AWS posts its fastest cloud growth since 2021 — confirming that inference-at-scale is now the primary demand driver, not just training workloads.
Chinese AI startup Moonshot is running frontier Kimi models on approximately 20,000 Nvidia Hopper chips accessed through Alibaba's cloud infrastructure, revealing that US-origin compute remains embedded in China's AI stack despite export controls — a significant strategic vulnerability in the US restriction regime.
The EU has issued a formal call for tender for seven supercomputing clusters under its AI gigafactory initiative, projecting at least €20 billion in unlocked private investment — the bloc's most concrete sovereign compute commitment to date.
Enterprise cloud infrastructure spending reached $143 billion in the most recent quarter, the eleventh consecutive growth quarter, confirming that AI-driven demand is structural rather than cyclical.
Key Developments
Hyperscaler Capex Commitments Cross $2.4 Trillion — Infrastructure Supply Chains Face Sustained Stress
The aggregate forward spending commitments of Amazon, Google, Meta, and Microsoft now stand at approximately $2.4 trillion, with the four companies having already deployed more than $1 trillion since 2023, according to Bloomberg and Tom's Hardware. Amazon alone is tracking toward $220 billion in annual capex, a figure Data Center Dynamics attributes in significant part to elevated AI memory costs — meaning HBM pricing and DRAM supply tightness are now directly legible in hyperscaler financial guidance.
The Tom's Hardware analysis flags a 'hidden debt' figure exceeding $1.65 trillion in operating lease and purchase commitments sitting off balance sheet across the four companies. This distinction matters for infrastructure analysts: these are contractual obligations to colocation providers, grid operators, and hardware vendors, not discretionary budget lines. The implication is that even in a demand downturn, construction pipelines will continue to execute, sustaining pressure on the transformer, cooling, and advanced packaging supply chains through at least 2028. Mizuho's analyst commentary on Amazon specifically highlighted accelerating demand for Amazon's custom AI chips — Trainium and Inferentia — as an additional signal that hyperscalers are actively diversifying away from pure NVIDIA dependency, though NVIDIA remains the dominant volume supplier Bloomberg.
Moonshot's Hopper Dependency Exposes the Limits of US Export Controls on China
Bloomberg reporting reveals that Moonshot AI's Kimi frontier models are trained and run on approximately 20,000 Nvidia Hopper-generation GPUs, accessed through a compute agreement with Alibaba Cloud rather than through direct procurement Bloomberg. Hopper chips — specifically the H800 variant — were available in China before the October 2023 export control tightening; the arrangement with Alibaba suggests these chips were stockpiled or accumulated prior to restriction and are now being shared across the cloud provider's enterprise client base.
The strategic reading here is twofold. First, US export controls have slowed but not severed Chinese frontier AI labs' access to high-end Nvidia compute — the inventory overhang from pre-restriction procurement is still being monetised. Second, Chinese hyperscalers like Alibaba are functioning as aggregators and redistributors of restricted compute, effectively extending the useful life of the installed base. This model complicates the US government's ability to assess effective compute ceilings for Chinese labs, since utilisation rather than ownership is now the operative variable. The semi-engineering week-in-review also flagged China's advancement in immersion DUV lithography as a parallel vector — less capable than EUV but potentially sufficient for mature-node domestic chip production at scale Semiconductor Engineering.
EU AI Gigafactory Tender Signals Serious Sovereign Compute Commitment
The European Union has issued a formal call for tender for seven supercomputing clusters under its AI gigafactory programme, with the initiative projected to unlock at least €20 billion in private co-investment Data Center Dynamics. This moves the EU's sovereign compute agenda from policy aspiration to procurement process — the tender stage implies budget allocation, technical specifications, and a competitive vendor selection process are underway. Seven clusters distributed across member states reflects the EU's characteristic approach of geographic balance over pure performance optimisation.
