Compute & Infrastructure

SoftBank Group's mobile unit has announced it will begin large-scale battery cell production at its Sakai, Osaka plant, explicitly targeting the power demands of AI data centre operations. This is confirmed as a plan with a named facility, but volume, timeline, and offtake agreements have not been publicly disclosed, so execution risk remains. The strategic logic is clear: Japan's grid infrastructure faces capacity constraints as AI workloads scale, and battery storage at or near data centre sites reduces dependence on peak-load utility supply and improves uptime guarantees. Bloomberg

What makes this notable from a structural standpoint is the vertical integration play. SoftBank is simultaneously a major AI investor through Vision Fund and now positioning its operating units as infrastructure suppliers to the same AI ecosystem. If successful, this creates a captive supply chain for battery storage that could be preferentially allocated to SoftBank-backed data centre and AI ventures. The broader signal is that conventional utility-scale power procurement is increasingly viewed as insufficient — hyperscalers and their allies are moving to own or control energy buffering assets directly.

Nebius, Lambda, and CoreWeave — all of which have received NVIDIA investment or have close commercial ties to the chip maker — have confirmed they are unlikely to purchase Google TPUs in the foreseeable future. Data Center Dynamics The explanation, while partly attributable to the NVIDIA relationship, also reflects genuine ecosystem lock-in: CUDA toolchains, existing model optimisation pipelines, and customer expectations are all calibrated to NVIDIA hardware. Switching costs are non-trivial even for technically sophisticated operators.

This development is a direct indicator of how durable NVIDIA's moat is at the inference cloud layer. Google's TPUs are purpose-built, highly capable, and cost-competitive for specific workloads — but they remain essentially captive to Google's own infrastructure. The failure to attract third-party cloud operators as TPU customers means Google cannot achieve the network effects and external validation that would challenge NVIDIA's ecosystem dominance. For sovereign infrastructure planners and enterprises evaluating multi-vendor strategies, this further narrows the credible non-NVIDIA options to AMD (with limited enterprise traction at scale) and custom silicon still in early deployment.

The HP ZGX Nano G1n review from ServeTheHome documents a broader pattern: GB10-based small-form-factor AI systems across multiple vendors are selling nearly as fast as they are produced. ServeTheHome This is a demand-pull supply constraint, not a demand problem, and it indicates that the inference market at the workgroup and edge tier is maturing faster than production capacity can accommodate. GB10 — NVIDIA's Grace Blackwell superchip in compact form factor — offers meaningful on-device inference capability, and its popularity suggests enterprises are actively deploying local AI rather than defaulting entirely to cloud API consumption.

The supply constraint at this tier is structurally different from the H100/H200 data centre shortage of 2023-2024. GB10 systems involve more complex integration across multiple OEM partners and require different packaging and cooling approaches than rack-mount accelerators. Sustained supply tightness here would reinforce cloud API dependency by default — the opposite outcome of what enterprise buyers seeking data sovereignty or latency reduction are pursuing. This is a chokepoint worth monitoring.

Two separate incidents this week underscore a widening gap between AI infrastructure expansion speed and regulatory oversight capacity. In Georgia, a QTS data centre project consumed 29 million gallons of water over 15 months without authorisation — only detected when nearby residents noticed low water pressure. Officials declined to issue fines. Tom's Hardware Separately, communities near data centres across multiple jurisdictions are raising formal complaints about infrasound — low-frequency mechanical noise from HVAC and UPS systems that does not register on standard decibel meters but causes reported physiological effects. Tom's Hardware

The water incident is particularly significant from an infrastructure planning perspective. Cooling water consumption is a known scaling challenge — evaporative cooling systems for large AI compute clusters can consume millions of gallons annually in steady-state operation, and construction-phase water use adds to that burden. The fact that 29 million gallons went undetected for over a year points to inadequate metering, reporting requirements, and enforcement at the municipal level. As data centre density increases in secondary markets with less experienced regulatory infrastructure, these incidents will multiply. The infrasound issue introduces a harder-to-quantify siting risk: affected communities are beginning to develop the language and legal frameworks to challenge permits, which could extend approval timelines in markets where data centre expansion has previously faced little opposition.