Compute & Infrastructure

SoftBank has announced plans to invest as much as €75 billion to build 5 gigawatts of AI data centre capacity in France, positioning the country as Europe's leading AI infrastructure hub according to Bloomberg. This is an announced plan, not a confirmed build — no construction contracts, grid connection agreements, or planning approvals have been publicly confirmed at this stage. The scale, 5GW, would represent a transformative addition to European compute capacity, but it also strains credibility against current French grid capacity and land availability timelines. For context, all of France's existing hyperscale data centre capacity is a fraction of that figure.

The announcement reflects a broader sovereign infrastructure competition playing out across Europe, where governments are actively courting hyperscale and AI-era investment with tax incentives, permitting fast-tracks, and energy commitments. France's relative nuclear power advantage gives it credible energy headroom compared to Germany or the UK. However, investors and infrastructure professionals should treat SoftBank announcements with historical caution — the firm has a pattern of headline-scale pledges that evolve significantly in execution. The 5GW figure and the €75 billion envelope should be tracked against actual land acquisition, grid connection filings, and construction milestones before being treated as committed capacity.

Huawei's Rotating Chairman used the unveiling of a new LogicFolding chip architecture to explicitly thank the United States for its export controls, stating they drove Chinese firms to invest in domestic R&D and build a competing technology stack according to Tom's Hardware. The LogicFolding architecture has not yet been independently benchmarked, and its manufacturing process node and yield characteristics remain undisclosed — critical unknowns for assessing competitive parity with NVIDIA or AMD.

The strategic implication is significant regardless of where LogicFolding actually sits on the performance curve. China is demonstrably advancing on novel architectural approaches rather than simply attempting to replicate Western designs at a process disadvantage. The supply chain chokepoints that the US export control regime relies on — advanced EUV lithography, leading-edge packaging, high-bandwidth memory — are being systematically targeted for domestic substitution. The pace of that substitution remains contested, but the direction is no longer in doubt. Investors in the Asian AI supply chain are already pricing this in, as noted in Bloomberg's reporting on capital flows toward Asian supply chain companies benefiting from the AI buildout.

A Goldman Sachs report cited by Tom's Hardware projects that AI agents could increase token demand by 24 times relative to current prompt-response usage patterns. This is an analyst projection, not confirmed infrastructure data — but it has direct implications for inference compute planning. Agentic workflows chain multiple model calls, tool invocations, and context windows in ways that compound token consumption non-linearly. Uber and Microsoft are among enterprises already flagging rising AI costs as a material concern.

The 24x multiplier, if it approaches reality, means that inference infrastructure — not training — becomes the dominant capacity constraint. This shifts the hardware priority from large GPU clusters optimised for training runs toward distributed, low-latency inference infrastructure optimised for throughput and cost-per-token. It also explains why custom silicon investments by hyperscalers (Google TPUs, Amazon Trainium, Microsoft's Maia) are accelerating: at 24x token volumes, even marginal improvements in inference efficiency translate to billions in cost reduction. The ASUS B300 GPU server reviewed by ServeTheHome — featuring 8x NVIDIA Blackwell Ultra GPUs and over 6.4 Tbps of networking — represents the current high-end inference and training hardware frontier, but its power and cost profile will be stress-tested as agentic workloads scale.

Erin Brockovich has launched a structured tracking initiative for AI data centre community impacts, aggregating over 2,700 reports from across the US according to Tom's Hardware. The concerns span water consumption, grid stress, noise, and land use. Brockovich's brand carries specific legal and regulatory weight — her involvement in the PG&E case resulted in a $333 million settlement and demonstrated that coordinated community documentation can translate into material liability.

For infrastructure planners, the significance is not the current report count but the institutionalisation of the opposition. A centralised, professionally organised database of community complaints creates a discoverable evidentiary record that can support regulatory interventions, permitting challenges, and litigation. Data centre developers in water-stressed regions — particularly the US Southwest and Southeast — face elevated risk. The combination of environmental pushback, grid interconnection queues, and local permitting resistance is already extending project timelines; Brockovich's initiative adds a litigation vector to that risk stack.

Semiconductor Engineering reports that curvilinear mask designs required for high-NA EUV lithography are pushing inspection and metrology systems beyond their current capabilities, requiring new model-based verification approaches and native curvilinear data flows to control defect escape rates and mask turn times according to Semiconductor Engineering. This is a technical chokepoint that sits upstream of wafer production — if mask quality cannot be assured at volume, high-NA EUV yield will be constrained regardless of tool availability.

The mask supply chain is one of the least-discussed but most concentrated chokepoints in advanced semiconductor manufacturing. A small number of photomask suppliers — including Toppan, Dai Nippon Printing, and Photronics — serve the entire leading-edge market. High-NA EUV adds a new layer of complexity to an already demanding process. ASML's high-NA tools are beginning to reach customer fabs, but the supporting ecosystem — mask blanks, inspection tools, metrology software — must scale in parallel. Delays in any element of this ecosystem can bottleneck the entire node transition, with direct consequences for the cadence of next-generation AI accelerator production at TSMC and Intel Foundry.