Compute & Infrastructure

SK Group chairman Chey Tae-won stated at Nvidia's GTC conference that global memory chip shortage will likely persist for another four to five years, with wafer supply trailing demand by approximately 20%. Micron Technology responded by raising fiscal year capital expenditure guidance above $25 billion, significantly exceeding the $22.4 billion analyst consensus, as reported by Bloomberg. The company cited insatiable appetite for memory chips as AI workloads drive unprecedented consumption of HBM and high-capacity DDR5 modules. Micron's autonomous vehicle projections call for 300GB of DRAM per vehicle, with humanoid robots requiring similar quantities, according to The Register.

The supply constraint reflects structural manufacturing bottlenecks rather than temporary mismatches. Bloomberg noted the rush to boost memory production carries significant climate costs, as expanding fab capacity requires substantial energy and water resources. Industry analysts view the shortage as evidence that AI infrastructure buildout is running ahead of semiconductor manufacturing capacity, creating strategic vulnerability for hyperscalers dependent on memory supply for training and inference workloads.

Google's Michigan data centre project includes a 20-year electricity contract requiring the company to cover full costs of adding new clean power generation, as reported by Bloomberg. This represents a significant departure from traditional utility-financed generation, signalling that hyperscalers are accepting direct infrastructure financing obligations to secure power access. The arrangement includes solar investment requirements alongside guaranteed electricity supply. NTT Global Data Centers announced plans to double capacity to 4 gigawatts, with Bloomberg noting the third-largest global provider outside China is responding to AI-driven infrastructure demand.

Regional resistance to data centre expansion is mounting. Ohio residents are proposing a ban on facilities exceeding 25MW capacity, as reported by The Register, while a Franklin County, Missouri planning meeting debated two data centre projects for 11 hours without resolution, according to Data Center Dynamics. Iron Mountain is pursuing a seven-building campus outside Austin, Texas — its first Lone Star State facility — per Data Center Dynamics, while AWS filed permits for $90 million San Antonio development.

Bloomberg reported that accelerated memory chip production to meet AI demand will meaningfully increase semiconductor sector climate footprint and raise emissions management costs. The analysis highlights tension between compute scaling ambitions and sustainability commitments as manufacturers expand fab capacity. Memory production is particularly energy-intensive due to cleanroom requirements and multi-stage fabrication processes involving extreme temperatures and chemical etching.

The climate impact extends beyond chip manufacturing to data centre operations. Google's Michigan facility structure — mandating clean power investment as prerequisite for grid connection — suggests utilities and regulators are imposing environmental conditions on new capacity approval. This creates execution risk for announced projects lacking secured power sources meeting emissions requirements. Industry faces potential scenario where compute demand is suppressed not by capital availability or technical constraints, but by inability to source sufficient clean electricity at buildout sites.

GMI Cloud launched a $12 billion sovereign AI infrastructure initiative in Japan targeting 1 gigawatt of capacity, as reported by Data Center Dynamics. The scale of investment reflects Japanese government prioritisation of domestic compute capability as strategic imperative. South Korea's SDT opened a Quantum-AI data centre deploying a 20-qubit quantum computer integrated with Nvidia DGX B200 hardware, according to Data Center Dynamics, marking the country's first commercial quantum computing centre.

These developments indicate middle powers are treating AI compute infrastructure as critical sovereignty issue comparable to semiconductor manufacturing or energy security. QScale in Canada is conducting strategic review with potential $1.5 billion investment from Goldman Sachs, per Data Center Dynamics, suggesting institutional capital recognises sovereign compute as attractive asset class. Nvidia CEO Jensen Huang stated the company will build multi-billion-dollar CPU business with single Vera processor SKU, as reported by Tom's Hardware, indicating hardware vendors are positioning for government procurement cycles.

Data Center Dynamics reported OpenAI is reorganising leadership and readjusting data centre strategy, embracing renting AI servers from cloud providers instead of building all its own capacity. This represents significant strategic pivot for company that previously emphasised infrastructure ownership for training largest models. The shift suggests either capital constraints or recognition that hyperscale operators possess advantages in power procurement, facility management, and hardware refresh cycles that AI-native firms cannot replicate efficiently.

Separately, Tom's Hardware reported Microsoft is considering legal action against OpenAI over Sam Altman's recent deal with Amazon, with dispute centred on Frontier multi-agent service exclusivity. The potential litigation highlights tensions around infrastructure access and commercial relationships as OpenAI diversifies compute sources beyond Microsoft's Azure platform. Combined developments indicate OpenAI faces pressure to optimise infrastructure spending while navigating complex partnership obligations.