The strategic context is the EU's concern about dependence on US hyperscaler infrastructure for AI workloads touching sensitive public sector data, defence, and regulated industries. The gigafactory framing — deliberately invoking industrial-scale production — signals an intent to build capacity sufficient for frontier model training, not merely inference. The €20 billion private investment unlock figure is a projection, not a confirmed commitment; the actual leverage ratio will depend on the tender structure and whether the EU offers sufficient regulatory or procurement incentives to attract hardware vendors and operators beyond the usual suspects.
Africa's Data Centre Moment — Energy Constraints Remain the Binding Variable
Teraco CEO Jan Hnizdo's Bloomberg interview positions Africa as an emerging data centre destination, driven by improving subsea cable connectivity, growing domestic cloud demand, and a young population generating AI-relevant data Bloomberg. Teraco operates the continent's largest carrier-neutral colocation facilities, primarily in South Africa. The investment thesis rests on serving regional demand rather than competing with US or European hyperscale campuses for global AI training workloads.
The binding constraint Hnizdo identifies is energy — South Africa's chronic grid instability (load shedding), combined with limited renewable buildout in data-dense regions, makes power availability the primary risk factor for any major compute expansion. This mirrors constraints seen in other emerging market data centre markets: the demand signal is real, but the infrastructure prerequisites — reliable grid power, high-capacity fibre, cooling water — are not uniformly present. The sustainability angle is particularly sharp: diesel backup generation, widely used in African data centres to compensate for grid unreliability, is incompatible with the ESG commitments hyperscalers have attached to their expansion plans.
Signals & Trends
AI Memory Costs Are Now a First-Order Input to Hyperscaler Capex Planning
Amazon's disclosure that AI memory costs are a primary driver of its $220 billion capex trajectory is a materially new signal. Historically, hyperscaler infrastructure spending was dominated by server compute, networking, and real estate. The explicit elevation of memory — meaning HBM3E and next-generation DRAM — as a budget line item visible at the CFO level indicates that the DRAM supply chain, concentrated at SK Hynix, Samsung, and Micron, is now a strategic chokepoint equivalent to GPU supply. The Semiconductor Engineering week-in-review flagged DRAM shortages as a current market condition, suggesting the constraint is active, not hypothetical. Analysts and procurement teams should track HBM allocation agreements — which are typically negotiated 12-18 months forward — as a leading indicator of hyperscaler training capacity trajectories.
Compute-as-a-Service Is Becoming a Vector for Export Control Evasion
The Moonshot-Alibaba arrangement is likely not unique. The pattern — a frontier AI lab accessing restricted hardware through a cloud intermediary that accumulated inventory before controls tightened — is replicable across China's hyperscaler ecosystem. As Tencent, Baidu, and ByteDance operate large pre-restriction GPU fleets, the effective compute available to Chinese AI labs through cloud agreements may substantially exceed what export control compliance reporting captures. This creates a structural intelligence gap for policymakers and a competitive intelligence gap for Western AI labs trying to benchmark Chinese training capacity. The logical regulatory response — extending controls to cloud compute services — would require extraterritorial enforcement mechanisms that do not currently exist at scale.
PCIe Generation Transitions Signal a Storage Bottleneck Forming in AI Infrastructure
The concurrent relevance of both PCIe Gen5 and the emerging PCIe Gen6 standard for AI storage — as analysed by ServeTheHome — reflects a transition period where the industry has not converged on a single interconnect standard even as AI cluster buildout accelerates. Gen6 offers double the bandwidth of Gen5 but requires new controller silicon and retimers that are not yet at volume production. The practical implication is that data centres being built today are making storage interconnect decisions under uncertainty, with some operators standardising on Gen5 for deployment certainty while others delay for Gen6 performance gains. At exascale cluster sizes, storage throughput becomes a training bottleneck — this generation transition is worth tracking as a potential constraint on the efficiency of clusters coming online in 2027.
Explore Other Categories
Read detailed analysis in other strategic domains